The control axis isn't fit prediction — it's motion classification. The 6-dimension ICP scoring (firmographic / vertical / revenue / tech / growth / intent) predicts T1/T2/T3 fit tier; gtm_motion_class predicts which sales motion to use. Different concept; deserves its own output.

Customer can move along the control axis as their team grows / shrinks ML capacity. Reclassification triggered by AGT-208 production_signal threshold breach, ConvIntelligence pattern shift, or RevOps manual override. Triggers AGT-503 v38.5 candidate signal for control-axis tier-migration play.

Opt-in per company config. Traditional B2B SaaS without consumption-pricing tiered product mix uses the existing 6-dimension scoring without classification. Turning on requires AGT-208 deployed + AGT-106 per-class AE-pool definitions.

Sales motion now matches the actual buying motion at the customer. A T1 fully_managed account doesn't get a low_level_control AE who'd scare them off with technical depth; a T1 low_level_control account doesn't get a generic AE who can't hold a research conversation.

Margin-axis play retires GP-overlay quota on existing-revenue GP uplift (v38 model). Control-axis play retires GP-overlay quota on uplift + new revenue from expanded scope (low_level_control plays often add scope beyond initial product).

margin-axis-only (infrastructure-led play, classic v38) / control-axis-only (research-led play, new v38.5) / combined (rare, multi-quarter, high-value, requires SLM review at play creation) / no-migration (no signal).

A customer on shared infrastructure (low_margin) AND fully_managed control can grow in two distinct directions: infra maturation (margin axis) or ML team maturation (control axis). The sales motion shape, AE pool, comp incentive, technical-prerequisite gate — all differ. Modeling as a single axis would mis-route the play.

First-class motion for control-axis tier migration. Reps now incentivized via AGT-101 v38 GP-overlay to drive both margin-axis and control-axis migrations with appropriate cadence per axis.

fully_managed competes against hyperscaler bundle (largest TAM, lowest margin, growth bet). off_the_shelf competes against vertical AI infrastructure (mid TAM, mid margin, path to high via control-axis migration). low_level_control competes against self-hosted vLLM + frontier-model-provider-direct (smallest TAM, highest margin per customer, deepest moat). Sizing them as one number obscures where to invest sales capacity.

fully_managed → off_the_shelf: 5–15% per year. off_the_shelf → low_level_control: 3–8% per year (lower because requires customer-side ML team maturation). Reverse migrations tracked as net-of-forward-flow; flagged for AGT-903 strategic-bet retrospective.

MarketAssumptions gains control_tier ENUM + per-cell tam_account_count_estimate, sam_account_count_observed, som_account_count_target, arpu_estimate_usd, gross_margin_estimate_pct, competitive_set JSONB, last_analyst_refresh_date.

AGT-903 capacity-reallocation queries can now return per-control-tier rep allocation recommendations. Without this matrix, AGT-903 would be reasoning over a blended TAM that obscures the actual strategic levers.

v38 added motion (traditional / plg-warm / vertical-led). v38.5 adds gtm_motion_class (fully_managed / off_the_shelf / low_level_control). Combined routing key per AGT-202: (segment × vertical × motion × gtm_motion_class). Most cells are sparse — most accounts won't trigger every dimension.

Consumption-pricing businesses have GP that varies per customer / SKU / backend deployment in ways subscription SaaS does not. Tier-migration motions need a deterministic decomposition of where each customer's GP actually comes from — otherwise migration plays are bet-by-vibes. AGT-503 v38 explicitly depends on TOOL-015 for tier-migration play scoring.

(1) Pricing: list-vs-realized + commit-tier discount. (2) Utilization: effective unit cost vs benchmark, classified tight/moderate/loose. (3) Backend: spend by region + provider, region-arbitrage GP projection. (4) Tier: spend + GP by product tier, current-mix label.

Tool projects tier-migration scenarios that the caller supplies. Tool never invents "what if we migrated 30% of Serverless to Dedicated" on its own. This bounds the reasoning surface and prevents drift.

Every tier-migration scenario carries a required articulation of the case for NOT migrating. Mirrors AGT-903's option-set discipline at the tool level. Forces honest framing for downstream callers.

Haiku-default, ~$0.005/call. Sonnet only on AGT-903 deep-mode cohort fan-outs (~$0.05–0.10/call).

Specced v38, build deferred. Build prerequisites: synth corpus extension to consumption-event data with backend cost feeds; AGT-503 v38 ExpansionLog schema landing; benchmark_unit_cost_usd reference data.

Closes the per-customer GP-attribution gap that consumption-pricing tier-migration motions need. Foundational for AGT-503 v38 tier-migration plays + AGT-101/103 v38 GP-overlay comp.

Developer-led / API-platform businesses have a funnel where customers enter as individual devs who sign up + start consuming, often hitting material adoption before any sales conversation. AGT-201 + AGT-202 are post-MQL designed; they don't see this pre-sales adoption signal. AGT-208 is the upstream complement.

Consumption velocity (30 pts), production signal (25), enterprise context (20), commercial intent (15), stakeholder breadth (10). Threshold-based tier mapping: 80–100 priority / 60–79 warm / 40–59 monitor / <40 stay-self-serve.

Per-corporate-domain rollup nightly. Sum-capped at 100. Multi-developer override: 3+ devs at same domain each ≥ 60 forces handoff-priority regardless of summed score (organizational-adoption-breadth signal).

Lead Router gains motion dimension. PLG-warm motion routes to AEs with PLG-track territory designation; cadence is technical-depth-first, longer dwell, no inbound-form etiquette.

When AGT-201 has rated an account T3 (poor fit) but AGT-208 score ≥ 60, surfaces for RevOps review. Doesn't auto-promote — disagreement is signal.

Specced v38, build deferred. Build prerequisites: synth corpus extension to product-event stream + developer roster; AGT-204 enrichment-domain feeds; AGT-202 motion-flag wiring.

First L2 service designed for PLG-led developer funnels. Reduces time-to-handoff vs cold-outbound by ~50% in spec target; surfaces enterprise-progression signal before customers manually request a call (typically too late).

Tier migration is a post-sale motion against an existing customer with an existing relationship; AM is the executor; success is measured in GP delta on the same revenue. That's structurally an expansion play. The deviation from classic expansion is that ARR may not move — which is why the new signal scores GP uplift and AGT-101 v38 introduces a parallel GP-overlay comp variant.

play_type ENUM gains tier_migration. New fields: tier_migration_from, tier_migration_to, projected_gp_uplift_pct, switching_cost_class, realized_gp_uplift_pct (post-migration retrospective).

TOOL-015 returns the case for NOT migrating; AGT-503 pre-fills it on the play card. AM has the contrary view in hand before the customer conversation.

2 quarters post-migration: AGT-103 writes realized_gp_uplift_pct back to ExpansionLog. If realized falls below 60% of projected, AGT-703 picks up as calibration signal → TOOL-015 system prompt reviewed.

First-class motion for the consumption-pricing margin-expansion strategy. Reps can now be incentivized (via AGT-101 v38 GP-overlay) to drive tier migrations even when the ARR delta is flat — aligning behavior to the company's actual margin-expansion thesis.

Comp incentivizes behavior. Classic ARR-only comp pushes reps toward fast-close logo volume, which is the wrong incentive when the company's margin-expansion strategy runs through tier migration on existing customers. GP-overlay realigns rep behavior to the actual strategy.

AGT-103 queries TOOL-015 Consumption-Margin Decomposer for the customer's realized GP over the period. Per-opp GP attribution = (opp_arr / customer_total_arr) × customer_realized_gp. For tier-migration plays specifically: GP attribution = realized_gp_uplift_pct × baseline revenue, attributed to the closing rep.

A rep on GP-overlay could suppress ARR growth in service of GP improvement (refusing to upsell low-margin volume). Guard fires when combined attainment > 100% but ARR attainment < 70% — flags to AGT-104 for manual review.

AGT-103 writes realized GP uplift back to ExpansionLog post-migration. AGT-703 picks up systematic mis-projections as calibration signal. Closes the brain-discipline loop AGT-903 calls for at the operational comp-plan level.

Opt-in per rep. Defaults disabled. Turning on requires CRO + Finance approval (same gate as comp plan changes). Reps without overlay use classic ARR-only attainment unchanged.

The single highest-leverage spec change in v38. Comp architecture is the lever that turns strategy into rep behavior. Without GP-overlay, tier-migration plays are aspirational; with it, they're aligned to how reps get paid.

PLG vs traditional sales motion is structural — different sales process, different cadence, different qualification criteria. Adding motion to AGT-202 routing prevents PLG-warm leads from being routed to traditional inbound AEs who'd treat them as cold MQLs.

v38: AGT-106 TerritoryDefinitions gains motion_track field. PLG-track territories carve from the same account universe but with explicit motion designation. AEs may serve one or both motions; assignment is per-territory not per-AE.

Routing-layer support for parallel sales motions. Required by AGT-208; will be used by future vertical-motion services.

Per the standing rule, all v38 doc references describe the target as "consumption-pricing developer-led platform" or "tiered-product-mix multi-cloud SaaS" — never a specific company name. The OS spec is a generic capability surface, not a custom build.

8-tab walk extends to mention v38 capabilities — tier-migration in Planning panel scenarios, GP-overlay comp talk-track, AGT-208 PLG funnel framing.

AGT-901's read contract is current-period brain-ready views; extending it with multi-quarter cohort projections would have bloated its system prompt and conflated current-pipeline-operational reasoning with strategic-portfolio reasoning — different stakeholders, different cost profile, different evaluation criteria. Cleaner separation: each brain has a distinct horizon, distinct action taxonomy, distinct write target.

Strategy brain-ready view extensions on AGT-201 (icp_outcome_brain_view), AGT-205 (strategy_brain_view — TAM/SAM/SOM), AGT-501 (cohort_brain_view), AGT-503 (strategy_brain_view), AGT-702 (strategy_brain_view — multi-quarter Magic Number/NRR/GRR/R40/CAC Payback), AGT-703 (strategy_brain_view — multi-quarter win-loss + forecast bias evolution), AGT-105, AGT-101, AGT-404, AGT-604. New contract obligation on each owning Tier 1 service. Where a view does not yet exist, AGT-903 declines and surfaces the gap rather than estimating from raw rows.

BrainAnalysisLog (own log, shared schema with AGT-901/902) + StrategyRecommendationLog (drafts only). Explicitly does NOT write SalesPlayLibrary — plays are AGT-901/902's job, downstream of an endorsed strategic recommendation. May write annual-planning narrative sections to BusinessReviewLog per AGT-704 charter extension.

Options-shaped: 2–4 viable options per memo with distinct hypotheses (not minor variants), tradeoffs matrix per option, four-class risk surface (market/execution/capacity/model_assumption), falsifiable assumptions-must-hold list. If brain finds one obvious answer, it is required to articulate at least one credible alternative + the conditions under which it would be correct. Eval criterion enforces option count ≥ 2.

Rare-but-heavy: 10–30 queries/month at 150K input + 10K output, Opus default. Per-query cap 200K input + 20K output. Monthly budget alert at $300 default (CRO-configurable). Refusal on missing/stale strategy view is the cheapest correct answer (~one tool-call's tokens).

draft → under_review (CRO pickup) → endorsed (CRO + CFO joint approval; CEO required if scope_severity = material). Endorsement triggers a planning workstream owned by the relevant Tier 1 service — never modifies a canonical table. Healthy cadence: 2–5 endorsed recommendations per year. Strategic churn is a known anti-pattern; brain availability should not lower the cost of strategic second-guessing.

Specced only. Build deferred until (a) critical mass of strategy_brain_view extensions ships and (b) at least one Tier 3 cohort tool (likely TOOL-013 cohort retention forecaster, TOOL-014 segment-LTV decomposer) exists. Spec-first to anchor the contract obligations.

L9 brain layer feature-complete at the spec level — three brains cover current-pipeline, per-account, and multi-quarter portfolio horizons. The CRO-conversation range expands from operational coverage to strategic bets.

Different consumer (no execution engine reads endorsed rows directly), different state machine (shelved as a non-terminal "set aside"), different approval roles (CRO/CFO/CEO vs SLM/RevOps), different volume discipline (no cap; cadence-as-limiter), longer retention (warm 4y / cold 10y vs SalesPlayLibrary's 18mo / 7y) because strategic decisions have longer relevance horizons.

options_enumerated (JSONB array, 2–4 items, each with hypothesis + projected_impact_range + required_investment + capacity_implications + tier1_dependencies), tradeoffs_matrix, risk_surface (four-class structured), assumptions_must_hold (with brittleness flag), suggested_workstream_owners, endorsed_option_label, scope_severity (routine/significant/material), data_staleness_acknowledged.

Endorsement rate per quarter (target band 10–30% — too low = options not credible; too high = brain reading leadership preference). Edits during review (heavy = drafts are starting points; light = brain is calibrated). Brittle-assumption surfacing rate (for failed endorsements: did brain pre-flag the failing assumption?). Retrospective outcomes vs projected impact range, 4–6 quarters out.

SalesPlayLibrary's execution wiring (AGT-302 reads active plays) is irrelevant here. Endorsed rows trigger workstreams in AGT-201/AGT-105/AGT-101/AGT-205/AGT-203/AGT-302/AGT-403/AGT-406/AGT-802 depending on action_type. Each owning service runs its normal redesign cycle — strategic recommendation is an input, not a substitute for governance.

Same architectural rule as AGT-901 / AGT-902: brains never recompute Tier 1 numbers. Cohort retention curves, multi-quarter NRR/GRR trajectories, win-rate evolution by segment — all are deterministic computations that belong on the owning Tier 1 service. Brain estimation would break source-trace integrity.

Each Tier 1 service owns its strategy_brain_view extension. AGT-201 owns icp_outcome_brain_view (correlation of 6-dimension ICP score to realized LTV/NRR/churn/win-rate by cohort). AGT-501 + AGT-503 own cohort_brain_view (signup-quarter cohort retention curves). AGT-702 + AGT-703 own multi-quarter strategy_brain_view (extended time window of existing brain-ready views). AGT-205 owns strategy_brain_view of MarketAssumptions with current penetration. AGT-105/AGT-101/AGT-404/AGT-604 own their respective extensions.

Two Tier 3 tools likely needed to make AGT-903 useful: TOOL-013 cohort retention forecaster (analogous to TOOL-004 consumption forecaster — projects cohort retention curves forward with confidence bands) and TOOL-014 segment-LTV decomposer (decomposes LTV by segment/vertical/ICP-tier to support capacity-reallocation and ICP-revision use cases). Neither specced yet.

Contract obligation specced; views not yet built. AGT-903 build-readiness depends on critical mass of these views shipping. Each service can ship its strategy_brain_view independently — incremental rollout supported (AGT-903 declines on any missing view rather than estimating).

CRO walkthrough of v35 surfaced the demand: "show me the bad case." The growth-arc-only demo answered "does the system work for expansion" but not "does it catch risk humans missed." Both arcs together give the demo CRO conversational range without doubling cost or complexity.

Inside the Follow-a-Lead tab, a two-button selector lets the viewer switch between arcs without leaving the page. Both narratives load on tab open; switching is instant. Stark Logistics (growth) is the default; clicking the Risk button swaps to Massive Dynamic.

13 timeline steps: lead → discovery → close → go-live → healthy quarter → champion departure (Day +210) → decline → brain diagnosis → SLM intervention → account-team rotation → brain handoff briefing → renewal negotiation → renewal closed. Three amber brain moments at Day +275 (renewal-risk diagnosis), Day +296 (handoff briefing for new AE), Day +280 (stale-data variant query).

eval/samples/follow_a_lead_growth.json + eval/samples/follow_a_lead_decliner.json (the original was renamed for parallelism). Explainer JS loads both at tab-open, sub-tab toggles between them.

Same shape as TOOL-008 — deterministic core (signal aggregation per contact_id, strength classification per signal type) + LLM characterization (movement_type label + recommended interventions, hard-rule-enforced). Per-call cost ≈ $0.004 (Haiku, ~1K input + 700 output). Spec model was Haiku; matched.

Synth corpus has no LinkedIn or email-bounce signals → external_signal_coverage="none". Tool reports overall_quality="low" and degrades confidence per spec — left_company classification cannot be high-confidence without external signals. Smoke test on champion_loss_decliner: tool correctly classifies "stopped_engaging" with confidence="medium" (not "left_company" / "high"), surfaces 4 explicit ungrounded_assumptions.

Registered in prototype/tools/registry.py as tool_010_champion_movement_detector. Per-tool input augmentation in dispatch_tool derives tracked_champions + attendance signals from conversation_intelligence_log at call time (no corpus extension needed; champion roster is inferred from conv intel rather than separately tracked).

AGT-902 system prompt updated: TOOL-010 listed alongside TOOL-004 + TOOL-008. Brain calls when diagnosing renewal risk on champion_loss-shaped accounts. Will surface in re-curated brain_outputs.json samples after the next eval sweep.

Champion-loss diagnosis is now a structured signal, not narrative speculation. The decliner-arc renewal-risk diagnosis becomes more concrete: brain calls TOOL-010, gets back stopped_engaging classification with cited evidence, factors that into the recommended interventions.

Anything that sounds like "X has Y at Z threshold" must be grounded in specs/AGT-NNN_*.html, not working memory. Architectural framing (tier structure, single-writer-per-table) is safe to write from memory; specific spec values are not.

