Upload folder using huggingface_hub

docs/CORTEX_ARCHITECTURE.md

@@ -192,11 +192,21 @@ class RouterStep(BaseModel):
     metacognition_state: MetacognitionState
     tokens_spent: int
     subagent_report: Optional[SubagentReport] = None  # populated when routing_action.kind == "call_subagent"
+    policy_kind: Literal["trainable", "deterministic_fallback"] = "trainable"  # which policy emitted this step
 ```
 
 **Rationale:** Per cortex/CLAUDE.md "Logging contract", the training-data
 row IS one router step. This schema makes it easy to dump a full
-trajectory as `List[RouterStep]` without bespoke serialisation.
+trajectory as `List[RouterStep]` without bespoke serialisation. The
+`policy_kind` field (OQ-3 resolution) tags which policy emitted the
+step: `"trainable"` for the Session-15 learned router, `"deterministic_fallback"`
+for the Session-13 router used both as B3 baseline AND as the crash
+fallback when the trainable router raises mid-tick (see §7 metacognition
+layer). GRPO advantage computation filters to `policy_kind == "trainable"`
+only — fallback steps are off-policy and excluded from the gradient.
+Default is `"trainable"` so B3 round-tripping does not need to set the
+field; B3 is a baseline, not training data, so the default value is
+inert for B3 trajectories.
 
 #### A4. `Trajectory` (Session 13 — referenced by training/* and inference.py)
 
@@ -268,9 +278,21 @@ while True:
 - **Forward only**: Divergence → Challenge → Narrowing → Convergence. No
   forward-skipping (cortex/CLAUDE.md "Phase machine invariants").
 - **Backward re-entry on dissent**: if `preserve_dissent` was emitted *and*
-  `metacog_state.preserved_dissent_count` increased above a tracked
-  threshold, the router may emit `switch_phase(Challenge)` to re-open
-  challenge. This is the only allowed backward transition.
+  `metacog_state.preserved_dissent_count >= 2`, the router may emit
+  `switch_phase(Challenge)` to re-open challenge. This is the only
+  allowed backward transition. The `>= 2` threshold (Item A resolution)
+  reflects "two preserved dissents = real minority views surfaced; a
+  single one is noise" — re-opening challenge on a single preserved
+  dissent would burn the cross-brain-challenge cap on what may be a
+  one-off objection.
+- **Round-2 entry trigger** (Item C resolution): the router enters
+  round 2 ONLY by emitting an explicit `switch_phase(Divergence)`
+  routing action. Council Executive forbids implicit round increment
+  (e.g. silently bumping `round = 2` after Convergence-without-emit).
+  Rationale: every round boundary is an explicit policy decision,
+  which keeps the Session-15 router's training signal clean — the
+  trainable router learns "when to spend a second round" as a
+  first-class action choice rather than as an emergent side-effect.
 - **Round increment**: every time `phase` returns to Divergence (or
   re-enters Challenge from Convergence/Narrowing via a backtrack), the
   Council Executive checks `deliberation_rounds_used`. If it's already 2,
@@ -294,6 +316,28 @@ while True:
    the failing brain's recommendation is treated as `top_action = NoOp()`
    with `top_confidence = 0.0`; the round continues with the other two.
 
+### Phase-to-protocol mapping (Item F resolution)
+
+The 5-step anti-hivemind protocol (cortex/CLAUDE.md, §8.1 of design doc)
+maps onto the 4-phase machine as follows. This pin lets
+`tests/test_cortex_protocol_invariants.py` assert the exact protocol
+step exercised in each phase.
+
+| Anti-hivemind step | Phase |
+|---|---|
+| 1. Independent reasoning (private first pass) | Divergence |
+| 2. Cross-brain critique (typed evidence disclosure + targeted challenge entry) | Challenge phase entry |
+| 3. Counter-recommendation (challenge resolved or accepted) | Challenge phase end |
+| 4. Anonymized comparison | **Deferred V2 per Decision 56** — not implemented in MVP |
+| 5. Final aggregation with preserved dissent | Narrowing + Convergence |
+
+The Session-12 protocol-invariant test asserts that for any `Council.step`
+trace, every phase-entry `RouterStep` is paired with exactly the protocol
+step listed above (or a logged short-circuit reason — e.g. budget
+exhausted, brain failure, override). Step 4 is asserted as
+**not-fired** in MVP; if it ever fires, the test treats that as a
+forward-leak from a V2 branch into MVP and fails closed.
