[Dream Cycle 2026-09-12] intelligence: findSimilar() conflates learned confidence with query-similarity, inverting distillLearning()'s reliability gate + capabilities,memory scan #3301

1. Tonight's Rotation

DATE=2026-09-12
DEEP=intelligence
SCAN=capabilities,memory
SLOT=2 (DAYINT % 5 == 2)
COMMIT=39e0b0540c9b018174955fc8a21f355bbac26c6a
No bonus deep-dive triggers (DAYINT%25=12, %75=37)

2. Ledger Check

Inspected last 14+ rows of docs/dream-cycle/LEDGER.md and verified GitHub state via MCP tools (no gh CLI in this environment) for the last 7 nights:

Trailing-7-night merge rate: 4/7 (up from the historical "0 of last 14" low). No forced small-patch bias triggered by STEP 1.1, but the general small-scoped-patch discipline was kept anyway. No 3x-duplicate-direction on the intelligence surface (Thompson-decay #3049, EWC-gate #3110, LearningBridge #3160, tier-forwarding #3221 are all distinct bugs) — no surface substitution needed.

Prior gist score (STEP 1.2): scored the 2026-09-09 gist 9/10 (2 benchmark-evidence, 2 competitor-rows [6 rows], 2 executable-recommendations, 2 valid-witness, 0 for <1500 words [it ran long], 1 for genuine novelty).

3. Deep Dive Findings

Five research roles ran in parallel (deep researcher, 2 scan researchers, competitor analyst, independent architecture reviewer — see docs/dream-cycle/dream-gist-2026-09-12.md for full detail).

Selected finding (architecture reviewer): LocalReasoningBank.findSimilar() (v3/@claude-flow/cli/src/memory/intelligence.ts:604-638, pre-fix) overwrote every returned pattern's confidence (learned reliability) with the per-query cosine similarity score — return { ...s.pattern, confidence: s.score }. This broke two things: (a) distillLearning()'s "only distill from high-confidence matches" gate (match.confidence < 0.5, line 357) tested query-similarity instead of reliability, inverting its stated intent; (b) the already-exported PatternMatch extends Pattern { similarity: number } contract (line 1195) that findSimilarPatterns() realizes could only ever return similarity === confidence, via an unsafe type-cast fallback ((r as unknown as {similarity?:number}).similarity ?? r.confidence ?? 0.5).

Deep researcher's 5 additional scored candidates (not selected tonight, see gist for full detail and scores): priorDecay distribution-shift fix built/tested/benchmarked but never wired to any config/env (score 4.55); MoE router missing an explore/exploit toggle unlike its sibling Q-learning router (4.20); determineEvolutionType() mislabels neutral/declining pattern evolution as 'improvement', duplicated in 2 files (4.15); ReasoningBank.distill() only distills successful/partial trajectories, never failure-derived memory despite the paper it's named after (arXiv:2509.25140) arguing for both (3.95); inconsistent cosine-similarity clamping across ~5 duplicate implementations (3.60). The selected finding scores 4.85 under the same rubric (5/5/5/4/5/5) — highest of all candidates, no override needed.

Independent re-verification: confirmed #3110 (EWC-gate wrong Fisher dims) and #3160 (LearningBridge reward-blindness) are still present on main exactly as the ledger states — both PRs remain open/draft.

4. Hypothesis

Given LocalReasoningBank.findSimilar() returning pattern objects whose confidence field is overwritten with the per-query cosine similarity score, when the candidate change stops overwriting confidence and instead returns the cosine score as a distinct similarity field (satisfying the pre-existing PatternMatch contract), then (a) distillLearning()'s gate should correctly test learned reliability rather than query-similarity, and (b) findSimilarPatterns() should return genuinely distinct confidence/similarity values, subject to: (1) memory-bridge.ts's search-ranking consumer, which wants query-relevance not reliability, is migrated to prefer the new similarity field; (2) endTrajectory()'s RL-update consumer (which already re-fetches pattern.confidence via bank.get()) is unaffected; (3) all existing tests remain green; (4) fully deterministic, $0 evaluation cost.

Frozen before evaluation; not modified after seeing results.

5. Evaluation Receipt

evaluated: accepted. Real evaluator: Vitest 4.1.8, deterministic, $0, zero LLM calls. New file: v3/@claude-flow/cli/__tests__/intelligence-confidence-similarity-conflation.test.ts (4 tests).

Baseline vs. candidate, isolated via git stash of the 2 source files (test kept): all 4 tests fail against baseline for the exact predicted mechanism (confidence overwritten by cosine 1.0/~0.45; DISTILL gate off by exactly the extra LoRA bump a wrongly-included pattern receives; public-API field conflation), pass on candidate.