14 brain output samples — 0 spec drift. All proposed_actions[].lever references match spec ownership (AGT-302, AGT-203, AGT-503, AGT-603, AGT-504, AGT-501, AGT-902). The brain narratives are constrained by sources_read citations, so they don't make spec-attribution claims. The vulnerability is hand-authored narrative content (eval/samples/*.json, README.md, follow_a_lead JSONs).

gtm_os_explainer.html (simulator JS), eval/samples/follow_a_lead*.json, README.md, prototype/README.md, schema/GTM_OS_Changelog.html, prototype/PORT_TO_CORPORATE.md. Exempt: brain runtimes (their system prompts ARE the source of truth for action taxonomies), fixtures.py (test cases not narrative), synth/*.py (generates data not quotes spec).

Process discipline. Prevents recurrence; gives any future contributor (human or AI) the rule + checklist + audit table.

Curation script prototype/eval/curate_brain_samples.py reads brain_analysis_log.jsonl + brain_eval_log.jsonl, picks the latest passing run per fixture, strips run-state-only fields, attaches fixture metadata, writes static eval/samples/brain_outputs.json. Re-run after fresh evals to refresh.

CRO showed-up reaction to the prior explainer was: 3 deterministic JS simulators are graphical recreations of spec logic, not proof of working reasoning. The gallery is direct counter-evidence — every card is a real prototype output, citation-resolved, with audit trail visible.

$0 to render (static JSON). Total spend across the curated samples (the actual eval runs) was $1.49.

Day +120 (consumption overage qualification — EVAL-Q04 real output) and Day +130 (cross-team siloed adoption diagnosis — EVAL-Q11 real output, with TOOL-008 cited inline). Both pulled from the curated brain_outputs.json and rendered with their full BrainAnalysisLog: narrative, sources, proposed actions, tool calls.

expansion_ready archetype was picked over champion_loss_decliner for the CRO-friendly story arc — growth posture (expansion + renewal at 161% NRR) rather than defensive (decline + escalation). Every step references real spec artifacts (CadenceEventLog, ConvIntelligence, MutualActionPlan, OnboardingLog, ExpansionLog, ChurnRiskLog, etc.).

eval/samples/follow_a_lead.json — narrative scaffold with brain_moment refs to fixture_ids in brain_outputs.json. Edit the JSON to refine the narrative; no code changes needed.

CRO question was: "How could I run a lead end-to-end through this, like does the system actually work?" This tab is the answer. A click-through walks one account through every relevant layer of the DAG, with two amber-highlighted moments where the brain actually reasons.

In the spec era they were the strongest demo of "the system has been thought through." Now that AGT-902 + AGT-901 actually run, those JS scoring functions are the WEAKER part because they're hardcoded math, not reasoning. Brain Outputs + Follow a Lead replace them with substance.

The 6-dimension ICP scoring + routing demo is genuinely useful for explaining how AGT-201's dimensional model works. Narrow, deterministic, calibratable — keeps its niche. Renamed + sub-line set the right expectation that it's narrow.

scorePlanLocal, runPlanSim, scoreForecastLocal, runForecastSim — ~15K chars of JS scoring logic deleted from the explainer, replaced with the gallery + walkthrough renderers (~13K chars). Net code reduction.

Only multi-touch sequence-shaped actions become drafts: open_expansion_play (AGT-902, scope=account_specific) and draft_play (AGT-901, scope=segment). Single-shot interventions (pull_qbr_forward, customer_communication, escalate_to_slm, brief_new_ae_or_csm) and information actions (recommend_human_query, flag_coverage_gap, none) stay in BrainAnalysisLog and route to their target service or human via existing mechanisms — they are not plays.

Per the SalesPlayLibrary schema state machine: brains insert draft rows; humans transition draft → under_review → active. SLM + RevOps joint approval required for activation. Volume cap enforced at activation, not at draft write.

originating_proposal_id, originating_analysis_id, originating_action_type, originating_lever, brain_confidence, supporting_source_indices. Enables cohort-level retrospective: brain-co-designed plays vs. plays designed without brain involvement.

The most important signal for the human reviewer is the brain's reasoning — captured unmodified. Reviewer edits cadence + criteria + name in under_review; the originating hypothesis stays as written for retrospective traceability.

First 5 play-shaped actions per brain run get TOOL-003 enrichment; remainder fall back to placeholder cadence + criteria. Defensive guard against pathological brain runs producing many drafts and burning Haiku budget unexpectedly.

Three-tier human-codefinition loop is end-to-end demonstrable. Prototype now shows: brain proposes → writer drafts → tool enriches → viewer renders → (human picks up, mocked).

Status: specced → prototyped. Implementation matches the spec contract.

Spec called for Sonnet; prototype uses Haiku. Rationale: the task is narrow (cadence composition from a 1-2 sentence hypothesis), output budget bounded (1.5K tokens), and Haiku's per-call latency keeps the SalesPlayLibrary writer fast enough that brains can produce drafts synchronously without timing out. Per-call cost ≈ $0.004 (vs. ~$0.05 at Sonnet). Promote to Sonnet if production output quality slips below acceptable.

Tool NEVER invents account-specific facts. If hypothesis mentions "champion engagement", cadence may say "champion-engaged touchpoint" but cannot fabricate names. When numbers are guesses, confidence drops to "low" and ungrounded_assumptions is annotated. Smoke test: 5 explicit caveats listed for a single-call output.

Registered in prototype/tools/registry.py as tool_003_sales_play_composer. Handler in TOOL_HANDLERS but NOT in TOOL_DEFINITIONS — invoked by writer, not brain. Anthropic TOOL_003_DEFINITION included for future-use consistency.

EVAL-Q11 (AGT-902 expansion play) + EVAL-P01 (AGT-901 cohort plays) — drafts now have populated cadence + criteria instead of pending_human_codefinition placeholders. Total eval cost unchanged at <$0.20 per fixture.

SalesPlayLibrary drafts are now review-ready out of the box. Reviewers see: brain hypothesis (verbatim) + suggested 6-touch cadence + target meeting rate ≈ 0.7 + explicit ungrounded_assumptions list — instead of empty stubs.

JSONL is gitignored (run state). The viewer renders empty state when accessed via GitHub Pages — that's intentional. Local users serve the prototype dir with python3 -m http.server after running an eval.

"Pick up for review" / "Reject" buttons are visual mocks — state transitions need server-side enforcement that's out of prototype scope. The buttons demonstrate the workflow without implementing it.

Slate palette matching the explainer (v33). L9 amber for writer badges (AGT-901 / AGT-902) consistent with brain layer theme. State pills: amber=draft, blue=under_review, green=active, gray=retired.

First visible demo of the human-codefinition workflow. Brain runs produce JSONL → viewer renders cards → demonstrates the promotion gate without needing a full workspace UI.

load_account_corpus(account_id) · account_exists(account_id) · account_data_freshness(account_id) · iterate_account_ids() · metadata(). Five methods are everything brains and tools depend on.

SynthCorpusSource (production-grade for the prototype, file-mtime staleness, ground_truth.json-driven account iteration) and WarehouseViewSource (stub class — every method raises NotImplementedError, documents what a real impl returns).

Source threaded through run_for_account, run_for_pipeline, call_brain, dispatch_tool, extract_brain_ready_view, extract_pipeline_view. Legacy file-IO branches preserved for backwards compat but unused under default execution. Eval verified: 14/14 fixtures pass through the seam.

prototype/PORT_TO_CORPORATE.md updated with concrete WarehouseViewSource subclass example. Brain-ready view extractors do NOT need to change when porting — only compression budgets per component.

C1 Invented action_type (AGT-902 enum) — validator must emit hard taxonomy issue. C1b AGT-902 action_type leaked into AGT-901 enum — verifies taxonomy= parameter is enforced. C2 Unresolved [src:99] citation — validator must emit hard citations issue. C2b Negative control — all citations resolve cleanly, validator silent (no false positives). Sanity default taxonomy is AGT-902 (legacy callers don't accidentally validate against AGT-901's enum).

_check_must_cite_tool in scorer now dual-checks dispatch-side tool_calls_made AND narrative-side sources_read. Catches fabricated citations (calibration probe finding from previous wave). Detail messages explain which side failed.

Add a new probe whenever a new validator rule is added to validate_all. (1) Construct synthetic output violating the rule. (2) Call the validator function directly. (3) Assert the rule's hard issue is in the returned issues list. (4) Add a negative-control probe for false-positive coverage.

Active nav state distinguished by bottom-border underline rather than full colored border + bg tint. Italic "complete" in headline rendered as bold-italic with slate underline (3px, 8px offset). Net effect: layer rainbow palette and L9 amber are the only color in the document; brand chrome goes monochrome.

Last L1-L8 spec name carrying legacy "Agent" suffix renamed (AGT-101 "Quota Setting Agent" → "Quota Setting"). All other L1-L8 service names already clean. L9 "Pipeline Brain" / "Account Brain" / "Reasoning (Tier 2 Agents)" preserved — those ARE agents.

Status: specced → prototyped. Runtime exists, evals pass, not yet wired to canonical Tier 1 tables in production.

11 fixtures across question types: churn_diagnosis (Q01, Q02, Q09), expansion_qualification (Q04, Q05, Q11), handoff_briefing (Q06), qbr_narrative (Q08), diagnosis_on_stale_data (Q07, Q10), stalled_onboarding (Q03). All 11 pass at $1.16/run.

pull_qbr_forward / open_expansion_play / brief_new_ae_or_csm / customer_communication / escalate_to_slm / recommend_human_query / none. Enforced by validator at eval time.

4096 — sufficient for per-account narratives. (Cohort brains require 8192; see AGT-901 entry.)

Three-tier architecture validated end-to-end on synthetic data. Brain-ready view contract works as the seam between Tier 1 (deterministic) and Tier 2 (LLM-native).

Status: specced → prototyped. New aggregate view extractor at prototype/aggregates.py rolls up segment / vertical / ICP-tier dimensions plus top-K drill-down anchors.

3 fixtures across question types: cohort_diagnosis (P01), expansion_prioritization (P02), coverage_gap (P03). All 3 pass at $0.32/run. New criterion must_cite_source verifies the brain cites the specific rollup table its claims came from.

draft_play / flag_coverage_gap / recommend_query_for_human / none. Distinct from AGT-902 — pipeline-shaped (cohort actions), not per-account.

8192 — cohort narratives are longer than per-account. First smoke test failed three retries at 4096 with consistent JSON truncation around char 12K. Captured as a calibration learning: per-account brains run on 4K, cohort/multi-source brains require 8K+.

Brain calls TOOL-004 / TOOL-008 with account_id from top_expansion_candidates / top_churn_risks / stalled_onboardings drill-down anchors. Validates cohort hypotheses against per-account evidence — proves the cross-tier pattern.

Two-brain architecture validated. AGT-901 (cohort) and AGT-902 (per-account) operate as complements, not duplicates.

Status: specced → prototyped. Implementation matches the spec: deterministic numerical core in code, narrow LLM characterization for pattern label + interpretation.

Registered in prototype/tools/registry.py as tool_004_consumption_forecast. Schema in TOOL_DEFINITIONS; handler in TOOL_HANDLERS; corpus augmentation in dispatch_tool reads UsageMeteringLog from the synth corpus.

EVAL-Q04 (real expansion confirmed), EVAL-Q05 (spike-then-crash trap), EVAL-Q11 (siloed expansion qualification). Brain consistently cites tool output as canonical source via [src:N].

UBB-specific cognition gap closed in prototype. Pattern: numerical work in code, characterization in LLM — applies to TOOL-008 and future tools.

Status: specced → prototyped. Required new corpus extension: 23-feature engagement telemetry per account (5 categories: core / advanced / integration / admin / experimental — modeled on a B2B API product with UBB pricing).

8/8 archetypes correctly classified including the spec's hard onboarding-aware rule. Deeply_integrated (ideal_power_user) and siloed_by_team (expansion_ready) discriminated by Gini concentration threshold; activating (stalled_onboarding) preserved for <60d accounts.

EVAL-Q11 with new must_cite_tool_008 criterion. Brain consistently uses tool output for adoption-depth diagnosis when consumption-volume signals are ambiguous.

Post-sales feature-level adoption gap closed in prototype. Three-tool post-sales activation stack (TOOL-008 + TOOL-009 + TOOL-012) now has its first member operational.

CSS rule .status-prototyped added to explainer (gold/amber: #FAEEDA bg, #7A4A0E fg, #D4A84C border). Tools_Index gets .tool-status-prototyped with same palette.

"Built" implies production deployment with real data, real users, real audit trail. The prototype runs against synthetic data and isn't wired to canonical Tier 1 tables. Calling it "built" would overstate readiness; calling it "specced" understates the work done. The third state captures the truth.

Documentation precision. Forward changelog entries can promote incrementally: specced → prototyped → built.

synth/archetypes.py (8 archetypes with usage/health/conversation/feature profiles), synth/main.py (corpus orchestrator), synth/usage.py / health.py / payments.py / feature_engagement.py (per-component generators), synth/conversations.py (LLM-generated call summaries with cache).

Seed-driven for reproducibility. Feature-seed derived via hashlib.sha256(account_id) rather than consuming from the main RNG — invariant: corpus regen does not shift account UUIDs, conversation cache stays stable.

Per prototype/PORT_TO_CORPORATE.md: synth/ generators are throwaway when porting to a corporate environment. Real Tier 1 services replace them. The boundary is the brain-ready view extractor.

Calibration probe on EVAL-Q11: temporarily removed TOOL-008 from TOOL_DEFINITIONS only. The brain still emitted tool_use calls for TOOL-008 because the system prompt described it; dispatch found the handler (still registered) and ran the tool successfully. False-pass on what should have been a failing rule. Real probe required removing handler too.

First smoke test of AGT-901 failed three retries in a row with Unterminated string at char 12383 (per-attempt). Bumping max_tokens from 4096 to 8192 fixed it on first try. The failure mode looks like a transient parser flake but it's actually a budget cap. First thing to check when a brain hits sudden JSON-decode errors.

Discipline for retiring tools and provisioning new brains. Both findings preserved in user's eval-design memory and in prototype/PORT_TO_CORPORATE.md's "risks worth flagging early" section.

Easier to write the bridge doc while the prototype is fresh than to reverse-engineer it from code later. Captures rationale, not just code.

Per the original architecture doc's recommendation: L8 first (UsageMeteringLog + AGT-804 revenue recognition), brain second. Brain reading from stub UsageMeteringLog produces fake-confident output — the worst possible failure mode for an audit-sensitive system.

New tool: TOOL-009. Haiku, 8K input + 1.5K output, $200/mo cap.

TOOL-008 is descriptive (current state of feature adoption). TOOL-009 is longitudinal (timing of milestone achievement vs benchmark). Both can be called for the same account; they answer different questions.

TOOL-009 measures observed delta vs benchmark on configured milestones. TOOL-012 (also v32) projects 6-month outcome from early signals. Three post-sales activation tools work as a stack: TOOL-008 (current state) + TOOL-009 (timing analysis) + TOOL-012 (outcome prediction).

Milestones configured in input (RevOps + Product define), tool measures against them — never invents milestones. Cohort-relative when baseline available, graceful degradation when not. Honest projection confidence (low/medium when activity signal is weak).

AGT-601 Onboarding Orchestrator (weekly batch on onboarding cohort), AGT-501 Customer Health (months 3–9 ramping accounts), AGT-902 Account Brain ("is this account activating on time?"), AGT-704 (cohort-level TTV trends for QBR retention section).

New tool: TOOL-010. Haiku, 10K input + 2K output, $250/mo cap.

Single weak signals do not produce high-confidence movement classifications. Title change without engagement decline may just be a promotion. Email bounce alone could be OOO misconfig. Tool fuses LinkedIn signals, email-bounce signals, AGT-407 ConvIntelligence call attendance, email engagement — produces classification with explicit contributing signals listed.

Five classifications: left_company, role_changed_internal, stopped_engaging, engagement_declining, no_movement_detected. Each maps to different intervention recommendations — champion who left vs. shifted internal role vs. disengaged are different problems.

Every classification stronger than no_movement_detected requires at least one contributing signal with strength ≥ moderate. Privacy-disciplined (only signals already in OS or from approved enrichment under signed data agreements). Single-signal restraint: weak single signals never produce high confidence.

AGT-502 Churn Risk (weekly batch on T-180 renewal cohort), AGT-401 Deal Health (when champion is economic buyer/strong influencer), AGT-902 Account Brain, AGT-501 Customer Health.

High-leverage early-warning signal. External enrichment required for full effectiveness — without LinkedIn/email-bounce signals, tool degrades to internal-only mode with lower confidence ceilings.

New tool: TOOL-011. Sonnet, 20K input + 2.5K output, $300/mo cap.

Four classifications: low_sensitivity (concession unnecessary — discount erodes margin without improving close), moderate_sensitivity (cohort-typical envelope), high_sensitivity (multiple revisions likely — consider non-price levers first), highly_constrained_budget (qualify out or reduce scope, don't discount to meet budget).

Output never contains specific discount percentages or unit prices — regex-enforced in eval. Tool produces classification + cohort context + observation patterns; human deal desk reviewer determines actual discount magnitude per AGT-406 spec. Cohort baseline degradation honest. Aggregate discount drift monitored quarterly — if pre-tool vs post-tool median discount drifts > 2pp, calibration sprint triggered.

do_not_concede_flags at least as valuable as the sensitivity classification. Examples: weak_qualifier (champion not confirmed — don't discount until qualified), discount_above_authority, contract_term_concession_higher_leverage (term length / payment terms / multi-year may have higher leverage than equivalent price concession; preserves list-price baseline for renewals).