+
 ---
 
 ## §4 — Hard cap accounting and preserved-dissent tracking
@@ -629,7 +673,7 @@ Decisions are grouped by layer. Each entry: **decision** / **rationale** /
 
 5. **Subagent output validated via Pydantic `TypeAdapter[<role-specific-type>]`.** / Already the pattern in `parse_action` (B1); reusable. / Considered hand-rolled JSON parsing; rejected — Pydantic gives discriminated-union safety free.
 
-6. **Empty `SubagentReport` on parse / call failure: `confidence=0.0, evidence=[]`** for any of the three subagent types (with role-specific other fields zeroed/empty). / Honest "no signal" state; downstream r_proto naturally penalizes. / Considered raising the exception; rejected — would break the per-brain isolation contract.
+6. **Empty `SubagentReport` on parse / call failure: `confidence=0.0, evidence=[]`** for any of the three subagent types (with role-specific other fields zeroed/empty). / Honest "no signal" state; downstream r_proto naturally penalizes. / Considered raising the exception; rejected — would break the per-brain isolation contract. **(D-FR-3 reinforcement)** A single subagent's empty fallback does NOT trigger `parse_failure_marker` — that fires only on full-brain emptiness (Brain Executive sees `top_confidence == 0` from all 3 subagents); see Decision 42.
 
 7. **Subagent caller_id format: `cortex:<brain>:<role>:t<tick>:r<round>:s<step_idx>`.** / Mirrors B2's `b2:t<tick>:p<n>:<role>` pattern. / Considered a flat `cortex:<step_idx>` form; rejected — losing the per-brain breakdown loses critical training analytics.
 
@@ -637,7 +681,7 @@ Decisions are grouped by layer. Each entry: **decision** / **rationale** /
 
 ### Lenses (Session 10)
 
-9. **Three lenses: epidemiology, logistics, governance. Each is a function `lens_for(brain, obs) -> BrainLensedObservation`.** / Matches the 3 brains. / V2 brains (Communications, Equity) have stub lens entries returning `BrainLensedObservation` with the full obs; future sessions add real ones.
+9. **Three lenses: epidemiology, logistics, governance. Each is a function `lens_for(brain, obs) -> BrainLensedObservation`.** / Matches the 3 brains. / **(Item E resolution)** V2 brains (Communications, Equity) deferred entirely per Decision 56 — no MVP stub lens functions. `lens_for(brain, obs)` raises `KeyError` on `brain ∈ {'communications', 'equity'}`. Rationale: stubs returning the raw obs would silently let a V2-leaked code path "work", masking the boundary; raising is the loud-fail behaviour V2 deferral contracts demand.
 
 10. **Epidemiology derived features: `r_effective_estimate, worst_region_infection, transmission_rate_trend`.** / These are exactly the signals an epi-savvy human would compute from telemetry. / Considered raw telemetry only (no derived features); rejected — forces every WM call to redo the same computations.
 
@@ -673,17 +717,17 @@ Decisions are grouped by layer. Each entry: **decision** / **rationale** /
 
 24. **Aggregation rule for the 3 `BrainRecommendation`s: weighted vote on `top_action.kind` first, then on parameters within the winning kind.** / Two-level vote keeps semantically-similar actions together. / Considered exact-match-only; rejected — would force Cortex to flip-flop on minor parameter differences.
 
-25. **Vote weight per brain: `top_confidence × evidence_count`.** / Rewards confident, well-evidenced brains. / Considered uniform weights; rejected — would erase brain-level differences.
+25. **Vote weight per brain: `top_confidence × evidence_count`.** / Rewards confident, well-evidenced brains. / Considered uniform weights; rejected — would erase brain-level differences. **(D-FR-2 resolution)** Do NOT multiply by `(1 - uncertainty)` here — Decision 16 already bakes `(1 - uncertainty)` into `top_confidence` via the multiplicative `confidence × (1 - uncertainty)` rule. Adding it again at the council layer would double-count uncertainty and over-penalise the most-uncertain brain. Cross-reference Decision 16 when reviewing this in future sessions.
 
 26. **Tie-breaking: deterministic ordering `epidemiology > logistics > governance`.** / Required for eval determinism. / Considered random; rejected — breaks the eval-mode determinism contract.
