dream(intelligence): #3475 expose fisherDecayRate via CLAUDE_FLOW_FISHER_DECAY env override (evaluated, ACCEPT-scoped)

[ruvnet]

1. Hypothesis

Given EWCConsolidator's fisherDecayRate (default 0.01) now correctly gating one consistent EMA direction across all three globalFisher-mutating sites (computeFisherMatrix(), recordGradient(), production-wired updateFisherFromConfidences() — post dream-cycle #3395), when a CLAUDE_FLOW_FISHER_DECAY env override is added — mirroring envPriorDecay()/envMaxUncertainty() in model-router.ts exactly (validated via Number.isFinite, range (0,1], silent fallback to the unchanged default on invalid/unset input) — then operators/future Dream Cycle nights should be able to experiment with different decay-rate values without a source change, subject to: (1) unset/invalid env leaves default behavior byte-identical; (2) an explicit constructor value still takes precedence; (3) all existing EWC tests remain green; (4) deterministic, $0 evaluation.

Deliberately scoped to making the parameter tunable, not finding a better value — see Darwin Results below for why the latter isn't claimed tonight.

Frozen before evaluation; not modified after seeing results.

2. Candidate

v3/@claude-flow/cli/src/memory/ewc-consolidation.ts (+21/-1, one file): adds envFisherDecayRate() (identical validation pattern to sibling envPriorDecay()), wires it as DEFAULT_EWC_CONFIG.fisherDecayRate = envFisherDecayRate() ?? 0.01. Confirmed by direct read that intelligence.ts:326's only production call site never passes an explicit fisherDecayRate, so the env var reaches production unshadowed. Includes a post-critique operator doc-note that the singleton reads the env var once at first construction (same behavior class as the already-merged envPriorDecay(), not a new regression).

v3/@claude-flow/cli/__tests__/ewc-fisher-decay-env-override.test.ts (new, 5 tests).

3. Evaluation Receipt

evaluated: accepted (ACCEPT-scoped). Real evaluator: Vitest 4.1.8, deterministic, $0, zero LLM calls, run against a freshly pnpm install-ed workspace (v3/'s pnpm workspace and @claude-flow/cli's dist/ were not pre-built in this checkout at session start).

Stash-isolated baseline (source reverted, test file kept): exactly 1/5 new tests fails against real baseline (0.158... not < 0.104... — the "wires the env var" discriminator), 4/5 pass both ways (correct non-discriminating controls). Candidate: 5/5 pass. Full @claude-flow/cli vitest suite: byte-identical 108 pre-existing failed-file set baseline vs. candidate (diffed, not just counted — the unbuilt-sibling-package/environmental class documented by nearly every prior dream-cycle night). tsc --noEmit: byte-identical 475 pre-existing errors, none in changed files. EWC-adjacent suites (ewc-fisher-ema-direction.test.ts, ewc-distill-confidence-gate.test.ts): 7/7 passing throughout, zero regression.

4. Baseline Comparison

	Baseline	Candidate
CLAUDE_FLOW_FISHER_DECAY=0.5, no explicit config	Ignored — default stays 0.01	Wired in — single-call retention ~50% vs. default's ~99%
Out-of-range (1.5) / non-numeric / zero env value	N/A (var doesn't exist yet)	Falls back to unchanged default (0.01)
Explicit { fisherDecayRate: 0.01 } with env=0.9 set	N/A	Explicit config wins (retention matches 0.01 behavior, not 0.9's)
Full package failed-file set	108 files	108 files (byte-identical)
tsc --noEmit	475 pre-existing errors	475 (byte-identical)
EWC-adjacent suites	7/7	7/7

5. Darwin Lineage

Skipped — for a different reason than every prior night. fisherDecayRate genuinely is now a real, bounded, correctly-directioned scalar — Darwin's real interface (npx ruvector harness darwin <config> --execute, confirmed available) could in principle sweep it, unlike every prior candidate (binary correctness fixes with no continuous parameter). Not invoked because it requires a fitness function tied to a real downstream outcome, which doesn't exist yet for Fisher-decay values specifically. Declined to fabricate a synthetic proxy just to produce a Darwin result — see the linked issue's §6 and the gist for the concrete named follow-up.

6. Flywheel Evidence

No .claude-flow/flywheel/ state or signed @metaharness/flywheel bundle applies to this deterministic config-fix (confirmed via npx ruvector harness flywheel --help). Evidence retained as: 5 new discriminating tests, the linked issue, the committed gist, the stash-isolated comparison, and the independent adversarial critique transcript.

7. Reward Hack Check

No standalone reward-hack CLI reachable. Manual checklist, independently re-verified by the adversarial critic: no existing test weakened (diff purely additive); no gold/expected data touched; no cherry-picking (full suite and full EWC-adjacent suite both reported); no seed manipulation (deterministic, re-run 4x with zero flakiness); zero cost.