AGT-406 CPQ & Deal Desk (every quote requiring approval — primary), AGT-401 Deal Health (proposal/negotiation stage), AGT-902 Account Brain, RevOps direct (pricing strategy work).

Augments AGT-406 deal desk decision support. Approval flow and discount authority remain unchanged. Calibration risk — aggregate use of the tool must not produce monotonic discount drift over time.

New tool: TOOL-012. Haiku, 10K input + 2K output, $200/mo cap.

Five classifications: strong_activation (cohort-leading signals), on_track_activation (cohort-typical), concerning_trajectory (one or two weak signals), early_stall_risk (multiple weak indicators clustered — multi-pronged intervention warranted), active_stall (effectively stopped, escalate to SLM, may require contract-level intervention).

TOOL-008 (descriptive: current adoption pattern) + TOOL-009 (longitudinal: timing vs benchmark) + TOOL-012 (predictive: project 6-month outcome). Three different questions for the same account; AGT-501 / AGT-601 / AGT-902 may invoke any combination depending on the question being asked.

Cohort-anchored predictions (without cohort baseline, projection_confidence drops). Intervention-window-aware (same weak signal at day 14 vs. day 60 has different intervention implications). Honest about projection uncertainty — AGT-601 weights interventions by confidence so low-confidence predictions don't trigger heavy interventions.

AGT-601 Onboarding Orchestrator (weekly batch on first-90-day cohort, primary), AGT-501 Customer Health (early forward-looking signal), AGT-902 Account Brain, AGT-704 (cohort-level onboarding trajectory in QBR retention section).

Closes predictive gap in post-sales activation toolkit. Onboarding intervention window is the highest-leverage moment in the customer lifecycle — getting prediction right early matters more than getting any other post-sales signal right.

Tools_Index header updated to "12 tools across 3 waves." Cost aggregation table extended with TOOL-009/010/011/012 budgets. Third-wave-candidates section replaced with future-candidates section listing 5 case-by-case ideas (Procurement Negotiation Pattern Recognizer, Multi-thread Quality Scorer, QBR Action-Item Outcome Tracker, Renewal Negotiation Risk Profiler, Industry-Specific Play Refiner) flagged for if/when operational signal warrants.

Tier 3 default budget: ~$4,800/mo across 12 tools. Combined with Tier 2 brain budgets ~$1,500/mo, reasoning + augmentation layer total ~$6,300/mo — roughly 1/3 of one RevOps analyst FTE. Bounded; trackable in single dashboard with per-tier 75% alerts.

"Add tools when gaps emerge from operational signal, not when they emerge from theory." 12 tools is already substantial maintenance load alongside the eval harness. Further additions trade off against catalog maintenance.

Tier 3 catalogue broadly complete relative to v26 architecture eval gaps. Restraint over volume going forward — the discipline is now to keep the catalog sharp, not to grow it.

New tool: TOOL-005. Haiku, 8K input + 2K output, $150/mo cap.

Recommendation grounding 100% (every recommended_action traces to a specific risk_flag). False positive rate on green periods ≤ 10% (over-flagging undermines the tool). Statistical significance check before flag generation; minimum send volume threshold prevents variance noise.

AGT-302 (daily batch, drives auto-throttle decisions with RevOps notification), AGT-303 (weekly batch, feeds risk-classified recommendations), AGT-901 Pipeline Brain (diagnostic queries), RevOps direct (workspace UI investigation).

Closes Domain 4 gap — AGT-302/303 cover sequence generation/execution/optimization but no component owned deliverability health observability until now.

New tool: TOOL-006. Sonnet (with Haiku fallback for low-stakes), 30K input + 1.5K output per invocation, ~30 invocations per customer call (60s cadence over 30 min) ≈ $3/call. Monthly cap $2,000/mo (~600 calls).

Live-call sidecar UI (recording-platform integration with Gong / Zoom / Chorus). Rep can pause/resume during off-the-record moments. No auto-action; rep decides whether to use any specific suggestion.

AGT-407 retrospective + TOOL-006 live = full conversation intelligence loop. CallGuidanceLog (new schema) FK's to ConvIntelligence; AGT-701 reads both for full coaching picture.

Closes Domain 2 (in-call) gap. Higher infrastructure complexity than other tools — live audio integration + sidecar UI surface required. Implementation gated on recording-platform partnerships.

New tool: TOOL-007. Sonnet, 15K input + 2K output, $200/mo cap.

KB grounding 100% (every talking_point cites a real KB section). Hallucinated competitor claims 0% (cannot extend beyond what's in the KB). Stale KB recognition 100% (entries > 90d old must surface staleness caveat).

AGT-403 owns the canonical CompetitiveKnowledgeBase. TOOL-007 composes deal-specific narratives from it — never extends or corrects. KB updates remain AGT-403's responsibility.

Talking points read differently from email copy from battlecard excerpts. Tool adapts tone + structure to narrative_target. Output includes a do_not_say field flagging anti-patterns specific to the competitor — equally important as the talking points.

Domain 2 quality lift — the difference between "here's our positioning vs. X" and "for this prospect at this stage with these specific concerns, here's how to frame the X comparison." Quality lift, not gap-closer.

New tool: TOOL-008. Haiku, 10K input + 1.5K output, $300/mo cap (high frequency — daily batch from AGT-501 across active accounts).

deeply_integrated (broad breadth + high users-per-feature + integrations active — strong renewal + expansion signal); surface_only (core features only, low per-feature adoption — competitor swap risk); siloed_by_team (decent breadth concentrated in one team — cross-team expansion signal but champion-loss risk); declining (abandoned > new — strong churn signal); activating (newly-onboarded with rapid breadth growth — onboarding success).

Onboarding-aware classification (account < 60 days into contract not labeled "surface_only" — must be "activating" or pre-pattern, 100% eval-enforced). Cohort-baseline graceful degradation (returns medium/low data_completeness; doesn't fabricate cohort comparisons).

AGT-501 Customer Health (daily batch — augments feature engagement dimension), AGT-902 Account Brain ("is this account getting value?"), AGT-503 Expansion Trigger (qualification augment, deeply_integrated/siloed boosts expansion signal, declining suppresses), AGT-601 Onboarding Orchestrator.

TOOL-004 is consumption volume (will overage hit, when, what trend). TOOL-008 is feature engagement (which features are used and how broadly). Complementary — both used together for full post-sales picture.

Closes post-sales feature-level adoption gap — AGT-501 had seat utilization but not feature engagement until now. That distinction predicts churn/expansion much earlier than seat metrics alone.

Rep Performance & Coaching, Specced (v21) → Built (v31). Role-parameterized across AE/AM/CSM/SDR. Manager approval gate on coaching delivery preserved. Reads ConvIntelligence with v23 call_owner_role filter for role-specific signals.

GTM Health Monitor, Specced (v21) → Built (v31). Magic Number, NRR, GRR, R40, CAC Payback. Promotion follows L8 buildout because Magic Number's denominator separates billed (AGT-802) from recognized (AGT-804) revenue — AGT-702 cannot be Built until those distinctions are deterministic.

Win-Loss & Forecast Accuracy, Specced (v21) → Built (v31). Produces calibration signals that propagate back into AGT-101 (quota), AGT-201 (ICP), AGT-402 (forecast), AGT-403 (competitive), AGT-404 (top-down). Without AGT-703 Built, OS had no closed-loop learning — this completes the loop.

Business Review Orchestrator, Specced (v21) → Built (v31). Closes L7. Reads from AGT-701/702/703 with hard staleness gates. v26 charter extension intact: brain narrative may write narrative sections of MBR/QBR but never metric sections. AGT-901 invokes for narrative jobs only.

Closed-loop learning operational. Calibration signals from AGT-703 → upstream services. Business reviews fully assembled with audit-grade source-trace. L9 brains can read AGT-702 MetricsCalc.brain_view + AGT-703 WinLossLog.brain_view + AGT-701 RepCoachingLog with confidence — the brain-ready view contracts on these sources are now backed by Built services.

AGT-107 Quota Plan Document, Specced (v23) → Built (v31). E-signature gate, three output formats (HTML/PDF/JSON), amendment recommendation logic. AGT-104 reads QuotaPlanDocLog.signature_received as one of its 12 governance audit controls.

Every L1 agent now Built. AGT-101 quota → AGT-102 comp → AGT-103 attainment → AGT-104 governance → AGT-105 capacity → AGT-106 territory → AGT-107 plan document. Hiring-quota-comp cycle is a complete operational loop.

Combined with L7 closure: 36 of 40 services Built across L1-L8. Remaining specced services: none. L0 infrastructure remains config-only (no agents).

SVG edge color logic in renderMap() updated: edges with source agent ID starting with AGT-9 render in #D4A84C (L9 layer color). All other edges render per existing type-color mapping (feeds=green, reads=blue, cross=orange).

Connection Map legend now lists 4 edge types: Feeds (Tier 1 → Tier 1), Reads (Tier 1 → Tier 1), Cross-layer (Tier 1 → Tier 1), Brain reads (L9 → Tier 1). Map instructions text updated to mention brain agents.

Cosmetic. The reach of L9 brains across the Tier 1 layers is visible at a glance — previously brain edges rendered identically to service-to-service reads, obscuring the architectural distinction.

AGT-801 promoted from Specced (v25) to Built (v30). Spec hardened with Production Controls section (8 controls) + Edge Cases & Failure Modes section (8 scenarios). Existing schema unchanged; ripple of v27 (originating_proposal_id) noted.

Two-track approval enforced (Pricing track + Legal track tracked independently); pricing correction Finance Director floor; amendment lineage preserved (originals never overwritten); auto-renewal clauses surfaced explicitly (non-NULL required); TCV computed deterministically (no manual override); billing contact validated (non-NULL on activation); SOW reference required when services on order; 7-year retention parallels RevenueRecognitionLog.

Customer-rescinds-before-signature; mid-term seat expansion mid-billing-cycle; pricing correction post-invoice (forces AGT-804 prior-period adjustment); AGT-802 offline at activation; handwritten markup discrepancy; two-track approval mid-flight imbalance; SKU-add with conflicting payment terms; auto-renewal trigger.

L8 buildout sequence per v26 plan: AGT-804 first (recognition is audit-critical), then AGT-801 (orders source-of-truth) before AGT-802 (billing depends on orders + recognition guarantees). Cascade complete with this v30 promotion.

L8 entry point Built. Order data is now contractually-locked source of truth for every downstream finance agent. External integration still required for go-live — spec is production target, not a deployed system.

AGT-802 promoted from Specced (v25) to Built (v30). Production Controls + Edge Cases sections added.

Invoice amounts traceable to OrderLog (every InvoiceLineItems carries order_line_item_id FK); consumption invoices read MAX(record_version) from UsageMeteringLog (corrections produce credit memos, not silent invoice updates); customer sign-off required for milestone billing (IE-only completion does not unlock); credit memo two-step gate for rep-submitted (manager + Finance Director); billing error → Finance Director floor; idempotent invoice generation (no duplicate billing on retry/replay); payment terms inherited not overridden (changes require AGT-801 amendment); daily external billing system reconciliation at ±$0.01 tolerance.

Customer disputes consumption invoice (UsageMeteringLog.audit_status = disputed flow); SOW partial completion (no partial billing); mid-cycle amendment pro-rated supplemental; annual upfront pre-payment (cash and recognition decoupled per ASC 606); external billing system offline (internal records held, retry on recovery); reconciliation discrepancy (AGT-803 holds payment tracking on affected invoices); externally-processed refund without AmendmentLog (forces back-fill); mid-period churn with prepaid balance.

Customer-facing invoice generation now production-spec'd. External billing system integration required for go-live. Architectural invariants preserved — AGT-802 doesn't decide what was sold or what's recognized; executes against AGT-801 + AGT-804.

AGT-803 promoted from Specced (v25) to Built (v30). Closes L8 buildout — every L8 agent now Built.

State transitions logged immutably (PaymentEventLog append-only with prior_state/new_state/transition_reason); per-account state aggregation (worst-applicable wins across multiple invoices); Suspended state requires Finance director-level authorization (no automatic transition; CSM/AM cannot trigger); retry exhaustion before AM notification (prevents storms on transient failures); AGT-501 modifier is read-only feed (writes only payment_health_status, not other CustomerHealthLog fields); external failure events idempotent (dedupe key: invoice_id + attempt_number + external_event_id); auto-recovery on payment to Current state; 7-year retention.

Multi-invoice account with mixed states (worst-applicable wins); Failed-invoice paid same day new invoice goes Overdue (no re-notify); external billing system payment-confirmation gap (held by reconciliation); customer changes payment method during Overdue; alternate payment methods on file (Retry 2/3 may use); mid-cycle Suspended cascade to AGT-501/502/503/504; reactivation from Suspended; payment reversal/chargeback recorded as distinct event; AGT-803 offline during failure event (recovery preserves accuracy).

Post-sales signal chain now fully Built end-to-end: UsageMeteringLog (production schema v26) → AGT-501 health (built) → AGT-502 churn (built) → AGT-503 expansion (built), with AGT-803 payment health gating the cap (built). The post-sales backbone is the strongest part of the OS — per the v26 architecture eval (95% domain coverage).

L8 fully Built. Deterministic backbone complete L1–L8. L9 brains can now read against trustworthy source data — the architectural prerequisite for production launch.

New schema artifact: Brain_Ready_Views_Contract.html. Defines the contract pattern + catalogues all 20 views (9 cross-population for AGT-901, 11 per-account components for AGT-902 composite).

Until v30, brain-ready views existed only as references in AGT-901 and AGT-902 specs. AGT-901 references 9 (MetricsCalc, Opportunities, ForecastLog, Accounts, AccountPriorityScore, CapacityPlan, WinLossLog, CadenceEventLog, VoCSynthesisLog). AGT-902 references 11 source views joined into a per-account composite. Without formal definitions, the brains were theory. Without compression discipline (10× typical), brain costs would explode.

Source table, owning service, projection rule, filter rule, refresh cadence, staleness threshold, token budget, used-by list, backwards compatibility rules. Every view in the catalogue has all 8 elements specified.

Three patterns: top-N + aggregate (most cross-population views); time-windowed slice (time-series views like UsageMeteringLog, MetricsCalc); per-account composite (precomputed denormalized snapshot for AGT-902). The composite is the operationally critical pattern — constructing it at brain query time would blow latency and token budget.

Every view exposes view_metadata.last_refresh_timestamp + staleness_threshold_hours + is_stale. Brain reads is_stale first; if TRUE, sets BrainAnalysisLog.data_staleness_acknowledged = TRUE and surfaces in narrative_output. Per-component staleness for the composite (AGT-902 may proceed with fresh components, surface specific stale dimensions).

Field additions: non-breaking. Field removals: breaking, requires coordinated brain prompt deploy. Projection rule changes: case-by-case. Refresh cadence faster: non-breaking (cost may rise). Cadence slower: possibly breaking (staleness threshold may need adjustment). Cadence + threshold changes logged in changelog.

Materialized view storage trivial. Daily refresh compute modest (piggybacks on existing AGT-XXX nightly jobs). Per-account composite refresh slightly higher (per active account, daily). View-driven brain cost reduction ~10× vs raw-table reads — saves >~$200/month at expected volumes. Net: views save more than they cost.

L9 brains have a real, formal read contract. Implementation work substantial — 10 materialized views + 1 composite refresh job. Eval linkage: BrainEvalLog.brain_view_contracts_hash detects view contract changes between runs; forces re-eval before brain promotion.

AGT-302 spec gains "L9 wire item (v27)" section parallel to existing L2 wire item. Documents two integrations: (1) reading active plays from SalesPlayLibrary alongside existing AGT-301/AGT-203 sequence sources, (2) writing play_id and originating_proposal_id to CadenceEventLog for cohort retrospective lineage.

v27 ripple was logged in changelog but never written into AGT-302's own spec. AGT-302 is the highest-traffic execution service in the OS — its spec needs to reflect every coordination wire, including the L9 one. Spec drift is itself a failure mode; fixing it before more services build against AGT-302 documentation.

All AGT-302 v22 behavior unchanged: 3-touch/week cap, ABM backstop, suppression compliance gates. Brain-co-designed plays do not bypass coordination guards. Both new CadenceEventLog columns are nullable; existing rows have NULL on both; existing AGT-301/AGT-304/AGT-203 paths continue without writing the new fields.

AGT-303 Cadence Intelligence reads CadenceEventLog including new fields — brain-co-designed plays appear in cadence performance analytics with play_id grouping (per AGT-303 v27 brain-ready view extension referenced in Brain_Ready_Views_Contract).

Documentation hygiene — spec now matches what changelog claims is true. Cohort retrospective is queryable — the CadenceEventLog originating_proposal_id column was theoretical until AGT-302's spec confirmed write behavior.

New artifact category established. Tools live alongside the layered services and agents but follow different rules. Stored in tools/ directory. Indexed by Tools_Index.html.

Tools are pure functions: stateless, no table ownership, no cadence, no approval gates. Treating them as layered components creates wrong governance overhead; treating them as opaque LLM calls creates wrong audit posture. Tier 3 splits the difference — callable contracts with eval discipline, but no layer-integration burden. Per the v26 architecture eval recommendation.

Every Tier 3 tool spec must define: purpose (1 sentence), input schema (strict JSON), output schema (strict JSON; schema changes are breaking), model tier (Haiku default unless task demands otherwise), called-by list, cost ceiling (per-call + monthly), eval criteria, declared failure modes.

Brain calls tool during query (most common); service calls tool as enrichment (e.g., AGT-201/AGT-204 calling TOOL-002); operator calls tool directly via workspace UI; tool chains (brain orchestrating multiple tools in sequence).