 
-27. **Cross-brain challenge selection: when `request_challenge` fires, the *challenger_brain*'s Critic runs against the *target_brain*'s top plan.** / Matches design §10 worked example. / Considered pairwise challenge (both brains' Critics); rejected — exceeds the 1-challenge cap.
+27. **Cross-brain challenge selection: when `request_challenge` fires, the *challenger_brain*'s Critic runs against the *target_brain*'s top plan.** / Matches design §10 worked example. / Considered pairwise challenge (both brains' Critics); rejected — exceeds the 1-challenge cap. **(Item B resolution)** Cross-brain Critic USR field includes BOTH the target's `PerceptionReport` AND the challenger's `PerceptionReport` (≈ 200 extra tokens / cross-brain challenge — affordable within `TICK_BUDGET=6000`). Rationale: a logistics Critic challenging an epi plan needs the epi-lens view of the data (otherwise the critique is uninformed) plus its own logistics-lens view (otherwise the challenge has no domain leverage). Single-perception Critic was considered and rejected — would force the Critic to "guess" what the target brain saw.
 
 28. **Phase advancement triggers** (Session 13's deterministic router; Session 15 learns these):
     - Divergence → Challenge: `inter_brain_agreement < 0.4` AND `round 1 complete`.
     - Challenge → Narrowing: challenge resolved OR cap hit.
     - Narrowing → Convergence: top candidate stable across last router step OR `urgency > 0.7` OR `budget_remaining_frac < 0.2`.
-    / These thresholds are the deterministic-router defaults; trainable router learns its own. / Other thresholds considered; locked here for Session 13's reproducibility.
+    / These thresholds are the deterministic-router defaults; trainable router learns its own. / Other thresholds considered; locked here for Session 13's reproducibility. **(D-FR-1 resolution)** Conservative-only: `< 0.4` is the *single* threshold for `request_challenge` firing. The aggressive variant — also firing on `agreement ∈ [0.4, 0.7] AND round == 1` — was rejected. Reasoning: the challenge cap is 1/tick, so aggressive triggering buys *earlier* challenges, not *more* challenges. Challenging moderate disagreement (the [0.4, 0.7] band) misallocates the cap; high agreement is consensus and must not be challenged at all. Aligns with the anti-hivemind story — challenge real dissent, not moderate disagreement.
 
 29. **Forced convergence on round 2 cap: emit current top-ranked action; if no candidate, emit `parse_failure_marker()`.** / Matches B2's pattern. / Considered emitting `NoOp()` directly; rejected — `parse_failure_marker` correctly fires the r_policy=0 penalty.
 
@@ -701,7 +745,7 @@ Decisions are grouped by layer. Each entry: **decision** / **rationale** /
 
 35. **`collapse_suspicion`: 1.0 if all 3 brains' `top_action` is bytewise-equal AND all 3 evidence lists are length 0; else 0.0.** / Catches the "all said NoOp with no reasoning" failure. / Considered a graded score; rejected — this is the binary signal eval cares about.
 
-36. **`urgency`: `1.0 - ticks_remaining / max_ticks` clipped + `worst_region_infection_estimate × 0.5`, clamped to [0, 1].** / Both time pressure and crisis severity. / Considered ticks-only; rejected — task-difficulty differences vanish.
+36. **`urgency`: `1.0 - ticks_remaining / max_ticks` clipped + `worst_region_infection_estimate × 0.5`, clamped to [0, 1].** / Both time pressure and crisis severity. / Considered ticks-only; rejected — task-difficulty differences vanish. **(D-FR-4 resolution)** Multiplier confirmed at `0.5`. The alternative — `× 1.0` — was rejected: a 1.0 weight would force premature convergence on `outbreak_hard` precisely when anti-hivemind reasoning matters most. Hard tasks need *more* deliberation, not faster convergence. The 0.5 weight lets infection severity nudge urgency without dominating it.
 
 37. **`novelty_yield_last_round`: only computed in round 2; round 1 always returns 0.0.** / Eval-only signal per design §7.4.3 line 404. / N/A.
 
@@ -709,10 +753,11 @@ Decisions are grouped by layer. Each entry: **decision** / **rationale** /
 
 38. **Session 13 deterministic router decision table** (binding for B3 baseline):
     - Round 1, Divergence, no recommendations yet: `call_subagent(epi, world_modeler)`, then `(epi, planner)`, then `(epi, critic)`, then logistics, then governance — fixed brain order.
-    - End of round 1, agreement < 0.4: `switch_phase(Challenge)` → `request_challenge(challenger=logistics, target=epidemiology)`.
