#3349 [Dream Cycle 2026-09-17] intelligence: wire CLAUDE_FLOW_PRIOR_DECAY env override for ModelRouter's dormant decay primitive + capabilities,memory scan

1. Tonight's Rotation

DATE=2026-09-17
DEEP=intelligence
SCAN=capabilities,memory
SLOT=2
COMMIT=6f0ed7112873eedc7cfe17281a2585188190b790
BRANCH=dream/2026-09-17-intelligence

2. Ledger Check

Last ledger row: 2026-09-12 (intelligence, #3301/#3302). Verified via GitHub MCP tools (not inferred): 2026-09-13/09-14 are genuine no-run gaps (no branch/issue/PR evidence, same class as the earlier 08-20..23 gap); 2026-09-15 ran (performance, #3329 — MoE gate load-balance loss computed but never applied to weights) and 2026-09-16 ran (security, #3338 — hive-mind consensus bypass). Trailing-14 PR merge rate: 14/14 MERGED (a bulk human-review sweep on 2026-09-14/15 cleared the entire backlog from #3086 through #3302) — STEP 1.1's "0-of-14 → bias small" rule does not apply; the actual signal is the opposite extreme, though noted as a bulk sweep (weaker scrutiny signal than a steady trickle). Prior gist (2026-09-12) scored 9/10 against the STEP 1.2 rubric (A/B benchmark evidence, 7 competitor rows, 5 file-pathed recommendations, valid witness, novel finding; lost 1pt only for exceeding 1500 words). intelligence surface findings on 05-27/06-02/08-27/09-02/09-07/09-12 are confirmed genuinely distinct sub-areas (EWC/Fisher, reward-blind consolidation, tier-routing, reasoning-bank confidence) — no reject-duplicate-direction trigger.

3. Deep Dive Findings

Tonight's selected finding: ModelRouter's discounted-Thompson-sampling priorDecay primitive (built/tested/benchmarked in #3049, 2026-08-17) was left permanently inert — DEFAULT_CONFIG.priorDecay hardcoded to 1 (disabled) with no env/config override, unlike its sibling maxUncertainty (envMaxUncertainty()). This was explicitly named as the recommended next step for a future intelligence night in the 2026-09-12 gist. Also surfaced but not selected (see Recommendation): MoE gate updateExpertWeights() has zero production callers despite an unsourced "Measured: 0.13→0.88" claim in two CLAUDE.md files (v3/@claude-flow/neural/src/moe-router.ts); SONAAdapter.forceLearningCycle() (integration/src/sona-adapter.ts:664) is never invoked — the production consolidation timer calls consolidatePatterns() directly, so learningCycles stays 0 forever and periodic pruning never runs automatically; a duplicate, fully-unwired AttentionCoordinator class sits in swarm/src/attention-coordinator.ts (zero importers repo-wide) alongside the one actually used (integration/src/attention-coordinator.ts).

4. Hypothesis

Given a long-running Ruflo deployment where .swarm/model-router-state.json accumulates Thompson-sampling history across many sessions, when priorDecay is exposed via a CLAUDE_FLOW_PRIOR_DECAY env override (mirroring envMaxUncertainty(), default unchanged at 1/disabled), then the router should recover faster from a real model-quality shift when an operator opts in, subject to: no default-behavior change when unset, no numeric-stability regression (existing PRIOR_DECAY_FLOOR untouched), and no material stationary-workload accuracy loss (pre-declared ±1pp tolerance from the original #3049 benchmark, not invented tonight).

5. Evaluation Receipt

evaluated: accepted (ACCEPT-scoped — see caveat below).

Discriminating test (stash-isolated): reverted only the source fix → 1/3 new tests failed exactly as predicted; restored → 16/16 pass (13 pre-existing + 3 new).
Full @claude-flow/cli suite: byte-identical failure sets baseline vs candidate (221 pre-existing failures, unbuilt-dist/network-dependent, unrelated); 2497→2500 passed (+3, exactly the new tests); 105 skipped both ways.
tsc --noEmit: 463 pre-existing errors, byte-identical error-line sets both ways (confirmed via diff, not just count).
Benchmark (prior-decay-benchmark.mjs, re-run tonight, byte-identical to the 2026-08-17 receipt): low bucket — non-stationary recovery 26.5→21.9 rounds (t=7.0), post-shift accuracy +1.3pp (t=5.9), stationary Δ=+0.02pp (t=0.7, noise) — real win. med bucket — non-stationary recovery/accuracy flat (t=−0.74/−0.33), stationary Δ=−0.08pp (t=−3.01), a small but statistically real regression accepted only under the original ±1pp tolerance band. Caveat carried forward verbatim from the adversarial critic: describe this as "safely exposes an existing opt-in knob, net-positive in the tested low-complexity bucket, negligible-but-real cost in med-complexity" — not an unqualified win. Default stays disabled; "no regression" for existing deployments is trivially true by construction (opt-in only), which is accurate but shouldn't be conflated with "proven beneficial when turned on."

6. Darwin Results

@metaharness/[email protected] available (ruvector harness status --json) but not invoked: the decay value (0.995) was already the subject of the original #3049 benchmark/selection; tonight's diff is a pure config-exposure change with no new free parameter for Darwin to search.