First-wave default budgets total ~$1,200/mo across the 4 tools. Combined with Tier 2 brain budgets (~$1,500/mo), reasoning layer total ~$2,700/mo. Tracked in single dashboard; alerts at 75% per tier.

Tools have own eval criteria, lighter than brain harness — typically 10–15 questions per tool. Schema compliance is hard requirement (100%); hallucination/grounding metrics are tool-specific. Tool eval results land in BrainEvalLog with flag distinguishing tool runs from brain runs.

Lightweight extension mechanism for narrow cognition tasks. New ongoing maintenance — tool eval suites + per-tool model selection discipline. No schema burden — tools don't write tables, so adding/removing a tool has zero schema migration cost.

New tool: TOOL-001_API_Doc_Play_Translator.html. Closes the largest single gap from the v26 architecture eval (Domain 3 was 20% covered).

API doc (OpenAPI spec / markdown URL / raw markdown), product context, current ICP summary, current active plays (for de-duplication), optional buyer persona hint, max plays constraint.

Up to 3 candidate plays with name, hypothesis, target_buyer_persona, api_capabilities_referenced (with doc citations), target_definition outline, suggested_cadence_outline, success_criteria_outline, confidence_self_rating, ungrounded_assumptions (hard requirement — assumptions must be separated from API-grounded claims).

No fabricated capabilities (eval enforces 0% hallucinated capability rate). Refuses if input docs too thin (returns 0 candidates with structured reason rather than fabricate). De-duplicates against current_active_plays.

AGT-901 Pipeline Brain (most common, after product launches), AGT-902 Account Brain (account-specific variant), RevOps direct via workspace UI.

Sonnet (synthesis-heavy; Haiku quality below threshold). 50K input + 5K output budget. ~$0.20-0.30 per call. Monthly cap $300/mo (~1,000 calls). Frequency expectation low.

First mechanism in the OS for translating product capabilities into sales plays. Often chains into TOOL-003 (Sales Play Composer) when a candidate is selected for refinement.

New tool: TOOL-002_Dev_Persona_ICP_Enricher.html. AGT-201's 6-dimension scorer materially under-weights developer-buyer signal; this tool augments without replacing.

Account basics, technical signals (job postings, GitHub org activity, technographic data from AGT-204, product telemetry from UsageMeteringLog), current ICP score breakdown.

dev_persona_score (0–100 composite), dev_persona_tier, buyer_persona_classification (developer / platform_team / compliance_engineer / data_engineer / non_technical_buyer / unclear), signals_present breakdown, key_drivers with citations, agt_201_recommended_adjustment (capped ±15pt on tech_stack and intent dimensions only), data_completeness assessment.

Bounded magnitude (cannot single-handedly flip Tier 3 to Tier 1). Sparse-data graceful degradation (returns "low" or "medium" confidence on sparse inputs, not "high"). Honest "unclear" classification when signals conflict. Hallucinated signals = 0% (key_drivers must cite real input fields).

AGT-201 ICP Scorer (on account update events with technical signals); AGT-901 Pipeline Brain (cross-account dev-persona analysis); AGT-204 Lead Enrichment (as enrichment subroutine).

Haiku (classification task; Sonnet tested no measurable lift). 10K input + 1K output budget. ~$0.01-0.02 per call. Monthly cap $200/mo (~10,000 calls). Frequency expectation high — called on AGT-204 enrichment cadence.

Developer-led GTM motion finally has a system home. AGT-201 keeps existing 6-dimension model; this is parallel signal stream. Backwards-compatible — accounts without enrichment scored exactly as before.

New tool: TOOL-003_Sales_Play_Composer.html. The lever that turns a brain proposal from "we should do something here" into "here's the candidate play, refine and approve."

Hypothesis from calling brain, scope (segment / account_specific), context_views (ICP profile, win/loss pattern, account priority distribution, current active plays, VoC themes), constraints (must_use_existing_lever_only=true, max_touch_count per AGT-302 limits, optional anchor_to_tool_001_output for chain), originating_proposal_id from BrainAnalysisLog.

draft_play (full SalesPlayLibrary draft schema: name, hypothesis, target_definition, suggested_cadence with step_outline, success_criteria with retire_signal), lever_grounding (every play element traces to existing agent spec section), duplication_check (overlap flagging vs current active plays), self_assessed_promotion_likelihood (calibrated honestly, not optimistically), ungrounded_assumptions.

100% lever-grounding required (ungrounded elements omitted, not fabricated). Schema compliance 100% (validates against SalesPlayLibrary draft schema). Must-use-existing-lever-only constraint non-negotiable; cannot propose capabilities AGT-302 doesn't support. Honest self-rated promotion likelihood — "low" rated drafts must actually have lower promotion rate than "high" by >20pp.

AGT-901 Pipeline Brain (most common). AGT-902 Account Brain (for scope=account_specific). Chains from TOOL-001 when API-doc candidate gets selected for refinement. RevOps direct via workspace UI.

Sonnet (composition with structural constraints; Haiku tested but lever-grounding accuracy below threshold). 30K input + 4K output budget. ~$0.15 per call. Monthly cap $300/mo (~2,000 calls).

Tool is stateless — calling brain writes the SalesPlayLibrary draft row with the tool's output. Brain owns the write because brain owns the BrainAnalysisLog row anchoring lineage. Promotion gate (draft → under_review → active) unchanged from SalesPlayLibrary v27 spec.

Closes Domain 1 gap from v26 eval. Brain proposals + TOOL-003 composition + human co-definition + AGT-302 execution = end-to-end "create plays, not just execute" pipeline.

New tool: TOOL-004_Consumption_Forecasting.html. Today the system can detect overage has happened (AGT-503 fires on threshold breach) but cannot predict when overage will happen or distinguish trend from spike.

account_id, sku_id, sku_type, metering_history (90–180 days from UsageMeteringLog brain-ready view, audit_status filtered to verified+pending_recon), contract context (start/end/renewal dates, current period commit, trailing overage count), forecast_horizon_days (default 60, max 180).

forecast_summary (predicted_overage_date with [low, high] confidence interval), trend_characterization (primary_pattern enum, growth_rate, volatility_score, seasonality_detected, anomalies_detected), interpretation_for_caller (is_likely_real_expansion / is_likely_one_time_spike / is_likely_seasonal_recurrence + rationale), data_quality assessment.

5 patterns enumerated: linear (steady growth), exponential (accelerating), seasonal (repeating cycle), cliff (sudden discontinuous step), flat (no trend). Pattern characterization is the most operationally important output — spike-vs-trend differentiation is the eval-targeted dimension at ≥ 80% threshold.

Refuses if < 30 days of history. Seasonality requires ≥ 2 full cycles (eval enforces 0% false-positive seasonality). Confidence intervals scale with history length (no narrow intervals on short data). LLM does not invent forecast numbers — numerical forecasting runs in code (exponential smoothing + regression with seasonality decomposition); LLM produces structured summary + interpretation.

AGT-902 Account Brain (most common — "real expansion or spike?" use case). AGT-503 Expansion Trigger (optional augment to existing 5-signal scoring — downgrades signal weight on cliff/spike pattern). AGT-402 Forecast Adjuster (weights expansion ACV probability in bottoms-up forecast by trend strength).

Haiku (LLM portion is interpretation only; numerical work in code). 15K input + 2K output budget. ~$0.02 per call. Monthly cap $400/mo (~20,000 calls). Highest-frequency tool of the four.

Closes UBB-specific gap from v26 eval. For a usage-based business: the difference between proactive expansion conversation (timed before overage hits) and reactive damage control (overage already painful for customer).

New artifact: Brain_Eval_Harness.html. Established methodology, scoring rubric, run cadence, human review process, drift detection signals, and pre-launch checklist.

7 dimensions. 3 hard criteria (any failure = harness fail): hallucination rate ≤ 2%, staleness recognition = 100%, lever-mapping correctness = 100% (AGT-902 only). 4 soft criteria: source citation rate ≥ 95%, diagnosis accuracy ≥ 0.70 mean, confidence calibration in healthy bands, narrative coherence ≥ 4.0/5.0 mean.

Computed via trigger from aggregated_scores. pass (all hard + all soft) / pass_with_notes (all hard + most soft, no soft > 15pp below threshold) / conditional (all hard but soft significantly below; brain may run shadow mode) / fail (any hard criterion fails). Decision is policy, not negotiation — manual override blocked.

Pre-launch (one time per brain, hard gate) · quarterly (drift detection) · material prompt change (hard gate) · new model promotion (comparison gate) · on-demand (operator suspicion, diagnostic only).

Hard requirement: human reviewer must not be the person who tuned the brain prompt. Trigger validates against an access-controlled prompt-tuner registry at write time. Independence preserves the eval as a measurement, not a target.

~$3/run in tokens (30 brain calls × Sonnet pricing). ~$25/yr token at quarterly cadence + 2–3 ad-hoc runs. ~24 reviewer-hours/year. Total < $50/yr token + ~40 hours analyst time. Cheap insurance.

Documented: catalog over-fitting (eval easy questions); brain "memorization" of eval (rotate question wording, hold private spot-check set); reviewer drift (calibrate periodically with golden-set questions). The eval is an artifact that needs maintenance, not a static check.

Pre-launch gate established. AGT-901 + AGT-902 cannot go to production without passing the harness. New ongoing operational burden — quarterly run + catalog maintenance + reviewer independence enforcement. Eval is the CFO defense — when costs come up, point at measurable accuracy gains. When auditors ask how brain quality is maintained, point at the trend.

New artifact: Brain_Eval_Question_Catalog.html. Scaffolded with 6 detailed question templates and 24 one-line summaries. RevOps owns filling in the remaining detailed templates with real anonymized historical fixtures before pre-launch.

AGT-901 (10): 4 plan diagnosis · 3 coverage gap · 2 quarterly play retrospective · 1 forecast bias attribution. AGT-902 (20): 8 churn diagnosis · 6 expansion qualification · 4 hand-off briefing · 2 QBR narrative. Volume favors AGT-902 because per-account ground truth is easier to construct retrospectively and the brain is higher-volume in production.

5 of 30 questions use deliberately stale Tier 1 source fixtures: EVAL-Q02 (forecast accuracy stale), Q07 (priority score stale), Q12 (composite view stale), Q24 (usage log stale), Q28 (conv intelligence stale during hand-off). All 5 must score 100% on staleness recognition for the harness to pass.

4 properties of a good retrospective question: ground truth exists (knowable in retrospect); source data is reconstructible (Tier 1 tables can be replayed to time-of-question state); multiple defensible answers possible but some clearly better; spans the brain's actual use cases. Hard rule: if source state can't be reconstructed, the question is not eligible — otherwise the eval measures hindsight bias not brain quality.

Mandatory before catalog commit. Account names fictionalized with consistent placeholders (mapping in access-controlled separate file). ACV/consumption rounded within 10%. Personnel names removed/fictionalized. Verified by second reviewer. BrainEvalLog 7-year retention makes anonymization a privacy + contract requirement, not a polish step.

Quarterly: rotate 3–5 questions (retire aged ground truth, add questions sourced from recent operational complaints). On material GTM context shift: refresh to span new ICP / product / segment. On model promotion: spot-check 5 questions on the new model, expand to full harness if divergence material.

Concrete pre-launch artifact — AGT-901 + AGT-902 launch is now blocked on a measurable, defined deliverable rather than vibes. Ongoing maintenance commitment — ~16 hours/year refreshing questions. Catalog is intentionally a scaffold, not finished — structure is set; RevOps fills in the historical detail before the first run.

New schema artifact: BrainEvalLog_Schema.html. Two tables — one row per run (aggregated), many rows per question per run (drill-down).

19 fields. Key: run_trigger, brain_under_test, model_id, system_prompt_hash (joinable to BrainAnalysisLog), brain_view_contracts_hash (detects view drift between runs), question_catalog_version, reviewer_user_id + reviewer_independence_verified (trigger checks against prompt-tuner registry), aggregated_scores (JSONB across 7 dimensions), hard_criteria_passed, decision (computed via trigger), decision_rationale, calibration_notes, token_cost_usd, reviewer_hours_spent.

17 fields per (run, question). Key: brain_analysis_id (FK to actual BrainAnalysisLog row produced for this eval question — full drill-down), uses_stale_fixture flag, citation_rate (auto-scored), hallucinations_detected + total_claims (human-reviewed), staleness_correctly_recognized (NULL if not stale-fixture; T/F otherwise), diagnosis_accuracy_score, lever_mapping_violations (AGT-902 only), confidence_distribution, narrative_coherence_rating (1–5), reviewer_comments.

Reviewer independence trigger: validates reviewer_user_id is not in the prompt-tuner registry at write. Decision logic trigger: computes BrainEvalLog.decision from aggregated_scores via the rubric — manual override blocked. Append-only enforcement: UPDATE blocked except for completed_at; DELETE blocked.

Deliberate. The brain should not learn what the eval expects — the eval is a measurement of general capability, not a target the brain optimizes for. RLS blocks Brain Agent read access to both eval tables. RevOps + auditors + the eval harness runner service are the only readers.

Per-brain quarterly trend; most-failing questions across runs; stale-fixture pass rate audit; reviewer calibration distribution; cost per run; production-vs-eval prompt drift (join on system_prompt_hash). All readable from the two tables without expensive joins.

7 years. Hot 0–24mo, cold 24mo–7yr. Parallels BrainAnalysisLog and UsageMeteringLog — eval results that informed brain-influenced decisions need the same retention floor.

Eval results are durable, queryable, auditable. New triggers to maintain — reviewer-independence trigger requires keeping the prompt-tuner registry accurate. Production-vs-eval drift detection becomes a real query, not a hope — system_prompt_hash joins detect when production is using a prompt that the eval hasn't validated.

New layer L9 — Reasoning. First time the system has agents in the LLM-native sense (distinct from the 40 GTM Services in L1–L8). L9 components are operator-invoked, not cadenced; they read Tier 1 service tables via brain-ready views and produce analysis with full source-trace metadata. No canonical writes from L9 — ever. The architectural commitment from the v26 evaluation made concrete.

Per the v26 architecture eval: brains require trustworthy Tier 1 data. L8 was promoted to Built first (UsageMeteringLog production schema + AGT-804 v26) so L9 reads from a real foundation rather than a stub. With L8 anchored, L9 can be specced without producing fake-confident output.

Color #D4A84C (gold) — chosen to visually distinguish L9 from the eight service layers. Positioned in the layer stack as the topmost layer, reading from L1–L8 below.

Components in L9 are Brain Agents, not GTM Services. The renaming from v26 reserved "Agents" specifically for this layer. The renderer labels L9 cards with "agents" while L0–L8 cards remain "services."

No L9 component runs on cadence by default. Invocation paths: operator query (chat-style), narrative jobs (called by AGT-704 / AGT-603 for narrative sections of MBR/QBR/QBR-prep), explicit batch triggers (quarterly play refresh, AE/CSM rotation hand-off, optional renewal-prep batch over T-90 cohort).

Default model: Sonnet tier. Opus reserved for explicit deep-analysis opt-in. Haiku for narrow Tier 3 tools (forthcoming). Prompt caching enabled. Brain-ready views (10× input compression typical) instead of full table dumps. Per-query budget caps + monthly tier alerts at 75%.

Open-ended operator queries become tractable. "Why is Q3 commercial soft?" no longer requires a 4-day RevOps analyst project. New audit surface. BrainAnalysisLog must demonstrate source-trace integrity to the auditor — a new capability but a known one. Tier 1 services unaffected at execution time. If L9 is offline, L1–L8 continues to run normally.

Step 3 of the four-step build sequence in the v26 architecture plan: deterministic backbone first (L8), framing rename second, Tier 2 agents third, eval harness fourth.

First Brain Agent in L9. Answers cross-functional pipeline questions: why a segment is off plan, what play would close the gap, which existing levers apply. Reads brain-ready views from AGT-702 (MetricsCalc), AGT-401 (Opportunities), AGT-402 (ForecastLog), AGT-201 (Accounts), AGT-206 (AccountPriorityScore), AGT-105 (CapacityPlan), AGT-703 (WinLossLog), AGT-303 (CadenceEventLog), AGT-604 (VoCSynthesisLog).

Plan diagnosis (operator query: "Why is segment X off plan?"); coverage gap proposal; quarterly play refresh; WBR/MBR narrative job (AGT-704 invokes); anomaly explanation after AGT-702 breach alert.

BrainAnalysisLog (own log) and SalesPlayLibrary (drafts only — never active). May write WBR/MBR/QBR narrative sections per AGT-704 v26 charter. Never writes Opportunities, ForecastLog, MetricsCalc, QuotaStore, CompPlans, RevenueRecognitionLog, ABMPlaybook directly.

draft → under_review: human pickup. under_review → active: SLM + RevOps joint approval. Hard volume cap: 3–8 active plays per segment per quarter (configurable, default 5). Cap enforced at write time on transition to active.

Source citation rate ≥ 95%; hallucination rate ≤ 2%; staleness recognition 100% (hard requirement); diagnosis accuracy ≥ 70% on 30-question retrospective harness; play survival rate ≥ 30%; cohort outcome lift tracked quarterly.

Closes Domain 1 gap from v26 eval — system can now create sales plays for human co-definition, not only execute them. No behavioral change to Tier 1 services until a brain-drafted play is promoted to active by humans, at which point AGT-302 executes per its existing v25 spec.