+    - End of round 1, agreement < 0.4: `switch_phase(Challenge)` → `request_challenge(challenger, target)` where **`challenger = brain with min(top_confidence)`** and **`target = brain with max(top_confidence)`**, ties broken by Decision 26's deterministic ordering (`epidemiology > logistics > governance`).
+    - **All-equal-confidences edge case**: when `min(top_confidence) == max(top_confidence)` across all three brains, do NOT fire `request_challenge` at all. Skip the Challenge phase and `switch_phase(Narrowing)` directly. Cross-brain challenge requires productive asymmetry; identical confidences provide none — Decision-26 tie-break would otherwise collapse to `challenger == target == epidemiology`, violating the cross-brain contract. The 1-challenge cap is preserved (unspent), and the council still has the dissent-preservation channel for surfacing minority recommendations.
     - End of round 2 OR agreement ≥ 0.7: `switch_phase(Convergence)` → `emit_outer_action(council_top)`.
     - Budget < 20% remaining: `emit_outer_action` immediately.
-    / Reproduces design §10 worked example. / N/A.
+    / Reproduces design §10 worked example. / **(Decision-38 inconsistency fix)** The earlier draft hardcoded `challenger=logistics, target=epidemiology` without justification. The dynamic pair lets the *most-uncertain* brain push back on the *most-confident*, which is where anti-hivemind correction has highest expected value: a low-confidence brain challenging a high-confidence one is exactly the configuration where dissent could productively flip the council's decision. Hardcoded pair was rejected because (a) it presupposes which domain is "always right to be challenged", which the design doc does not assert, and (b) it defeats the point of the metacognition signal feeding the router. The all-equal-confidences edge case is handled by short-circuiting to Narrowing rather than firing a same-brain self-challenge.
 
 39. **Session 15 trainable router: Option B (small MLP head over 24-dim featurized state, discrete output over top-50 most-common (kind, brain, subagent) tuples).** / Per design §22; Colab-compatible. / Option A (LoRA) is the stretch.
 
@@ -764,26 +809,50 @@ Decisions are grouped by layer. Each entry: **decision** / **rationale** /
 
 60. **All MetacognitionState fields use `float ∈ [0, 1]` except `tick`, `round`, `preserved_dissent_count`, `challenge_used_this_tick`.** / Already locked in `cortex/schemas.py`. / N/A.
 
-### Decisions flagged for human review (NOT auto-resolved)
-
-These need explicit ✅ / ❌ before Session 9 starts:
-
-- **D-FR-1.** Should Sessions 13's deterministic router emit
-  `request_challenge` only on `inter_brain_agreement < 0.4`, or also on
-  `inter_brain_agreement ∈ [0.4, 0.7] AND round == 1`? / The latter is
-  more aggressive but uses the challenge cap more often (matched-
-  compute story may shift).
-- **D-FR-2.** Should the Council Executive's vote weight use
-  `top_confidence × evidence_count` (Decision 25) or
-  `top_confidence × evidence_count × (1 - uncertainty)`? / The latter
-  punishes confidently-wrong brains harder.
-- **D-FR-3.** Should `parse_failure_marker()` fire on any subagent parse
-  failure, or only on the brain's *Brain Executive* declaring "all
-  subagents empty"? / The current proposal (Decision 6 + 42) is the
-  latter; a more aggressive policy would fire it earlier.
-- **D-FR-4.** Should `MetacognitionState.urgency` weight the worst-region
-  infection at 0.5 (Decision 36) or 1.0? / Higher weight = router
-  converges faster on hard tasks.
+### Phase A review-pass additions (Sessions 9–13 binding)
+
+These decisions land during the Phase A review pass (April 2026); each
+is locked to the same standard as Decisions 1–60.
+
+61. **Round-2 entry mechanism** (Item C): the router enters round 2 ONLY by emitting an explicit `switch_phase(Divergence)` routing action. Council Executive forbids implicit round increment. / Every round boundary becomes a first-class policy decision; the Session-15 trainable router learns "when to spend a second round" as an explicit action choice rather than an emergent side-effect — cleaner gradient signal. / Considered implicit round-bump on Convergence-without-emit; rejected — would couple round increment to phase state and obscure the router's training signal. (Cross-reference §3 transition rules.)