Full @claude-flow/cli suite both ways: baseline 91 failed/149 passed/1 skipped (241 files), candidate 90 failed/150 passed/1 skipped — the only failed-file-set difference is the new test file itself; the other 90 are identical pre-existing environmental failures (unbuilt monorepo sibling packages, e.g. @claude-flow/neural Vite-resolution failures). tsc --noEmit: 463 pre-existing errors, byte-identical count/lines both ways.

6. Darwin Results

Skipped — binary correctness/API-contract fix, not a continuous parameter with a fitness gradient for Darwin's real interface (npx ruvector harness darwin --help, confirmed available, @metaharness/[email protected]) to search over.

7. Flywheel Evidence

No .claude-flow/flywheel/ state or signed @metaharness/flywheel bundle in this repo (confirmed via npx ruvector harness flywheel --help). Evidence retained as: 4 new tests, gist, this issue, the stash-isolated comparison, and the independent adversarial critique — consistent with every accepted night since 2026-08-18.

8. Reward Hack Check

No standalone reward-hack CLI reachable (@metaharness/weight-eft is a LoRA-distillation tool, not a diff scanner). Manual checklist, independently re-verified by the adversarial critic: no test weakened, no gold data touched, no cherry-picking (all 90 pre-existing failures disclosed), no seed manipulation (deterministic hand-crafted embeddings), $0 cost, discriminating power confirmed real (fails on baseline for the stated reason).

9. Security Review

Not security-sensitive: in-process ranking/confidence bookkeeping only, no new I/O/network/credential/filesystem surface. Independently confirmed by the adversarial critic. The memory-bridge.ts migration's ?? fallback chain was specifically checked against a || footgun on a legitimate 0-similarity value — ?? is correct.

10. Scan Findings: capabilities

Ruflo's swarm permission/audit system (v3/@claude-flow/cli/src/permission/{permission-set,permission-audit}.ts) writes grants but can only ever emit a 'granted' audit event (3 of 4 declared event kinds — checked/denied/revoked — are dead code, zero call sites in v3/), and nothing reads .swarm/permissions.jsonl back to gate any tool call, path, or network host. The module's own doc comment is candid it's "a METADATA + AUDIT layer, not a runtime sandbox," but the audit-log shape implies more enforcement than exists. Comparison: OpenAI Agents SDK's tool_input_guardrail/tool_output_guardrail decorators run inline in the actual tool-call path (Grade A). Not selected tonight (SCAN surface); flagged for a future capabilities night.

11. Scan Findings: memory

@claude-flow/memory's same-key upsert path is broken across all three store layers: generateMemoryId() never produces a deterministic/reusable ID, so AgentDBAdapter.store() never evicts the prior occupant of a (namespace,key) pair; SQLiteBackend's schema has no UNIQUE(namespace,key) constraint, so repeated writes accumulate duplicate rows; HybridBackend.getByKey() is hard-wired to the (unordered) SQLite path and can return a stale copy. A working precedent already exists in the same repo (memory-initializer.ts's real UNIQUE(namespace,key) + removeHNSWEntriesByKey()) that the newer package never carried forward. External comparison: Mem0's v3 ADD-only pipeline hits the documented same failure shape (Grade A, mem0ai/mem0 issues #4956/#5867/#4896). Not selected tonight; key-collision pairs are noted as a free, ground-truth-labeled corpus for PR #3232's still-open near-dup-threshold tuning gap.

12. Competitors Reviewed

LangGraph, Microsoft Agent Framework/AutoGen, CrewAI, OpenAI Agents SDK, Google ADK (mandated floor); Letta, GEPA/DSPy, ACE (specific non-floor comparisons); Mem0, Qdrant/Weaviate/Milvus/LanceDB (memory scan); OpenAI Agents SDK guardrails, LangGraph third-party permission middleware (capabilities scan). Full graded table in the gist — re-verifies "no mainstream framework ships human-gated evolutionary self-improvement" with 2026-dated primary sources; notes the ecosystem arguably regressed (OpenAI sunsetting Agent Builder/Evals 2026-11-30).

13. Gist

docs/dream-cycle/dream-gist-2026-09-12.md (this branch). No gist-creation tool is reachable in this execution environment (GitHub MCP tools cover issues/PRs/repos, not gists) — committed to docs/dream-cycle/ instead, matching established repo convention for these reports.

14. Witness

15. Recommendation

Merge the linked draft PR (human review required). Pick up priorDecay wiring (candidate 1) for a future intelligence night. Prioritize merging the growing backlog of ACCEPT-scoped/ACCEPT drafts (#3110, #3152, #3160, #3232 all still open). Consider a themed memory night combining the upsert-bug fix with PR #3232's near-dup corpus gap. Consider a capabilities night to wire real enforcement (or explicitly document audit-only status) for the swarm permission system.