Per-account counterpart to AGT-901. Where AGT-901 reasons across pipeline-wide signals, AGT-902 reasons across the full L4 + L5 + L6 picture for a single account. Reads a composite per-account brain-ready view spanning 11 Tier 1 sources (CustomerHealthLog, ChurnRiskLog, ExpansionLog, UsageMeteringLog, Opportunities, ConvIntelligence, QBRLog, OnboardingLog, TechnicalMilestoneLog, PaymentEventLog, Accounts).

"What's the move?" account queries; AE/CSM hand-off briefing on rotation; renewal risk diagnosis after AGT-502 surfaces a Critical/High account; expansion qualification after AGT-503 fires; QBR prep narrative section (AGT-603 invokes).

All actions map to existing Tier 1 levers — no invented actions. Enum: pull_qbr_forward, open_expansion_play, brief_new_ae_or_csm, customer_communication, escalate_to_slm, recommend_human_query, none. Eval enforces 100% taxonomy compliance.

Within a freshness window, repeat queries on the same account hit cache. Saves cost on iterative meeting-prep workflows where someone asks 3+ follow-ups about the same account.

Conservative: 30 queries/day, 30K input + 3K output per query, Sonnet ≈ $4/day ≈ $120/mo. Per the v26 cost model.

Cross-source per-account synthesis becomes routine. Today this requires the AM/CSM to mentally union 11 source systems. Acceptance rate is a calibration signal, not a quality signal — too high suggests reading leadership/AM bias, too low suggests missing context. Track outcome lift cohort-wise.

New schema artifact: BrainAnalysisLog_Schema.html. The mechanism by which L9 earns audit-grade reasoning out of a non-deterministic substrate.

22 fields. Key: proposal_id (lineage anchor for cohort retrospective), writer_agent_id (single-writer enforced), sources_read (JSONB array with last_refresh_timestamp per source), narrative_output (with inline [src:N] citations), confidence_flags (high_confidence / multi_source / inference / speculation), data_staleness_acknowledged (hard CHECK requires staleness disclosure phrase in narrative when TRUE), cost_usd_estimate (audit-stable).

Staleness disclosure trigger (stale + no disclosure phrase = reject); source citation trigger (every [src:N] must reference a real source_index); action enum trigger (proposed_actions[].action_type must be in taxonomy). Append-only via row-level security — UPDATE and DELETE blocked.

Every brain output gets a stable proposal_id. Downstream actions (promoted plays, executed sequences, scheduled QBRs, sent comms) inherit it. At quarter end, cohort retrospective filters Tier 1 records on WHERE originating_proposal_id IS NOT NULL for the brain-influenced cohort. Pattern-level analysis, not per-deal attribution — per the v26 evaluation's "cohort-level retrospective" model.

7 years. Hot 0–90d, warm 90d–18mo, cold 18mo–7yr. Parallels UsageMeteringLog and RevenueRecognitionLog.

Brain Agents read own writes only. RevOps full read. Auditors full read on audit windows. AM/CSM/AE: own queries + assigned accounts. Tier 1 services intentionally have no read access — preserves the auditability invariant that canonical metrics never depend on brain outputs.

Audit reproducibility — given any row, auditor can re-fetch cited sources at cited timestamps and re-evaluate defensibility. New trigger maintenance burden — staleness disclosure trigger and source citation trigger must be kept aligned with prompt evolution. Eval ground truth — the 30-question retrospective harness reads BrainAnalysisLog directly.

New schema artifact: SalesPlayLibrary_Schema.html. Operationalizes the v26 commitment: brains propose, RevOps and sales leaders refine into a small curated set, AGT-302 executes.

4 states with enforced transitions. draft (Brain writes) → under_review (human pickup) → active (SLM + RevOps joint approval) → retired (terminal). No backward transitions. Trigger validates allowed transitions only.

Hard cap enforced at trigger level on transition to active. Default 3–8 per segment per quarter (configurable per segment via SalesPlayLibraryConfig in L0; default 5). Account-specific plays: 1 active per account, hard. Cap met → activation rejected with error suggesting retire-an-existing-play first. No silent overflow.

AGT-302 reads SalesPlayLibrary WHERE state = 'active' at sequence-generation time. Drafts and under-review plays invisible to AGT-302 by design. AGT-302 writes play_id and originating_proposal_id to CadenceEventLog for every event generated under a brain-co-designed play. v27 ripple change to AGT-302's schema (two new fields).

Three signals: promotion rate (active / draft ratio per quarter per brain — eval target ≥ 30%), edits-during-review volume (light edits = brain calibrated; heavy edits = drafts are starting points), retrospective outcomes vs. success_criteria (cohort lift quarterly). Brain reads retrospective_outcomes in next play-refresh batch.

Closes Domain 1 "create plays, not just execute" gap from the v26 evaluation. RevOps + SLM workflow change — explicit promotion gate process, joint approval, quarterly retrospective on cohort outcomes. AGT-302 v25 behavior preserved — touch caps, suppression, ABM backstop still apply on top of any active play execution.

Six Tier 1 services gain a nullable originating_proposal_id column. Always nullable — actions taken without brain involvement carry NULL. Cohort retrospective queries filter on WHERE originating_proposal_id IS NOT NULL for the brain-influenced cohort. AGT-302 also gains a play_id column for direct play-to-event lineage.

AGT-203 ABMPlaybook (play promoted from SalesPlayLibrary carries proposal_id) · AGT-302 CadenceEventLog (sequences executing under brain-co-designed plays carry proposal_id + play_id) · AGT-503 ExpansionLog (expansion plays initiated via brain proposal) · AGT-603 QBRLog (pull-forwards triggered by brain proposal) · AGT-405 OpportunityBriefLog (briefs from brain-recommended meetings) · AGT-504 CommBriefLog (comms approved after brain recommendation).

Fully backwards-compatible. Existing rows have NULL originating_proposal_id — they predate L9 and were not brain-influenced. Existing service behavior unchanged. New rows populate the field only when triggered by a brain-promoted play or brain-proposed action.

Per the v26 evaluation: in enterprise sales-led motion you cannot prove the brain caused a deal loss; what you can do is compare cohort outcomes between brain-influenced and non-brain-influenced actions over a quarter. The originating_proposal_id column is the load-bearing piece of that measurement.

Six service schemas updated. Migration is straightforward (nullable add). No reads change in those services. Brain output never affects how Tier 1 services compute, only how RevOps reports retrospective outcomes.

AGT-603 charter extended (parallels the AGT-704 v26 charter extension): AGT-902 may write per-account narrative sections of QBR prep documents. AGT-603 still owns the artifact, gates publication, and pulls fresh from canonical Tier 1 metric tables for any number-bearing section.

AGT-902 may write narrative content describing account history, relationship dynamics, recent ConvIntelligence trajectory, open commitments — into designated QBRLog narrative fields. Always with source-trace metadata.

AGT-902 may never write metric sections of QBR prep — health score, churn risk tier, expansion ACV, usage trends. Those pull fresh from AGT-501 / AGT-502 / AGT-503 / AGT-804 at QBR generation time. Same architectural commitment as AGT-704.

QBR prep narrative quality lift — synthesis across 11 sources is now automatic for AM/CSM consumption. Metric integrity preserved — every number in QBR still traces deterministically to Tier 1 service tables.

AGT-804 promoted from Specced (v25) to Built (v26). Spec hardened with two new sections: Production controls & audit (8 controls covering reconciliation, deferred-revenue tie-out, MAX(record_version) discipline, prior-period adjustment flagging, milestone sign-off gating, 7-year retention, restatement workflow) and Edge cases & failure modes (8 scenarios: reconciliation failure, prior-period correction, mid-term amendment, customer dispute, SKU type change, mid-term churn, source replay, ingestion outage). UsageMeteringLog section rewritten to reference the new production schema document instead of duplicating field-level detail.

The architecture evaluation (May 2026) called L8 the gating dependency for any Tier 2 / brain layer. The brain reading from a stub UsageMeteringLog produces fake-confident output — the worst possible failure mode for audit-sensitive metrics like Magic Number, NRR, and CAC Payback. Building reasoning on top of an unbuilt L8 is incoherent. AGT-804 goes first within L8 because it is the audit-critical one: ASC 606 recognized revenue feeds AGT-702's MetricsCalcLog, which feeds AGT-704's MBR/QBR.

Reconciliation against product DB (monthly tolerance ±0.01%); recognition method immutable within a period; deferred revenue ties to balance sheet within ±$1; MAX(record_version) discipline enforced in agent code; prior-period adjustments flagged distinctly so AGT-702 can sum them separately; milestone billing requires customer_sign_off = TRUE (IE-only completion does not unlock recognition); 7-year retention via cold storage tier; restatement workflow defined for auditor-driven closed-period changes.

If a closed-period figure must be restated, AGT-804 emits a RevenueRecognitionLog row with restatement_flag = TRUE and reason. AGT-702 surfaces restatements as a separate line in MBR/QBR rather than blending them into current-period activity. Preserves auditability + avoids polluting trend lines.

If UsageMeteringLog ingestion is offline >24h during a close period, AGT-804 holds the close pending data catch-up. No partial-data recognition. Cascades to AGT-702 staleness and AGT-704 MBR staleness gate per existing AGT-704 spec.

No schema additions to RevenueRecognitionLog or DeferredRevenueSchedule beyond the v25 spec. Two existing fields gain explicit semantics: prior_period_adjustment (BOOL, set TRUE on rows derived from UsageMeteringLog corrections to closed periods) and restatement_flag (BOOL, set TRUE on auditor-driven closed-period restatements). UsageMeteringLog itself is now governed by a separate production schema document — see related schema entry below.

First L8 agent in Built status. Magic Number, NRR, and CAC Payback now have a fully specced canonical source. AGT-704 staleness cascade gains a new branch — UsageMeteringLog reconciliation failure now blocks AGT-702 metric recompute, which blocks AGT-704 MBR/QBR. No behavioral changes for AGT-501 / AGT-503 / AGT-802 / AGT-402 — they continue to read UsageMeteringLog as before; the production schema doc formalizes contracts already in use.

Architecture evaluation document (May 2026) explicitly directed L8 buildout before adding any reasoning/brain layer. Step 1 of the four-step build sequence in the architecture plan.

New schema artifact: UsageMeteringLog_Production_Schema.html. Standalone document because the table is consumed by four agents (AGT-804 / AGT-802 / AGT-503 / AGT-501) for materially different decisions (recognition vs. billing vs. expansion vs. health), making data quality on this one table the highest-leverage data-quality investment in L8. Embedding the production contract inside any single agent spec would obscure the cross-cutting nature of the dependency.

26 fields with types, nullability, constraints. Natural key is (account_id, sku_id, period_start, period_end, source_system, ingest_event_id, record_version). Surrogate PK is usage_id. New fields beyond the v25 spec: sku_type (consumption/seat/hybrid), period_granularity, unit_price_usd + overage_unit_price_usd (captured for audit reproducibility), seat_utilization_pct (computed), source_system, ingest_event_id, ingest_batch_id, received_at, effective_at, record_version, prior_version_id, correction_reason, audit_status.

UPSERT on (source_system, ingest_event_id). Source replay is a no-op. Corrections insert new rows with record_version + 1 and prior_version_id pointing to the predecessor; original rows are preserved. Consumers always filter to MAX(record_version) per natural key. UsageMeteringLog_current view exposes this as the default read path.

9 rules at ingestion. HARD failures (reject to DLQ): negative units_consumed, inverted period bounds, future-dated period_end, missing ingest_event_id, FK orphan, out-of-range seat_utilization_pct, non-monotonic record_version. SOFT warnings (accept + emit): late arrival >24h, effective_at drift, anomaly >10x trailing 30d median.

Daily ±0.1%, monthly ±0.01% (audit-grade), quarterly ±0.001%. Monthly failure holds AGT-804 close. audit_status state machine: pending_recon → verified on pass; transitions to disputed on customer challenge. Verified rows feed all consumers; disputed rows excluded from recognition.

Hot 0–13mo (operational), warm 13–36mo (reporting), cold 36mo–7yr (audit-only restored on demand), purge after 7yr with legal-hold check. ASC 606 + tax retention requirement. Tier transitions move bytes, never modify values; record_version chains preserved across tiers.

Source emits to ingestion queue (Kafka/Kinesis/SQS) → ingestion service deduplicates + validates + UPSERTs → threshold publisher emits to UsageThresholdEvents for AGT-503 real-time subscription → reconciliation jobs run daily/monthly/quarterly → nightly tier-transition job. DLQ for HARD failures with >1h-depth pager.

10 gating items including: source system instrumentation, ingestion service deployment, DLQ + monitoring, UsageThresholdEvents end-to-end test, AGT-804 reading real (non-synthetic) data for one full close, daily reconciliation passing 7 consecutive days, monthly reconciliation within tolerance 3 consecutive months, external auditor sign-off (SOC 2 / ASC 606), 7-year retention path tested with sample restore, runbook reviewed by oncall.

L8's data foundation is now formally specified. Brain layer (planned Tier 2) can be built against a known contract rather than against ambient assumptions. Engineering work to make the schema operational is non-trivial — the document is the production target, not a deployed table. Going forward, every L8 agent build references this schema for usage data; no agent queries the product system directly. No behavioral changes to existing agents — the production schema formalizes contracts already in use; field additions (source_system, record_version, etc.) are append-only against the v25 schema as documented in GTM_OS_Schema_v25.xlsx.

Joint: Finance + Data Engineering. Single-writer is the ingestion service. AGT-804 / AGT-802 / AGT-503 / AGT-501 are read-only consumers.

Net new agent. Generates legally deployable quota plan documents from approved AGT-101/102 outputs. Execution-only — never sets quota or modifies plan terms. Three triggers: AGT-102 final approval (annual), new rep hire (onboarding), AGT-104 exception path (mid-year amendment). Three output formats per generation: URL-based HTML for rep view + e-signature, PDF generated at signature time as legal record, structured JSON payload for downstream system consumption.

Quota targets (annual + quarterly + monthly where applicable), payout curve, kickers + accelerators, draw terms (section omitted if no draw), territory/named account list, plan governance summary. OTE split excluded (governed by offer letter/HR). SPIFs excluded (separate short-term documents).

Plan not legally active until e-signature received. AGT-103 attainment computes but AGT-104 holds payout pending signature. Signature overdue (5 business days) → manager + HR notified, flagged in CompAuditLog. Integration: v1 hook (DocuSign/equivalent in live deployment; manual HR fallback in pre-integration).

Agent recommends addendum (quota-only changes) vs full reissue (structural term changes: new kickers, territory reassignment, draw terms, payout curve change). RevOps makes final selection before generation begins — agent never auto-selects. Override rationale required if RevOps deviates from recommendation. All selections logged in QuotaPlanDocLog.

AGT-104 plan approval gate now reads QuotaPlanDocLog.signature_received. Reps with pending_signature held individually — not the full cohort. Signature completeness audit added to AGT-104's 12 controls.

QuotaPlanDocLog — one row per document generation event. 21 fields: doc_id, rep_id, plan_id, quota_store_id, trigger_type, doc_type, doc_version, doc_status, html_url, structured_payload, pdf_url, generated_at, signature_received, signed_at, signature_overdue_notified, superseded_by_doc_id, amendment_recommendation, amendment_selection, amendment_selection_by, amendment_override_rationale, exception_log_id.

Planned at L1 original build (v11–v16). Specced now as the last L1 agent. No earlier sessions had the full AGT-102/104 design locked to spec against.

Closes the legal deployment gap in L1 — quota and comp plan design were fully specced but the mechanism for producing a legally binding rep-facing document was absent. AGT-104 signature completeness audit is a new control — existing comp governance now covers the full lifecycle from quota approval through signed plan delivery. AGT-103 payout behavior unchanged — attainment computes regardless of signature status; AGT-104 controls the payout hold, not AGT-103.

Net new agent. Hybrid top-down/bottoms-up quota methodology. eRep ramp model (0.25→0.50→0.75→1.0). Nine guardrails. Four-gate approval chain (SLM→RevOps→CRO/FP&A/HR→publish). Segment-differentiated quotas (AE: SMB $400K, MM $800K, Ent $1.6M).

QuotaStore (625 rows), RampSchedules, AccountQuotaTargets, QuotaAmendmentLog

Layer 1 original build.

Foundational. All downstream productivity ratios (AGT-105, AGT-205, AGT-702) derive quota targets from QuotaStore.

Net new agent. Translates AGT-101 quota outputs into legally deployable comp plans. OTE splits, gate/multiplier curves, quadratic usage curve, kickers, SPIFs, recoverable/non-recoverable draw. Soft approval gate.

CompPlans extended (cols 22–31), TIPolicy, TIBudgetModel, Kickers, SPIFs

Layer 1 original build.

Foundational. TIBudgetModel feeds AGT-105 loaded HC cost model. TIPolicy feeds AGT-104 TI% variance control.

Net new agent. Daily batch payout calculation. QuotaRetirementLedger as transaction layer (one row per deal/usage event). DrawLedger for recoverable draw tracking. AttainmentSnapshot deliberately excluded — intra-period attainment is a display problem.

QuotaRetirementLedger, PayoutHistory, DrawLedger

Layer 1 original build.

Foundational. PayoutHistory feeds AGT-102 TI budget calibration (Option B), AGT-105 TI spend in loaded HC model, AGT-702 GTM Health Monitor S&M spend.

Net new agent. 12 controls: payout accuracy, draw integrity, kicker stacking, quota amendment audit, TI% variance, ROE violations, plan approval gate, SOX chain, duplicate detection, clawback triggers, usage gate audit, policy exception log. Clawback writes DrawLedger recovery rows.