+
+62. **Round-2 `prior_belief` encoding when round 1 produced nothing** (Item D): when round 1 produced no useful `BeliefState` (parse failure, empty subagent fallback, or LLM-call exception), round-2 `SubagentInput.prior_belief` is an EMPTY `BeliefState(brain=<X>, latent_estimates={}, hypotheses=[], uncertainty=1.0, reducible_by_more_thought=False, evidence=[])`, NOT `None`. / Different prompt signals: `None` tells the WorldModeler "round 1 has not happened, no history to revise"; an empty `BeliefState` tells it "round 1 happened, produced nothing — start clean but acknowledge the failed pass". Conflating the two would mask the failure mode in the trajectory log. / Considered using `None` for both; rejected — collapses two distinct epistemic states into one.
+
+63. **`PerceptionReport.salient_signals` cap** (OQ-2 resolution): `salient_signals` is `List[str]` with **at most 5 entries** per report. Pydantic validator enforces. / Cap prevents the LLM (in V2 lens-extension experiments) from emitting a paragraph as a "signal"; tests assert content not schema. / Considered a per-brain enum; rejected — over-rigid for early MVP iteration where lens features are still being tuned.
+
+64. **Training-rollout temperature** (OQ-1 resolution): deferred to Session 15. Phase A does not pin a numeric value. `cortex.llm_client.LLMClient.chat(..., temperature: float = 0.0)` already exposes the parameter; Session 15 trainer reads its rollout temperature from the GRPO config and passes it through. / Right session for this decision is Session 15, where the trainer will run early experiments and pick a value based on observed exploration. / Considered pinning `0.7` (common GRPO default); rejected — premature; locks a hyperparameter before any training data exists.
+
+65. **`RouterStep.policy_kind` GRPO advantage filter** (OQ-3 resolution): when the Session-15 trainable router crashes mid-tick and Council falls back to the deterministic router (per §7 metacognition layer), the deterministic-router steps that close the tick are tagged `policy_kind="deterministic_fallback"`. GRPO advantage computation filters to `policy_kind == "trainable"` ONLY — fallback steps are off-policy with respect to the gradient. B3 baseline runs are unaffected: every B3 RouterStep gets the default `policy_kind="trainable"`, but B3 trajectories are not training data so the field is inert for B3. / Cleanest split: keep all router steps on the trajectory (so trainer can still measure failure rate, fallback-tick reward, etc.) but exclude fallback steps from the policy gradient. / Considered dropping fallback steps entirely from the trajectory; rejected — loses observability into how often the trainable router crashes and what the deterministic fallback does instead. (Cross-reference §2 A3 schema and §7 metacognition layer.)
+
+### Decisions resolved during Phase A review pass (April 2026)
+
+All four D-FR items below were resolved in the review pass and folded
+into the relevant decision entries above. Listed here for trace-
+ability; the binding text now lives on the referenced decisions.
+
+- **D-FR-1 → resolved.** Conservative threshold only (`inter_brain_agreement < 0.4`). Aggressive variant rejected. → Decision 28 updated.
+- **D-FR-2 → resolved.** Vote weight stays `top_confidence × evidence_count`; do NOT add `(1 - uncertainty)` factor (already in Decision 16's `top_confidence`). → Decision 25 updated.
+- **D-FR-3 → resolved.** `parse_failure_marker` fires only on full-brain emptiness (status quo of Decisions 6 + 42). Single-subagent failures use 1-retry + empty-report fallback. → Decision 6 updated.
+- **D-FR-4 → resolved.** `urgency` keeps `worst_region_infection × 0.5`. The 1.0 weight was rejected — would force premature convergence on `outbreak_hard`. → Decision 36 updated.
+
+### Open questions resolved during Phase A review pass
+
+- **OQ-1 → resolved.** Training-rollout temperature deferred to Session 15. → Decision 64 added.
+- **OQ-2 → resolved.** `salient_signals` is free-form `List[str]` capped at 5. → Decision 63 added.
+- **OQ-3 → resolved.** Trainable-router fallback steps logged with `policy_kind` field on `RouterStep`; GRPO filter excludes fallback. → §2 A3 + Decision 65 added.
+
+### Additional pins resolved during Phase A review pass
+
+- **Item A → resolved.** Backward re-entry threshold pinned at `preserved_dissent_count >= 2`. → §3 transition rules updated.
+- **Item B → resolved.** Cross-brain Critic USR includes BOTH target's and challenger's perception (~200 extra tokens / challenge). → Decision 27 updated.