7. Flywheel Evidence

@metaharness/[email protected]'s gate <evidence> probed directly (gate.js read in full): its frozen conjunctive rule expects {candidate,baseline}.{primary,noopRate,costPerWin} — an executor commit/no-op-rate shape built for a different class of candidate (an agentic policy with commit/no-op actions), not a bandit-router config change. Declined to force our receipt into it rather than fabricate a noopRate that doesn't exist for this candidate. flywheel verify <bundle> similarly needs a signed replay bundle; .harness/README.md confirms WITNESS_SIGNING_KEY isn't configured in this repo, so formal signed replay isn't available — substituted with a manual, independently-repeated reproduction (the adversarial critic re-ran the benchmark script from scratch a second time and got byte-identical numbers).

8. Reward Hack Check

Independent adversarial critic ran the full STEP 10/11 checklist against the actual diff (not the description) by re-executing every claim itself, including deliberately breaking the fix via git stash to confirm the new test is load-bearing. Verdict: CONFIRMED-WITH-CAVEATS. All 10 reward-hack checklist items (test weakening, benchmark weakening, evaluation leakage, hardcoded outputs, metric substitution, selective task removal, seed manipulation, hidden preprocessing, error suppression, cost hiding): not found, each individually verified.

9. Security Review

Not a security-sensitive change. envPriorDecay() reads process.env.CLAUDE_FLOW_PRIOR_DECAY via Number() (no eval/exec), validates Number.isFinite + explicit (0,1] range, falls back safely to the disabled default on any garbage input — same rigor as the existing envMaxUncertainty() pattern it mirrors. No injection surface, no cost/pricing path touched, no credential/filesystem/network scope change.

10. Scan Findings: capabilities

plugins install (cli/src/commands/plugins.ts → plugins/manager.ts) fetches a registry PluginEntry only to print its name/version, then unconditionally npm installs it — never checking declared permissions/trustLevel/securityAudit against PluginStoreConfig.allowedPermissions/minTrustLevel/requireVerification, despite that being fully modeled in store/types.ts and named in SECURITY.md as the privilege-escalation mitigation. --verify (checksum, "default: true") is parsed and never read again. Not selected (SCAN surface, lighter-weight); real gap for a future capabilities/security night.

11. Scan Findings: memory

hnsw-index.ts's binaryQuantize()/scalarQuantize() produce bit-packed/min-range-polluted arrays; distance() only special-cases quantization for 'product' (the already-fixed #3093/#3094 bug) — 'binary'/'scalar' fall through to generic cosine/euclidean/dot computed directly on the packed representation, numerically meaningless, with zero recall test coverage for either. Also: CLAUDE.md's "RaBitQ Quantization — Measured, 32x/0.60ms" claim has zero matching implementation under grep -ri rabitq in memory/src — flagged, not chased further tonight. Not selected (SCAN surface); strong candidate for a future memory DEEP night.

12. Competitors Reviewed

AutoGen/AG2, OpenAI Responses/Assistants API, CrewAI/LangGraph (context), Qdrant, Weaviate, Milvus, LanceDB, Zep/Graphiti (bi-temporal fact memory — genuine open gap, A-grade source, strongest future-night candidate), MCP 2026-07-28 spec + OpenAI Agents SDK strict function-calling (capabilities scan).

13. Gist

docs/dream-cycle/dream-gist-2026-09-17.md (this branch). External gist publication (gh gist create) was not attempted — no gist-creation tool is exposed in this environment's GitHub MCP toolset (repo/issue/PR tools only); this is a tooling-availability gap, not a gh auth failure, so per the STOP CONDITIONS section this is treated the same as FALLBACK for gist publication specifically (all other steps proceeded normally since repo/issue/PR GitHub access is available). The full report is committed to the repo at the path above instead.

14. Witness

Field	Value
Session commit	`6f0ed7112873eedc7cfe17281a2585188190b790`
Gist SHA-256 (pre-witness content)	`e97f610c7699f729fbb4c738486f0201e21230b185006dca2da6116da8d5938a`
Witness stamp	`91d8edc3bf54252f847c593d1e2bd7a7f4e50afae5a18bdf558d9329ebf57715`

Verifier: fetch docs/dream-cycle/dream-gist-2026-09-17.md from this branch, strip the witness table's filled values back to PENDING, SHA-256 the result, concatenate with the session commit above, SHA-256 again — must equal the witness stamp.

15. Recommendation

ACCEPT-scoped — recommend human review and merge. Small (~79 lines across 3 files, 1 conceptual change), fully reviewable, zero regressions across code tests/full suite/typecheck, independently adversarial-reviewed. Human reviewer should weigh the disclosed med-bucket caveat (small statistically-real stationary-accuracy cost, accepted under a pre-existing tolerance band) before deciding whether to also recommend a default non-1 value in a future night — tonight's patch only exposes the knob, it does not change any default. Next-night candidates, ranked: (1) SONAAdapter.forceLearningCycle() 1-line wiring — smallest, safest, ready; (2) MoE updateExpertWeights() + computeLoadBalanceLoss() regularizer as one themed pair (not sequential, per ST-MoE stability literature); (3) HNSW binary/scalar quantization distance dispatch, same bug class as #3093/#3094.

Merge policy: human review required. Do not self-merge. Do not autonomously promote Flywheel state.