CompAuditLog, ExceptionLog. DrawLedger extended: new draw_type = clawback_recovery.

Layer 1 original build.

Net new agent. Three productivity layers (eRep, fully loaded GTM HC, GTM + Marketing HC). Magic Number, R40, and CAC Payback guardrails run independently and reconcile to a single hiring recommendation. FP&APlan as top-down constraint source.

FP&APlan, CapacityPlan, HiringPlan, AttritionLog, MarketingHC

Layer 1 original build.

FP&APlan is foundational for Layers 4 and 7. AGT-402 and AGT-404 reference revenue targets. AGT-702 references metric targets. AGT-205 TAM/SAM references plan for feasibility check.

Net new agent. Role-differentiated carve: AM/CSM/Enterprise AE on named account lists; SMB/MM AE on rules-based routing (segment × geo × vertical). routing_eligible flag uses behavioral capacity signals (strict AND logic).

TerritoryDefinitions, TerritoryAssignmentLog

Layer 1 original build.

TerritoryDefinitions is the routing authority for AGT-202 Lead Router v2. routing_eligible gate is used by AGT-202 and feeds AGT-206 dormancy flags back to AGT-106.

Segment bands updated: SMB 5–200, Mid-Market 201–2,000, Enterprise 2,001+. Dimension 2 vertical taxonomy aligned to TerritoryDefinitions (FinTech, HealthTech, SaaS, RetailTech, HR Tech, EdTech, Cybersecurity, Logistics). Segment removed as a scoring dimension — confirmed as routing input only.

Original bands predated segment redefinition decision. Vertical alignment needed so ICP scoring and territory routing reference the same 8-vertical taxonomy.

Scoring recalibration expected on existing accounts. Accounts near tier boundaries may shift T1↔T2. AGT-202 routing tier SLAs unaffected (T1=2hr, T2=24hr remain).

AGT-106 Territory Design (v14) introduced TerritoryDefinitions as the authoritative account ownership source. AGT-202 v1 was built before this existed and used RepBook instead. SDR motion decision: SDRs are purely outbound — PLG accounts should never hit a cold outreach sequence.

Leads extended: gtm_motion_at_routing, routing_rule_applied, named_account_hold, vertical_match, workload_flag, first_contact_timestamp, tier_change_post_contact

Routing behavior change. PLG leads previously hitting SDR sequences now route directly to AE. Accounts with named rep ineligible now hold rather than fall back to segment manager immediately.

Net new agent. Target account selection only (not outreach execution). T1 accounts only. Enterprise: 1:1 and 1:few eligible. MM: 1:few only. SMB: SLM nomination only. Agent proposes → rep confirms 1:1 → FLM confirms 1:few → SLM approves.

TargetAccountList, ABMPlaybook, ABMEngagementLog. Accounts extended: abm_active, abm_tier

ABM status creates downstream behavior changes in 3 agents: AGT-202 named account lock (no fallback routing), AGT-301 deeper research pass, AGT-304 nurture suppression.

Net new agent. Pre-score enrichment (30s timeout gate) + signal-differentiated periodic re-enrichment. Intent/news: daily (T1). Contact: monthly. Firmographic/technographic: quarterly. Confidence-weighted overwrite. ICP auto-recalculates on scored field update. 30s timeout: AGT-201 scores with available data, enrichment completes async.

EnrichmentLog, EnrichmentQueue, Contacts (net new cross-cutting table). Accounts extended: 15 enrichment fields including intent_score, funding_amount_usd, crm_vendor, enrichment date fields.

Unblocks AGT-201 Dimensions 4+5 (25 pts previously data_gap). crm_vendor and technographic fields feed AGT-403 competitor pre-population at deal open (see v17→v18 AGT-403 ripple). Contacts table feeds AGT-305, AGT-401, AGT-405 stakeholder mapping.

Net new agent. TAM/SAM/SOM × segment × vertical × geo × product family (288 cells). Hybrid methodology. FP&APlan feasibility check on every run. Whitespace flags to AGT-203. SOM feeds AGT-105 "what would have to be true" scenarios.

MarketAssumptions (RevOps-owned, versioned), MarketSizeModel

Feasibility check creates a live validation against FP&APlan targets. If FP&APlan Y2 target exceeds SOM, AGT-205 flags the gap — first agent to challenge the financial plan from a market sizing perspective.

Net new agent. 3-dimension model: fit (40%), intent (40%), opportunity (20%). Intent decay: score × e^(-hours/72). Event-triggered re-rank (intent spike, stage change, enrichment). Semi-annual full-universe scoring feeds AGT-106 territory rebalancing. Pin-only rep override with 30-day expiry.

AccountPriorityScore, PriorityWeightsConfig

Priority score feeds AGT-301 queue order (highest priority sequences drafted first). Semi-annual scores feed AGT-106 territory rebalancing. Dormancy flags identify accounts for removal from rep books.

When Accounts.abm_active = TRUE: agent runs a deeper enrichment pull (all 6 enrichment categories + recent news/triggers) before drafting each step. ABM sequence template and step cadence unchanged — personalization depth only.

AccountPriorityScore.composite_score determines draft order when multiple sequences are queued. Highest score drafted first. No effect on step timing, cadence structure, or sequence content.

AGT-203 ABM Target Selection (Layer 2) added abm_active and abm_tier to Accounts. AGT-206 Account Prioritization (Layer 2) added composite_score as a rep-facing priority signal that should influence agent work queue.

None. No new tables or fields. Prompt-only delta.

First check before any nurture enrollment: IF Accounts.abm_active = TRUE → exclude from nurture. Exclusion logged in NurtureRecommendations with reason = 'ABM active suppression'. Account re-enters nurture eligibility when abm_active resets to FALSE on TAL exit.

AGT-203 ABM Target Selection creates bespoke outreach plays for T1 accounts. Generic nurture cadences running in parallel would create channel conflict and undermine the ABM play quality.

None. Prompt-only delta. NurtureRecommendations already has a reason field.

Behavior change. Any account currently in nurture that gets added to TAL will be suppressed from nurture on next enrollment check. Existing nurture sequences in-flight are not retroactively stopped — they complete and re-enrollment is blocked.

Net new agent. Triggered by meeting_booked from outreach cadence (no deal exists). 7-section brief: account overview, pain hypothesis, discovery questions, stakeholder context, talk track, competitive context, ABM playbook alignment. Draft on booking + 4am rep local time refresh on meeting day. AGT-407 scores adherence + hypothesis accuracy post-call. Option 4 feedback loop activates when ≥30 scored calls per vertical.

DiscoveryBriefLog (all 7 sections + rep edits + nullable post-call score fields)

Creates DiscoveryBriefLog as a new data source for AGT-407 post-call scoring. Option 4 feedback loop will improve pain hypothesis quality per vertical over time — closes a learning loop that currently doesn't exist in the OS.

AGT-306 Rep Daily Briefing Agent removed from GTM OS agent registry. Daily briefing is a personal workflow preference, not a system-level function.

Reps have different briefing preferences. A system-mandated daily brief creates maintenance overhead and doesn't account for role variation (AE briefing vs CSM briefing vs SDR briefing are materially different). Better served by rep-built personal agents.

Rep Agent Builder Guide published — documents which GTM OS tables and agents reps can reference when building personal agents. Includes 4 example prompts (AE, CSM/AM, single-account deep dive, intent spike alert).

AGT-301 now writes CadenceEventLog.sequence_type at sequence draft time. Derived from Accounts.abm_active + abm_tier: abm_bespoke (1:1 ABM), abm_cluster (1:few ABM), nurture (marketing nurture), standard (all other). Classification label only — no behavior change to sequence content, timing, or channel mix.

AGT-301 now writes CadenceEventLog.priority_score_at_queue at queue registration time. Snapshotted from AccountPriorityScore.composite_score at the moment of queue — does not update as score changes during the sequence. NULL if no AccountPriorityScore record exists for the account. Enables AGT-303 to analyze whether high-priority accounts received timely sequence execution relative to their score.

ABM deeper research pass and priority queue order were wired in v18 (behavioral), but neither was traceable in CadenceEventLog. Without these fields, AGT-303 cannot distinguish ABM sequences from standard sequences or correlate priority score to sequence timing in analysis. Schema fields close the traceability gap.

CadenceEventLog extended: sequence_type (enum), priority_score_at_queue (float, nullable)

AGT-303 can now segment cadence performance by sequence_type — ABM vs standard sequence efficacy becomes measurable for the first time. Priority score correlation to sequence timing enables AGT-303 to flag when high-priority accounts are queued but executed late relative to their score.

AGT-302 adds an explicit ABM nurture backstop as Gate 2 in its coordination sequence (after compliance check, before touch cap). If sequence_type = 'nurture' AND Accounts.abm_active = TRUE: block activation, write suppressed_reason = 'abm_active_suppression' to CadenceEventLog, notify originating agent. AGT-302 also writes suppressed_reason for all other gate blocks: touch_cap_exceeded, compliance_block, conflict_resolved_queued.

AGT-304 is the primary ABM suppression gate for nurture enrollment. But AGT-302 previously had no visibility into whether future routing paths might bypass AGT-304 (e.g. a marketing automation trigger). The backstop ensures the OS-level constraint holds regardless of origination path. The suppressed_reason field provides AGT-303 with a unified suppression audit trail across all gate types and sequence origins.

CadenceEventLog extended: suppressed_reason (string, nullable). Written exclusively by AGT-302. AGT-304 continues to log suppression in NurtureRecommendations — AGT-302 writes to CadenceEventLog so suppression is visible in the unified cadence audit trail.

No behavior change for accounts not flagged abm_active — existing conflict resolution and touch cap logic unchanged. AGT-303 gains a suppression audit trail — can now report on how often ABM suppression fires, touch cap holds, and conflict resolution blocks occur across the cadence universe.

FP&APlan versioning (v21) introduced plan_version, reforecast_as_of_date, and is_board_plan. AGT-402's pct_of_target join is now explicitly specced: reads FP&APlan WHERE is_board_plan = FALSE ORDER BY reforecast_as_of_date DESC LIMIT 1 — the current operating plan, not the board plan. Board plan displayed as secondary reference only.

Expansion ACV from ExpansionLog.acv_estimate (AGT-503) added as a third forecast component alongside new logo and renewal pipeline. Weighted by churn risk tier: Low = 80%, Medium = 50%, High/Critical = 0%. Forecast output adds an expansion revenue line to the breakdown. pct_of_target denominator uses the full FP&APlan revenue target including planned expansion.

None. Both FP&APlan versioning fields and ExpansionLog already exist. Prompt-only delta.

Expansion revenue is now visible in the bottoms-up forecast — NRR upside is quantified in the forward revenue view for the first time. FP&APlan version join prevents the operating plan from being compared against the original board plan mid-year when reforecast has occurred.

Two new CompetitiveKnowledgeBase feed sources: (1) VoCSynthesisLog (AGT-604) — monthly synthesis surfaces competitor themes from CSM calls and QBR outcomes. Post-sale competitive signal tagged signal_source = 'voc'. Read on each monthly AGT-604 synthesis run. (2) CalibrationSignalLog (AGT-703) — quarterly win-loss competitive dimension provides outcome-verified win/loss rates by competitor × segment × vertical. Tagged signal_source = 'win_loss'. win_loss entries take precedence over call_mention entries when signals conflict on a competitor's strength/weakness.

Original KB only had pre-deal and mid-deal signals (technographic enrichment + live call mentions). Post-sale and outcome-verified signals are higher quality for displacement angle generation — a competitor that wins deals against us is more important to track than one that's merely mentioned.

None. VoCSynthesisLog and CalibrationSignalLog already exist. Prompt-only delta adding two read sources to KB maintenance logic.

CompetitiveKnowledgeBase becomes a full-lifecycle competitive signal repository — pre-deal (technographic), mid-deal (calls), post-sale (VoC), and outcome-verified (win-loss). Displacement angles generated by AGT-403 for AGT-301/305/405 are now informed by actual win/loss outcomes, not just mentions.

Same FP&APlan versioning fix as AGT-402. AGT-404's feasibility comparison now joins on the current operating plan version (non-board, most recent reforecast). Board plan displayed as a secondary reference line. Every comparison output line now carries a plan_version label (e.g. "3+9") so the reader knows which plan iteration the % refers to — critical when operating plan has been revised mid-year.

None. Prompt-only delta.

MutualActionPlan (MAP) was introduced in v20 with AGT-601. AGT-405's deal terms section (Proposal/Negotiation/Close only) now reads MutualActionPlan where opp_id matches. Section 6 gains: MAP status (draft/customer_review/signed/not_applicable), open MAP action items and owners, outstanding rep commitments from MAP. If MAP not started for an eligible deal: brief flags "MAP not started — required before close." Nullable — omitted without error if no MAP record exists.

MAP is created pre-close for Enterprise/high-ACV deals. A late-stage meeting brief that doesn't reference MAP commitments leaves the rep without visibility into what the selling team has formally committed to. Open MAP items are often the decisive close blockers.

None. MutualActionPlan already exists. Prompt-only delta on Section 6 of AGT-405 brief generation logic.

Late-stage meeting briefs now surface MAP status and open commitments — rep enters close meetings with full visibility into what's been promised and what's still open. No impact on early-stage briefs — MAP is only read when deal is at Proposal or later.

AGT-501 Customer Health Monitor uses a ConvIntelligence adjustment from CSM-owned calls specifically — not all sales conversations. Without a call_owner_role field, AGT-501 cannot distinguish CSM calls from AE/SDR/AM/SE calls. AGT-407 now writes call_owner_role (enum: AE / SDR / AM / CSM / SE) on every ConvIntelligence record, derived from rep role in RepActivity or TerritoryDefinitions. Existing rows carry NULL until agent is re-run.

AGT-501's ConvIntelligence adjustment was specced in v19 but the filtering mechanism was never created. CSM call sentiment feeding into a customer health score is materially different from AE call sentiment on a deal — conflating them would corrupt the health signal. call_owner_role is the filter key that makes this distinction possible.

ConvIntelligence extended: call_owner_role (enum: AE / SDR / AM / CSM / SE, nullable on existing rows)

Unblocks AGT-501 ConvIntelligence adjustment — CSM call signals can now be correctly filtered and applied to customer health scoring. Without this field, AGT-501 had to skip the ConvIntelligence dimension or apply it incorrectly using all rep types.

stakeholder_engagement_breadth dimension now reads from Contacts table (which persona types are engaged: economic buyer, champion, technical evaluator) rather than raw RepActivity contact count. More precise signal — persona coverage matters more than raw contact count.

When Accounts.abm_active = TRUE: minimum score for Green threshold raises from 75 → 80. Economic buyer engagement becomes a required dimension — score 0 on this dimension if economic buyer is not engaged, regardless of other scores.

AGT-204 Lead Enrichment created the Contacts table (v17) with persona-typed records. AGT-203 ABM status signals a higher-scrutiny deal that warrants stricter health standards.

None. Reads from existing Contacts table. Prompt-only delta.

ABM deals will score harder to reach Green. Expected and intentional — ABM accounts warrant higher scrutiny. Non-ABM deals unaffected.

Forecast output adds pct_of_target field: AI forecast ÷ FP&APlan.revenue_target_usd by segment. Shows where each segment's pipeline trajectory sits relative to annual plan — not just absolute numbers.

FP&APlan table (introduced in v16 with AGT-105) contains annual revenue targets by segment. The bottoms-up forecast gains meaning when expressed as % of target rather than just a dollar figure in isolation.

None. FP&APlan already exists. Prompt-only delta adding one derived field to forecast output.

When an opportunity opens, AGT-403 checks Accounts.crm_vendor and technographic enrichment fields against CompetitiveKnowledgeBase. If a competitor is detected: Opportunities.competitor_detected = TRUE pre-set at deal creation. Previously this only fired when a rep manually flagged a competitor mid-deal.

AGT-204 Lead Enrichment (v17) added crm_vendor and technographic stack fields to Accounts. This data makes it possible to know the competitive landscape before the deal starts — rep shouldn't have to discover this themselves in a discovery call.

None. Reads from existing Accounts.crm_vendor. Prompt-only delta.

Competitive context available from deal open, not mid-deal. Reps enter discovery knowing the competitive landscape. CompetitiveKnowledgeBase can be pre-loaded with displacement angles before the first call.

Top-down forecast output adds a FP&APlan comparison line: "Current pipeline trajectory vs annual plan target" by segment. Shows whether the top-down model is on track to hit plan, and by how much it is over or under.

FP&APlan table (v16) contains annual and 3-year revenue targets. The top-down forecast produces a revenue projection — expressing it as % of plan target closes the loop between the financial model and the sales forecast.

None. FP&APlan already exists. Prompt-only delta.

Net new agent. 4 triggers: stage advance, meeting scheduled, on-demand, deal health drop. 7 sections including deal terms (Proposal/Negotiation/Close only). Standard format for early stages, differentiated for late stages. Draft on trigger + 4am refresh on meeting day. Rep + internal attendees + manager (Amber/Red) receive brief. AGT-407 scores objective completion and deal advance quality.

OpportunityBriefLog (deal-active counterpart to DiscoveryBriefLog)

AGT-305 scores hypothesis accuracy (did pain prediction match reality?). AGT-405 scores deal advance quality (did deal move stage, did rep execute objectives?). Different feedback loops on the same scoring infrastructure.

Health drop trigger creates automatic recovery prep — manager notified immediately. AGT-407 deal advance scoring creates a new coaching signal for AGT-701: which reps execute stage objectives vs which advance deals without meeting stated objectives.