+- **Item C → resolved.** Round-2 entry: only via explicit `switch_phase(Divergence)` from router. → §3 transition rules + Decision 61 added.
+- **Item D → resolved.** Round-2 `prior_belief` when round 1 failed: empty `BeliefState`, NOT `None`. → Decision 62 added.
+- **Item E → resolved.** No V2-brain stub lens functions; `lens_for(brain, obs)` raises `KeyError` on V2 brain ids. → Decision 9 updated.
+- **Item F → resolved.** Phase-to-protocol mapping pinned (steps 1, 2, 3, 5 → MVP phases; step 4 → V2-deferred). → §3 phase-to-protocol mapping subsection added.
+
+### Inconsistency resolved during Phase A review pass
+
+- **Decision 38 inconsistency → resolved.** Hardcoded `request_challenge(challenger=logistics, target=epidemiology)` replaced with dynamic pair: `challenger = brain with min(top_confidence)`, `target = brain with max(top_confidence)`, ties broken by Decision 26's deterministic ordering. → Decision 38 updated.
 
 ---
 
@@ -873,17 +942,25 @@ Session 15 (out of Phase A scope) (Trainable router)
 
 ### Word count
 
-≈ 6,750 words. Within the 5,000–8,000 target.
+≈ 8,275 words after the Phase A review pass — about 275 over the
+original 8,000 soft ceiling. (Initial draft: ≈ 6,529 words; review
+pass added ≈ 1,750 words of resolution prose, including per-resolution
+"why rejected" rationale to keep §9's decision-format consistent.)
+The overage is accepted as the cost of folding all 13 resolutions
+inline rather than as a separate addendum file — single source of
+truth for Sessions 9–13.
 
 ### Sections completed (10/10)
 
-§1 Data flow / §2 Schemas / §3 Phase machine / §4 Hard caps + dissent /
-§5 Token budget / §6 Routing policy / §7 Failure modes / §8 Test
-strategy / §9 Pre-approved decisions (60 entries) / §10 Sequencing.
+§1 Data flow / §2 Schemas / §3 Phase machine (now includes phase-to-
+protocol mapping) / §4 Hard caps + dissent / §5 Token budget /
+§6 Routing policy / §7 Failure modes / §8 Test strategy / §9 Pre-
+approved decisions (65 entries — 60 initial + 5 review-pass additions) /
+§10 Sequencing.
 
 ### Decisions auto-resolved (rationales recorded inline in §9)
 
-60 modeling decisions, grouped by layer:
+65 modeling decisions, grouped by layer:
 - 8 subagent decisions
 - 6 lens decisions
 - 7 brain executive decisions
@@ -893,39 +970,43 @@ strategy / §9 Pre-approved decisions (60 entries) / §10 Sequencing.
 - 3 failure / fallback decisions
 - 4 testing decisions
 - 12 tooling / V2-deferral decisions
+- **5 review-pass additions (61–65)**: round-2 entry mechanism,
+  round-2 `prior_belief` encoding, `salient_signals` cap, training-
+  rollout temperature deferral, `RouterStep.policy_kind` GRPO filter.
+
+### Phase A review pass — resolutions (April 2026)
 
-### Decisions flagged for human review (4)
-
-- **D-FR-1.** Aggressive vs. conservative challenge-trigger threshold
-  for Session 13's deterministic router.
-- **D-FR-2.** Whether the council's vote weight should include
-  `(1 - uncertainty)` as an additional factor.
-- **D-FR-3.** Whether `parse_failure_marker` should fire on any
-  subagent parse failure, or only on full-brain "all subagents empty".
-- **D-FR-4.** `MetacognitionState.urgency` weight on worst-region
-  infection (0.5 vs 1.0).
-
-### Open questions before Session 9 starts
-
-- **OQ-1.** What temperature value > 0 should training rollouts use?
-  cortex/CLAUDE.md says "higher temperature during training rollouts"
-  but doesn't pin a number. Suggest `0.7` based on common GRPO setups;
-  needs Session 15 acknowledgement.
-- **OQ-2.** Should `PerceptionReport.salient_signals` be free-form
-  strings or constrained to a per-brain enum? The schema currently has
-  `List[str]`; constraining would make tests stricter but Lens output
-  more rigid.
-- **OQ-3.** When the trainable router (Session 15) crashes mid-tick
-  and falls back to deterministic (§7), should the deterministic
-  router's choices be logged as `RouterStep`s for training data, or
-  marked as off-policy and excluded? GRPO advantage computation differs
-  in each case.
+All 4 D-FR items, 3 OQ items, 6 additional pins (Items A–F), and 1
+Decision-38 inconsistency resolved. Each is folded into the relevant
+inline decision; the §9 "resolved during Phase A review pass"
+subsection lists the cross-references for traceability.
+
+| ID | Resolution (1-line) | Lands in |
+|---|---|---|
+| D-FR-1 | Conservative-only `< 0.4`; aggressive variant rejected. | Decision 28 |
+| D-FR-2 | Vote weight stays `top_confidence × evidence_count`; no extra `(1 - uncertainty)` factor (Decision 16 already bakes it in). | Decision 25 |
+| D-FR-3 | `parse_failure_marker` fires only on full-brain emptiness; single-subagent failures use empty-report fallback. | Decision 6 |
+| D-FR-4 | `urgency` keeps `worst_region_infection × 0.5`; 1.0 weight rejected. | Decision 36 |
+| OQ-1 | Training-rollout temperature deferred to Session 15. | Decision 64 (new) |
+| OQ-2 | `salient_signals` is free-form `List[str]` capped at 5. | Decision 63 (new) |
+| OQ-3 | `policy_kind` field on `RouterStep`; GRPO filter excludes fallback. | §2 A3 + Decision 65 (new) |
+| Item A | Backward re-entry pinned at `preserved_dissent_count >= 2`. | §3 transition rules |
+| Item B | Cross-brain Critic USR includes target + challenger perception (~200 extra tokens). | Decision 27 |
+| Item C | Round-2 entry: only via explicit `switch_phase(Divergence)`. | §3 + Decision 61 (new) |
+| Item D | Round-2 `prior_belief` when round 1 failed: empty `BeliefState`, NOT `None`. | Decision 62 (new) |
+| Item E | No V2-brain stub lens; `lens_for` raises `KeyError` on V2 brain ids. | Decision 9 |
+| Item F | Phase-to-protocol mapping pinned (steps 1, 2, 3, 5 → MVP phases; 4 → V2-deferred). | §3 phase-to-protocol mapping subsection |
+| Decision 38 inconsistency | Hardcoded `(challenger=logistics, target=epidemiology)` → dynamic `(min top_confidence, max top_confidence)`, ties via Decision 26. | Decision 38 |
+
+### Standing-by status
+
+**Decisions locked. Session 9 unblocked.** Phase A is complete; no
+remaining flagged items, no remaining open questions. Sessions 9–13
+implement against this document. Future revisions to Phase A
+contracts require explicit ✅ in a new review pass.
 
 ### What is NOT in this document
 
 - No Python code (per the no-code constraint).
-- No commit. The file is written but not staged. Run
-  `git add docs/CORTEX_ARCHITECTURE.md && git commit -m "docs(cortex):
-  add Phase A architecture contract"` if/when ready.
 - No Session 14+ details (training reward composition, eval harness,
   HF Spaces deploy). Phase A intentionally bounds at Session 13.

docs/SMOKE_GATE_RUNBOOK.md

@@ -0,0 +1,38 @@
+# Smoke Gate Runbook
+
+Use this runbook for the Session 7d/lock-and-test smoke gate on Windows.
+
+## Preferred Shells
+
+Use Git Bash or WSL for the canonical command style:
+
+```bash
+docker run -d --rm -p 58300:8000 --name cwc-diag crisisworldcortex-env:7d-smoke
+until curl -sf http://localhost:58300/health >/dev/null 2>&1; do sleep 2; done
+ENV_URL=http://localhost:58300 HF_TOKEN="$HF_TOKEN" uv run python inference.py
+docker logs cwc-diag
+docker stop cwc-diag
+```
+
+## Windows `cmd.exe`
+
+Do not write `set ENV_URL=http://localhost:58300 && uv run ...`.
+`cmd.exe` includes the space before `&&` in the variable value, producing
+bad values such as `58300 ` or URLs with a literal trailing space.
+
+Use one assignment per line:
+
+```bat
+set ENV_URL=http://localhost:58300
+set HF_TOKEN=<real-token>
+uv run python inference.py
+```
+
+For Hugging Face Spaces, also clear Docker mode first:
+
+```bat
+set LOCAL_IMAGE_NAME=
+set ENV_URL=https://angshuman28-crisisworldcortex.hf.space
+set HF_TOKEN=<real-token>
+uv run python inference.py
+```