Net new agent. Analysis layer — transcription handled by Gong/Zoom/Chorus via webhook (hook in v1). 5 dimensions: sentiment, next step extraction, competitor detection, objection pattern extraction, email analysis. All conversations in scope regardless of deal stage. Per-conversation processing on transcript receipt.

ConvIntelligence extended with AGT-407 analysis fields: sentiment scores, next_steps (JSON), competitors_mentioned, objections_raised, conv_intelligence_score (0–3 for AGT-401), post_call_summary.

Highest-impact new agent in Layer 4. Unblocks the full deal health score and activates AGT-305's feedback loop. Every downstream agent that reads ConvIntelligence gains richer signal.

Layer reorder: L0→L1→L2→L3→L4→L5 (post-sale retention, was L6)→L6 (customer lifecycle, was L7)→L7 (measurement & feedback, was L5). Agent numbering updated accordingly: AGT-501–504, AGT-601–603, AGT-701–703.

Measurement is a feedback layer that reads from all other layers and closes the loop back to L1. Placing it before post-sale layers was architecturally awkward — it implied measurement happens before customer management. Post-sale layers are sequential; measurement is horizontal.

Canonical agent registry established. Every agent has: unique number (AGT-XXX), human-readable name, purpose, layer, status. Layer-prefix convention: 1xx=L1, 2xx=L2, 3xx=L3, 4xx=L4, 5xx=L5, 6xx=L6, 7xx=L7, 8xx=L8. Text-only IDs (AGT-ABM, AGT-ENR, etc.) replaced with numbered equivalents.

Mixed naming convention (some numbered, some text) made cross-references ambiguous and prevented self-documenting dependency tracking. Numbered IDs with human names provide both machine-readable precision and human-readable context.

L8 Revenue Operations defined as a distinct module within the OS framework (same agent/schema architecture) but Finance-owned. GTM OS stops at signed quote. L8 covers order → invoice → payment → revenue recognition. CPQ (AGT-406) lives in L4 — it's used during active deals. L8 starts at AGT-801 Order Management, triggered by signed quote output from AGT-406.

Invoicing and billing interact with billing infrastructure (Stripe, Zuora, NetSuite) rather than CRM/GTM tooling. Finance owns billing; GTM owns L1–L7. Clear ownership boundary reduces governance ambiguity while maintaining schema coherence for integration points (usage data→AGT-503, revenue recognition→AGT-702).

Net new agent. Deal-active counterpart to AGT-305. 4 triggers: stage advance, meeting scheduled on opportunity, rep on-demand, deal health drop. 7 sections with stage-differentiated format for Proposal/Negotiation/Close (deal terms section added). Draft on trigger + 4am refresh on meeting day. Recipients: rep + internal calendar attendees + manager when Amber/Red. AGT-407 scores objective completion and deal advance quality post-meeting.

OpportunityBriefLog — deal-active counterpart to DiscoveryBriefLog. Includes nullable post-meeting scoring fields: objective_completion_rate, deal_advanced, deal_advance_quality_score, commitment_follow_through_score.

AGT-305 scores hypothesis accuracy (did pain prediction match reality?). AGT-405 scores deal advance quality (did deal move stage, did rep execute stated objectives?). Different feedback loops on the same scoring infrastructure.

Layer 4 spec session — net new agent.

Health drop trigger creates automatic recovery prep — manager notified immediately regardless of normal band threshold. AGT-407 deal advance scoring creates a new coaching signal for AGT-701: which reps execute stage objectives vs which advance deals without meeting stated objectives.

Net new agent. Supports 4 pricing models: seat license, consumption commitment, pay-as-you-go, professional services. Seat + consumption can coexist as separate line items, not bundled. Agent proposes configuration, rep explicitly confirms before quote is generated. Quote is URL-based HTML — not a PDF. URL in pending state until all approvals clear. 4-tier approval chain (rep → manager → deal desk → CRO/VP) auto-routed based on deal configuration vs PricingConfig thresholds. Per-viewer trackable links for email delivery. Forwarding signal detection (new external IP = possible procurement loop-in).

Quote is a pricing proposal, not a legal document. Acceptance triggers AGT-801 (L8) which generates the formal order form from quote data. GTM OS boundary is at accepted quote. Billing contact on the OF (built by L8) is the target for AGT-504 Customer Communications on future price change events.

PricingConfig (L0 — Finance + RevOps owned, rate cards + discount thresholds), QuoteLog (state machine + line items), QuoteApprovalLog (per-tier approval audit trail), QuoteViewLog (per-viewer interaction tracking)

Layer 4 spec session — net new agent.

QuoteViewLog forwarding signal feeds AGT-401 deal health — buyer engagement and procurement loop-in are positive deal signals. PricingConfig is a new L0 policy table read by AGT-405 (deal terms section) and AGT-504 (price change comms) in addition to AGT-406.

Net new agent. Analysis layer only — transcription handled upstream by recording platform (Gong/Zoom/Chorus) via webhook (hook in v1). 5 analysis dimensions: sentiment scoring, next step extraction with commitment quality, competitor mention detection, objection pattern categorization, email thread analysis. Processes all sales conversations regardless of deal stage. Per-conversation processing on transcript receipt.

ConvIntelligence extended with AGT-407 analysis fields: overall_sentiment, sentiment_score, next_steps (JSON), next_step_committed, competitors_mentioned, objections_raised, unaddressed_showstopper, conv_intelligence_score (0–3, feeds AGT-401 deal health dimension), post_call_summary. Nullable fields — existing 100 rows carry NULL until agent is live.

Layer 4 spec session — net new agent. Identified as dependency for AGT-401, AGT-305, and AGT-405.

Highest-impact new agent in Layer 4. Unblocks the full deal health score. Activates the AGT-305 hypothesis accuracy feedback loop. Every downstream agent reading ConvIntelligence gains richer signal as call volume accumulates.

Net new agent. Daily batch scoring for all active customer accounts. 7-dimension model (100 pts base): product usage trend (22), exec sponsor engagement (20), seat utilization (18), NPS/CSAT (14), engagement recency (12), support ticket trend (8), competitive threat (6). ConvIntelligence adjustment from CSM-owned calls adds up to 12 pts (sentiment, talk ratio, next-step quality — trailing 30-day pattern). Payment health modifier (cap/floor) applied last. Writes CustomerHealthLog as canonical daily record — AGT-502 and AGT-503 read exclusively from this table.

Current: no effect. Overdue: score capped at 77 (cannot reach Low tier). Failed: score capped at 62 (cannot reach Medium). Suspended: score floored at Critical regardless of all other dimensions. Sourced from AGT-803 (L8) read-only feed. Non-blocking — if AGT-803 not live, modifier skipped.

Exec sponsor dimension reads from Contacts table persona-typed records (economic_buyer, champion) rather than raw RepActivity contact count. Same pattern as AGT-401 v18 update.

AGT-803 payment_health_status → CustomerHealthLog.payment_health_status. Read-only feed. Non-blocking dependency.

CustomerHealthLog — one row per account per day. Fields: raw_health_score, health_score (post-modifier), score_delta, payment_health_status, payment_modifier_applied, per-dimension scores (7 fields), conv_intelligence_adj, active_flags (JSON), data_gap_dimensions (JSON), sponsor_departure_flag.

Foundational for L5. CustomerHealthLog is the single source of truth consumed by AGT-502 (churn risk) and AGT-503 (expansion). Sponsor departure flag passes to AGT-502 immediately on same run — does not wait for next AGT-502 cycle.

Net new agent. Runs daily after AGT-501 batch. Reads CustomerHealthLog health_score and applies renewal proximity multiplier: ≤30 days ×0.85, 31–60 ×0.92, 61–90 ×0.96, >90 ×1.0. Effective risk score drives tier assignment: Critical (0–44), High (45–62), Medium (63–77), Low (78–100). Immediate alerts on: score drop ≥15 pts, tier transition, sponsor_departure_flag (fires same run as AGT-501 detection). Weekly digest for all other accounts — one per AM, one per CSM, segment-level to SLM, org-wide NRR risk to CRO (feeds AGT-702).

AM is relationship and revenue owner — primary alert recipient on all churn risk events. CSM is adoption influencer — notified concurrently as context, not action owner. Finance notified first on payment-status alerts (billing event before relationship event), then AM + CSM simultaneously. This corrects the original L6 spec which positioned CSM as primary.

ChurnRiskLog — one row per account per run. Fields: effective_risk_score, churn_risk_tier, renewal_multiplier, renewal_proximity_days, alert_type, recommended_action, recovery_plan_logged, downstream_triggers (JSON), expansion_candidate.

Weekly org-wide NRR risk summary feeds AGT-702 GTM Health Monitor. Role correction changes alert routing for all existing customers — AM inbox receives what previously went to CSM as primary.

Net new agent. Two entry paths: AGT-502 Low-tier pass (scheduled) and AGT-804 usage events (bypass scheduled cycle — overage/seat signals are immediate). Churn risk gate runs first: High/Critical = suppress, Medium = CSM confirm required, Low = proceed. Open opp check before any new play — if open expansion opp exists, enrichment path only (append context, single contextual update to AM, no new play created). Five signals: consumption overage +40 pts (immediate), seat utilization >80% for 2+ months +30 pts (immediate), ConvIntelligence feature inquiry + positive sentiment +30 pts (immediate), new stakeholder engaged +20 pts (digest), cross-sell peer cohort gap +15–25 pts (digest). Peer cohort: segment × vertical × size band, minimum 5 accounts, ≥60% match rate.

AM routing uses TerritoryDefinitions as authoritative account ownership source. Original spec used RepBook. Same change applied to AGT-202 Lead Router in v17. Routing conflict fallback (AM unavailable) → segment manager, same chain as AGT-202.

AGT-804 usage data feed replaces generic Revenue table as source for consumption overage and seat utilization signals. AGT-804 also bypasses the scheduled daily cycle — usage threshold events trigger AGT-503 directly.

ExpansionLog — one row per signal detection. Fields: expansion_score, play_type (new_play / enrichment / suppressed), suppressed_reason, signal_sources (JSON), expansion_play_type, acv_estimate, open_opp_exists, enriched_opp_id, routed_to_am, alert_type, csm_confirm_required, csm_confirmed_at. Accounts extended: expansion_score, expansion_play_type, expansion_acv_potential, expansion_routed_to, expansion_last_detected.

Expansion ACV potential feeds MetricsCalc expansion revenue forecast — first agent to quantify NRR upside in the OS. Open opp check prevents duplicate plays and rep confusion — enrichment path preserves rep context without creating noise.

Net new agent. Execution-only — never initiates the decision to communicate. Dual-track: (1) internal brief to each AM scoped to their affected accounts, priority flags, and proactive outreach CTAs; (2) external transactional comm to appropriate contacts, gated by human approval. Always sends externally regardless of whether AM has contacted the customer first — internal brief is prep, not a replacement.

Hybrid. Price increases: PricingConfig scan detects future-dated rate delta — auto-initiates process, no manual entry required. EoS/EoL and ToS: manual entry to CommEventQueue (L0, RevOps-owned). AGT-504 polls CommEventQueue for pending records.

Price increase → billing contact on order form (AGT-801 source). ToS → billing contact + legal/signatory personas from Contacts table. EoS/EoL → all active users in Contacts table for affected accounts; falls back to billing contact if no Contacts records exist. Generic billing email detection: 1-business-day delay + AM flagged to identify human contact. Send proceeds after delay regardless of resolution. AGT-204 enrichment triggered as persistent fix.

Accounts flagged as priority in internal brief (requiring proactive AM outreach): churn risk tier ≥ Medium (AGT-502), renewal proximity ≤90 days, ABM-active, ACV ≥ segment P75 (configurable per event via CommEventQueue.acv_priority_threshold).

Human approval required before external comm sends — always. RevOps approves all comm types; Legal/Comms also required for ToS and EoS/EoL. Approval and lead time window run in parallel (both must clear). External comm never auto-sends.

Default 3 business days between internal brief delivery and external send. Configurable per event (CommEventQueue.lead_time_days). Clock starts on brief delivery to AMs.

CommEventQueue (L0) — event anchor record, auto-created on PricingConfig detection or manually entered. Fields: comm_type, status, effective_date, affected_skus, change_summary, lead_time_days, acv_priority_threshold, approval fields, auto_detected flag, pricing_config_ref. CommBriefLog — one row per AM per event. CommDeliveryLog — one row per recipient per event, includes generic_email_flag, unresolved_generic_flag, delay_applied_days.

PricingConfig (L0) — reads effective_date and rate_delta for price change detection and external comm content. CommEventQueue (L0) — RevOps-owned intake table for non-price comm events.

Closes the last unmanaged customer touchpoint in the OS. Price changes, product sunset, and legal notices previously had no systematic AM prep path — priority customers could receive a transactional email before their AM knew it was coming. The internal brief window eliminates that gap. CommDeliveryLog provides first audit trail for customer-facing transactional comms — useful for compliance and dispute resolution.

Net new agent. Pre-close trigger fires when deal health ≥ threshold AND stage ≥ Proposal for eligible deals (Enterprise, ACV ≥ threshold, implementation services on quote, SAM ≥ threshold). Post-close trigger fires for all segments at order execution (AGT-801 signal). Pre-close outputs: onboarding plan draft, internal kickoff brief (AM + CSM), resource requirement flag, post-sales risk alert with reject gate. Active onboarding: 7 configurable milestones (M1–M7). TTV = first value event + adoption threshold both required; RevOps defines value events per product in ProductValueEvents; selling team selects + dates in MAP; inheriting team signs off or escalates. Escalation triggers: customer unresponsive (2+ attempts), TTV trajectory miss. Onboarding closes: all milestones complete + CSM manual gate.

If post-sales rejects deal handoff or MAP sign-off not obtained: AGT-601 writes exception_type = 'map_signoff_hold' to ExceptionLog. AGT-104 picks up on next daily run and holds both ACV commission and usage realization payout. Hold lifts automatically on MAP sign-off. SLM notified immediately. SLM override allows deal to proceed despite rejection.

OnboardingLog written continuously (not a closing summary). Every milestone, CSM activity, and exec touchpoint writes a record. AGT-501 reads OnboardingLog for accounts where onboarding_status = 'active' — substitutes OnboardingLog signals for dimensions where live data doesn't exist yet. Health records flagged in_onboarding = TRUE. Exec sponsor engagement, QBR recency, usage, and seat utilization dimensions all have defined OnboardingLog substitutes. NPS/CSAT stays data_gap — no proxy, by design.

Reads ConvIntelligence and OpportunityBriefLog to extract technical win details and rep commitments into MAP. Reads QuoteLog for product configuration context. These are read-only — no writes back to L4 tables.

MutualActionPlan — shared L4/L6 handoff artifact with comp_hold_active field wiring to AGT-104. OnboardingLog — continuous milestone + activity log. ProductValueEvents (L0, RevOps-owned) — first value event definitions per SKU. OnboardingMilestoneConfig (L0, RevOps-owned) — milestone defaults per segment + product. Accounts extended: onboarding_status, onboarding_start_date, onboarding_complete_date, ttv_achieved_days, ttv_target_days, map_id (6 fields).

Closes the institutional memory gap at handoff — technical win details, rep commitments, and customer stakeholder map transfer from selling team to post-sales systematically. Comp hold gate creates a new governance dependency — reps must ensure post-sales signs off on MAP or both ACV and usage comp are held. OnboardingLog eliminates NULL health scores for newly closed customers — AGT-501 has signals from day one.

Net new agent. Activation gate: QuoteLog contains implementation services line item OR ImplementationSOW record exists. Not active = AGT-601 covers M1 (technical setup) with CSM as owner using a simplified checklist. When active: runs in parallel with AGT-601 from close. Technical owner = Implementation Engineer when services purchased; CSM when not. Covers 6 technical milestones: environment provisioning, integration build, data migration, technical UAT, technical go-live sign-off, technical handoff to CSM. Reads MutualActionPlan for technical win details and customer technical stakeholder map. Writes TechnicalMilestoneLog — feeds AGT-601 M1 status. CSM sees AGT-602 status as simplified summary via AGT-601 — IE owns technical detail, CSM owns customer communication about it.

ImplementationSOW (L0) — SOW contract for purchased services: contracted hours, hours_consumed, roles_required, timeline, scope_summary, IE assignment, status. TechnicalMilestoneLog — per-milestone tracking with blocker_detail and customer_sign_off fields. feeds_agt601_milestone field links each technical milestone to AGT-601 M1 progression.

Gives Implementation Engineers their own agent workspace — previously no agent in the OS served this role. Also useful to SMB/MM CSMs as a technical checklist even without dedicated IE — AGT-601 exposes simplified AGT-602 view when implementation services are not purchased.

Net new agent. Dual trigger: quarterly auto (fires X days before QBR due date, default 10 business days) OR meeting booking (whichever first). Dual output: internal prep brief (all 7 sections, full context including internal-only risk and renewal language) + customer-facing summary artifact (sections 1/2/3/6/7 only — expansion ACV and churn risk language are internal-only). Brief is split by role: AM owns sections 4 (expansion) and 5 (renewal); CSM owns sections 1 (usage), 3 (health), 6 (open issues); both own sections 2 (ROI vs MAP) and 7 (next commitments). Customer artifact: CSM edits before sending — agent drafts, human always reviews.

Section 2 (ROI narrative) reads MutualActionPlan.selected_value_events — requires AGT-601 MAP to be populated. Section 4 (expansion) reads ExpansionLog — requires AGT-503 to be active. Section 5 (renewal) reads ChurnRiskLog + Opportunities. Section 3 (health scorecard) reads CustomerHealthLog — translated to executive narrative, not raw scores.

QBRLog — internal brief content (JSON), customer artifact content (JSON), trigger type, health score at prep, expansion signals included flag, renewal proximity days, post-QBR outcome (CSM-logged), commitments agreed (JSON).

First agent in the OS to produce a customer-facing artifact (all prior agents produce internal outputs only). QBRLog.qbr_outcome becomes a new VoC input for AGT-604. QBRLog.commitments_agreed feeds next QBR's MAP alignment check.

Net new agent. Six input signals: NPS verbatim + score, CSAT + comments, support ticket themes, ConvIntelligence (feature requests, objections, competitor mentions), CSM call notes + QBR outcomes, product usage patterns (adoption gap analysis). Monthly deep synthesis: full theme clustering across all sources. Weekly digest: anomaly flags and emerging themes from new signals only. Five consumer outputs: Product (feature themes, adoption gaps), RevOps (ICP refinement signals), Marketing (messaging gaps from objections), Finance/RevOps (pricing feedback for PricingConfig review), CRO (monthly org-wide executive narrative).

AGT-201 ICP Scorer: ICP refinement signals flag when observed customer outcome patterns diverge from current scoring weights — input for RevOps-driven weight recalibration. AGT-205 MarketAssumptions: customer adoption patterns feed market sizing assumption updates. PricingConfig (L0): pricing objection patterns and value perception signals routed to Finance for review — VoC surfaces signal, Finance decides. AGT-702 GTM Health Monitor: customer sentiment trend included as a health signal in L7.

VoCSynthesisLog — per synthesis run: theme clusters (JSON), ICP refinement signals (JSON), pricing feedback (JSON), messaging gaps (JSON), CRO summary (text, monthly only). VoCSignalLog — individual signal records before synthesis. Denormalized segment + vertical fields on each record for cohort analysis without joins.

First agent in the OS to create a feedback loop from customer outcomes back to L0 policy tables and L2 scoring models. Closes the NRR loop: customer experience signals now systematically inform ICP model, pricing decisions, and market assumptions — not just reactive health monitoring. VoCSignalLog.included_in_synthesis field provides full traceability from raw signal to synthesized output.

Net new agent. Role-parameterized across AE, AM, CSM, and SDR — shared skeleton, role-specific metric inputs. SDR Coaching folded in as a role module (not a separate agent). Monthly digest: manager-approved before delivery to rep, contains skill gap assessment, trend commentary, and recommended next action. Daily flag: fires only on critical threshold breach (attainment <50% with ≤6 weeks remaining in period). Differentiated artifacts: rep receives coaching, manager receives pattern summary across their book.

Attainment <50% with ≤6 weeks remaining → SLM notified concurrently with manager. Same pattern 3+ consecutive weeks unresolved → automatic SLM escalation regardless of manager action. Manager cannot suppress the auto-escalation.

Four new coaching signals available from L6: TTV performance vs target (AGT-601 OnboardingLog), onboarding milestone velocity (AGT-601), QBR outcome quality (AGT-603 QBRLog), MAP commitment follow-through (AGT-601 MutualActionPlan). These supplement the existing ConvIntelligence objection handling and AGT-405 deal advance quality signals from L4.

RepSkillAssessmentLog — per-rep per-period skill gap detection across role-specific dimensions. RepCoachingLog — coaching pattern tracking, manager approval state, delivery timestamp, rep acknowledgment.

Layer 7 spec session — net new agent. SDR Coaching previously earmarked as a separate agent; folded in as role-parameterized module to avoid redundant infrastructure.

RepCoachingLog becomes a new input for AGT-703 bias detection — sustained coaching patterns on specific deal types or verticals can surface systematic forecast bias. SDR consolidation reduces agent count from planned 40 to 39. AGT-703 gap preserves the intentional blank.

Net new agent. Computes and persists five canonical GTM efficiency metrics on cadence: Magic Number (revenue basis — never ARR), Rule of 40 (three variants: EBITDA, FCF, operating margin), NRR (>110% target), GRR (>85% target), CAC Payback (<18 months target). All computed values written to MetricsCalcLog as the canonical persisted source — downstream agents read from MetricsCalcLog, never recompute. Snapshot model only: weekly (NRR trajectory + pipeline coverage), monthly (Magic Number, NRR/GRR actuals, R40), quarterly (CAC Payback, trailing averages). No continuous monitoring.

Every MetricsCalcLog row includes four reference points: FP&APlan current version (0+12 → 1+11 → 2+10 rolling), board-approved plan (is_board_plan flag), prior period actuals from MetricsCalcLog history, AGT-402 forward forecast vs plan target.

RevOps only — they triage and escalate. AGT-702 never routes directly to CRO, Finance, or reps. Calibration recommendations (quota, capacity, ICP weight recalibration) surface via AGT-704 MBR synthesis — not AGT-702.

MetricsCalcLog — one row per metric per period per snapshot run. Fields: snapshot_date, cadence, metric_name, metric_value, plan_version, plan_target, board_plan_target, prior_period_value, pct_of_plan, data_gap, data_gap_reason. FP&APlan extended: plan_version (0+12 format), reforecast_as_of_date, is_board_plan fields added.

Layer 7 spec session — net new agent. FP&APlan versioning extension propagates to AGT-402 and AGT-404 which now reference current plan_version for pct_of_target comparisons.

MetricsCalcLog is the first OS table to persist computed efficiency metrics as queryable data — prior period comparisons and trend analysis now available without recomputation. FP&APlan versioning enables accurate plan-vs-actual comparisons as the operating plan rolls forward through the year. AGT-703 reads MetricsCalcLog for historical metric context in forecast accuracy analysis.

Net new agent. Two analysis modes: (1) Win-loss pattern analysis on closed opportunities — loss reason clustering, competitive pattern detection, segment/vertical win rate divergence. (2) Forecast accuracy measurement — rep commit vs closed outcome, stage-weighted vs actual, AGT-402/AGT-404 forecast vs realized. Reads MetricsCalcLog for historical metric context.

CalibrationSignalLog generates typed signals when analysis reveals systematic divergence: quota calibration (persistent miss/overachieve patterns → AGT-101), ICP weight recalibration (win rate divergence by segment/vertical → AGT-201), forecast model adjustment (systematic over/under-forecast → AGT-402/AGT-404), competitive KB update (new competitor patterns → AGT-403). Signals are notifications — humans decide whether to act. Dismissed signals are not re-opened; if the pattern recurs, a new signal is generated.

Rep-level forecast accuracy (commit vs closed) routed to AGT-701 as a coaching input. This is not a public accuracy leaderboard — it feeds the coaching engine to improve rep forecasting behavior.

Minimum sample sizes enforced before pattern conclusions are drawn. Win-loss requires N closed opps per cohort (configurable, default 10). Forecast accuracy requires 2+ completed periods. Below threshold: data surfaced as directional only, not as calibration signals.

WinLossLog — per analysis cycle: win/loss patterns by segment, vertical, competitor, loss reason clusters. ForecastAccuracyLog — per period per forecast method: predicted vs actual, variance, directional accuracy. CalibrationSignalLog — typed calibration signals with target agent, signal strength, supporting evidence, status (pending/acknowledged/actioned/dismissed).

Layer 7 spec session — net new agent.

Closes the loop between forecast methodology and outcomes — the OS can now systematically measure whether its own forecasting agents are accurate and surface recalibration signals when they aren't. Win-loss patterns feed AGT-403 competitive KB with structured loss reason data rather than ad-hoc rep notes. AGT-704 surfaces CalibrationSignalLog in MBR section 6 for leadership review.

Net new agent. Three cadence artifacts: WBR (5 sections — weekly, FLM + SLM + RevOps full + Finance read-only on sections 1 and 5), MBR (8 sections — monthly, broader leadership), exec QBR (7 sections — quarterly, board-ready). Every cadence produces three output formats: structured document, slides, and BI payload. Pre-flight staleness gate runs before every artifact generation — if any upstream source is stale, AGT-704 flags the gap and holds until data is current or human overrides.

Hybrid. WBR narrative roll-up carries forward into MBR as section 1 context. Remaining MBR sections pull fresh from source tables (MetricsCalcLog, ChurnRiskLog, ExpansionLog, CalibrationSignalLog). This avoids a full recompute while ensuring metrics are current at MBR time.

Action items auto-assigned to owners via routing rules derived from item type and content. Routing rules: metric miss → RevOps owner; rep coaching → SLM owner; customer risk → AM owner; calibration signal → owning agent team. Humans confirm before distribution — never auto-sent. Action item lifecycle tracked in ActionItemLog.

AGT-703 CalibrationSignalLog is surfaced in MBR section 6. AGT-704 contextualizes calibration signals (quota, comp, ICP, forecast model adjustments) for leadership review — it does not make the recalibration recommendation autonomously. Humans decide; AGT-704 presents the case.

AGT-704 stays on cadence regardless of mid-period events (major deal wins/losses, leadership changes, market events). No unscheduled exec artifacts. If an urgent situation requires a brief, that is a human-authored document — not an AGT-704 output.

BusinessReviewLog — one row per cadence run. Fields: cadence type, artifact versions (doc/slides/BI payload links), staleness_gate_result, upstream_sources_checked, override_by (if staleness overridden). ActionItemLog — one row per action item. Fields: source_review_id, owner, routing_rule_applied, status, due_date, confirmed_at, closed_at.

Layer 7 spec session — net new agent.

First agent in the OS to produce board-ready artifacts — exec QBR output is designed for external consumption. ActionItemLog closes the accountability gap in business reviews — action items now have owners, due dates, and tracked resolution. Staleness gate creates a forcing function for upstream agents to stay on schedule — AGT-704 will not paper over stale data.

Net new agent. Converts AGT-406 accepted quote into a legally binding order form. Order form adds to the quote: legal entity names + billing address, payment terms, auto-renewal clause + notice period, governing law, MSA reference, SOW reference (if services), commit frequency (monthly/quarterly/annual, rampable), all billable SKUs with unit prices + commitment amounts, total contract value. AGT-801 re-scores the deal (0–100, R/Y/G) against the order form terms — not inherited from AGT-406's quote score.

Two independent tracks: (1) Pricing/deal terms: Green = auto-approve; Yellow = Finance Director; Red = Finance VP + CRO. (2) Legal terms: Green = auto-approve; Yellow = Legal Director; Red = Legal VP + CRO. Both tracks must clear independently. Pricing corrections always require Finance Director minimum regardless of score.

Three amendment triggers: seat expansion, SKU add, pricing correction. Each generates an AmendmentLog record and re-routes through the approval chain. Original order is versioned, not overwritten.

Billing contact written to OrderLog by AGT-801 is the target for AGT-504 Customer Communications on price change events. AGT-601 reads QuoteLog + OrderLog for pre-close onboarding plan context.

OrderLog, OrderLineItems, OrderApprovalLog, AmendmentLog

Closes the quote → legal contract gap — the OS now covers the full commercial lifecycle from pricing proposal (AGT-406) through signed legal order form. AGT-504's billing contact routing now has a reliable data source — OrderLog is the authoritative record of who the billing contact is.

Net new agent. Three billing types: recurring (seat license — monthly/quarterly/annual per contract), consumption (always monthly — reads UsageMeteringLog for units consumed vs commit, overage billed separately), SOW milestone (event-driven — fires on TechnicalMilestoneLog customer sign-off or manual Finance entry). Amendment deltas billed on amendment approval (pro-rated seat expansion, credit for pricing correction).

Four trigger sources: customer dispute, billing error, contract downgrade, rep submission (manager approval → Finance Director approval — two-step gate). Finance Director minimum approval on all credit events. Billing errors always require Finance Director regardless of amount.

Customer sign-off required (TechnicalMilestoneLog.customer_sign_off = TRUE) before AGT-802 invoices. IE completion alone does not trigger billing. Manual Finance entry accepted as equivalent signal when AGT-602 is not active.

InvoiceLog, InvoiceLineItems, CreditMemoLog

Consumption and seat billing are now unified in the OS schema — separate invoices per billing type per account are tracked in InvoiceLog and visible to AGT-803 for payment monitoring. Rep credit submission path creates a governed channel for sales-initiated billing adjustments without Finance having to manage ad hoc requests.

Net new agent. Tracks payment status across all invoices. 3 retry maximum: Retry 1 at 10–20 minutes, Retry 2 next business day, Retry 3 following business day. Finance notified on Retries 1 and 2. Finance + AM notified simultaneously on Retry 3 failure. Suspended state is Finance-initiated (not automatic) — Finance + AM + SLM all notified simultaneously.

Current (no modifier) → Overdue (score capped at 77) → Failed (score capped at 62) → Suspended (score floored at Critical). States are per-account — one overdue invoice puts the account in Overdue state regardless of other invoices. Resolves to Current only when all outstanding invoices are paid.

Writes payment_health_status to CustomerHealthLog on every status change. Non-blocking — if AGT-803 not live, AGT-501 skips the modifier and scores on behavioral dimensions alone. Modifier applied last in AGT-501 scoring sequence so cap/floor is visible as a distinct adjustment.

PaymentEventLog — one row per payment event (attempt, failure, retry, success, escalation)

AGT-501 payment modifier is now fully operational — specced in v19 but the feed was always a planned integration. AGT-803 makes it live. AM notification at Failed status means relationship owners now know about payment issues before customers lose access — previously this happened after Finance had already escalated externally.

Net new agent. Three recognition methods: seat licenses = straight-line ratable (TCV ÷ term months, begins on term_start regardless of invoice payment); consumption = usage recognized as consumed per UsageMeteringLog (no deferred revenue); professional services = milestone delivery on customer sign-off (each milestone is an independent performance obligation). AGT-804 writes DeferredRevenueSchedule for accounting system journal entries — it computes the schedule, the accounting system posts journals.

One row per account per SKU per metering period. External product system is the data source — AGT-804 receives the feed. Multiple OS agents read from UsageMeteringLog: AGT-804 (recognition), AGT-802 (consumption invoicing), AGT-503 (expansion signals — overage +40 pts, seat utilization >80% for 2+ months +30 pts), AGT-501 (seat utilization dimension, 18 pts max). Single external integration point; all consumers downstream of it.

AGT-702 reads RevenueRecognitionLog for recognized revenue in Magic Number calculation — not billed revenue. AGT-503 usage threshold events bypass AGT-503's scheduled daily cycle and trigger directly. AGT-501 seat utilization dimension replaces any prior usage data references with UsageMeteringLog.

RevenueRecognitionLog, DeferredRevenueSchedule, UsageMeteringLog

AGT-702 Magic Number now uses recognized revenue — the canonical efficiency metric is now computed on the correct basis (ASC 606 recognized, not billed). UsageMeteringLog is the single metering integration point — AGT-501 seat utilization and AGT-503 expansion signals no longer require separate product system integrations. L5/L6 ripple pending: AGT-501 and AGT-503 specs should be updated to reference UsageMeteringLog explicitly, replacing any prior generic usage data source references.

Seat utilization dimension (18 pts max) now explicitly reads UsageMeteringLog WHERE account_id = [account] AND sku_id LIKE 'SKU-SEAT%'. Prior spec referenced "AGT-804 usage data feed" without naming the table. Non-blocking — if no UsageMeteringLog records exist for the account, dimension remains data_gap as before.

ConvIntelligence adjustment now explicitly filters call_owner_role = 'CSM' (v23 field). Selling-role calls (AE, SDR, SE) excluded from customer health adjustment — they are deal activity signals, not customer health signals.

None. UsageMeteringLog and call_owner_role already exist. Prompt-only delta.

Seat utilization dimension is now fully operational — the data source is explicit and AGT-804 populates it. CSM call filter prevents selling-role conversations from inflating customer health scores.

Two expansion signals now explicitly source from UsageMeteringLog: (1) Consumption overage (+40 pts): UsageMeteringLog WHERE overage_units > 0 AND period_end = [most recent closed period] — immediate trigger, bypasses scheduled daily cycle. (2) Seat utilization >80% for 2+ months (+30 pts): UsageMeteringLog WHERE (units_consumed / commit_units) > 0.80 for 2+ consecutive periods — immediate trigger. AGT-804 threshold events push UsageMeteringLog writes in real-time; AGT-503 listens and fires on new records rather than waiting for the next scheduled run.

None. UsageMeteringLog already exists. Prompt-only delta.

Usage-driven expansion signals are now fully wired end-to-end — AGT-804 receives product metering data, writes UsageMeteringLog, AGT-503 reads it and fires immediately on threshold breach. No intermediate manual step or batch delay.

Rep Agent Builder Guide updated from schema v17 (L1–L4 only) to v25 (L1–L7). New use case cards: customer health alerts (AM/CSM), expansion opportunities (AM/CSM), QBR prep (AM/CSM), onboarding status (CSM/AM), coaching plan (all roles). New example prompts: customer health + renewal focus, QBR prep for specific account, coaching plan summary. Schema reference extended with L5 (CustomerHealthLog, ChurnRiskLog, ExpansionLog), L6 (OnboardingLog, MutualActionPlan, QBRLog, VoCSynthesisLog), and L7 (RepSkillAssessmentLog, RepCoachingLog, MetricsCalcLog, BusinessReviewLog, ActionItemLog) tables. New guardrail added: don't share internal QBR content externally.

AM and CSM roles now have first-class use cases in the guide — v1 was predominantly AE/SDR facing. No schema or agent behavior changes. Enablement artifact only.