8. Security Review

Not security-sensitive. envFisherDecayRate() uses the identical safe pattern as sibling envPriorDecay()/envMaxUncertainty() (Number() + Number.isFinite + range check, silent-safe fallback, no eval/exec). Independently confirmed: no request-body/API/MCP-input path can set CLAUDE_FLOW_FISHER_DECAY — operator-set process env only, same trust boundary as the existing CLAUDE_FLOW_PRIOR_DECAY.

Independent adversarial critique (separate subagent, no authoring context): independently re-read the diff, ran all 5 new tests, independently reproduced the stash-isolated baseline-vs-candidate proof itself, re-ran the existing EWC suites (7/7), and traced intelligence.ts:326 to confirm no shadowing call site exists. Verdict: CONFIRMED-WITH-CAVEATS → both caveats closed same-session (operator doc-note added for the sticky-singleton env-read behavior; the gist itself is now committed to this branch, since the critic initially couldn't verify its own Darwin-scoping reasoning from a checkout that didn't yet have the file).

9. Regression Analysis

Byte-identical failed-test-file set (108 files) and tsc --noEmit error set (475) baseline vs. candidate across the full @claude-flow/cli package. intelligence.ts is the only production caller of getEWCConsolidator() and never passes fisherDecayRate explicitly (grep-confirmed across hooks-tools.ts/intelligence.ts/index.ts), so the env var reaches production in every real call path without being shadowed.

10. ADR

None — scoped, additive config-exposure change to an already-decided mechanism (the EMA direction was already fixed and decided by #3395), matching repo convention for this exact fix class (cf. #3350's identical priorDecay exposure, which also created no ADR).

11. Research Gist

docs/dream-cycle/dream-gist-2026-09-27.md (this branch) — full SOTA report: 5-role parallel research fan-out (deep researcher, competitor analyst, capabilities scan, memory scan, independent architecture reviewer), 3 competitor-comparison tables (intelligence/capabilities/memory, 4-7 rows each with real sourced grades), 5 scored intelligence candidates (this one selected, 4.65/5), and scan findings for capabilities (a real, tested, but completely unwired PluginIntegrityVerifier — ADR-145 P1, still Proposed) and memory (two still-live, independently-confirmed inefficiencies distinct from already-merged fixes).

12. Issue

Closes #3475 (full 15-section report: ledger check with GitHub-verified fates for the last 8 nights plus a newly-confirmed 09-23/09-25/09-26 no-run gap, research, evaluation receipt, adversarial critique, reward-hack check, security review, capabilities/memory scan findings, competitor comparison, witness).

13. Witness

Field	Value
Session commit	a5adb5aadc8378169c7ad9ddee8c942bc3f9f9a1
Gist SHA-256 (pre-witness content)	5a2cf06014446a1dea65bab2b0a73ba24df8c8c8a08442116e0c2e704cca7011
Witness stamp	538729fe5dd14cc49fa2decc1f7475de51d05367b3e04740318c2a635513323a

Verifier procedure: fetch docs/dream-cycle/dream-gist-2026-09-27.md from this branch, strip the witness table's three filled values back to PENDING, SHA-256 it, concatenate with the session commit above, SHA-256 again — result must equal the witness stamp. Self-verified end-to-end against the committed file before this description was written.

14. Merge Policy

Human review required. Do not self-merge. Do not autonomously promote Flywheel state.

---

🤖 Generated with RuFlo

https://claude.ai/code/session_017NSCBLZggg81jUaa8RvZV7

---

Generated by Claude Code

[ruvnet]

Backlog sweep 2026-10-09: still valid against main (fisherDecayRate is hardcoded to 0.01 in cli/src/memory/ewc-consolidation.ts:158), but DIRTY, so it needs a rebase before it can be considered. It targets the cli EWC consolidator and is independent of the neural-mode EWC rework (#3879). It adds a new env var, so it also needs its audit-env-var-precedence entry and a minor-bump decision from the integrator.

[ruvnet]

Rebased onto current main (commit 63f4d97d, force-pushed) — the LEDGER.md conflict from the 2026-10-09 backlog sweep is resolved (kept both sides' rows/notes, nothing dropped), and confirmed clean of conflicts with the neural-mode EWC rework (#3879), which touches a different package entirely.

On the audit-env-var-precedence entry: that was already added in this PR's 3rd commit (7c7274a, pushed 2026-09-27, in response to CI's own first-run failure on this exact gate) — CLAUDE_FLOW_FISHER_DECAY is registered in KNOWN_ESCAPE_HATCHES right next to CLAUDE_FLOW_PRIOR_DECAY, whose exact justification it mirrors. Re-ran node scripts/audit-env-var-precedence.mjs against the rebased head just now — still passes. Re-ran the full EWC-adjacent test suite (12/12) against the rebased head too, no regressions.

Leaving the minor-version-bump call to you as the integrator, per your comment.

---

🤖 Generated with RuFlo

---

Generated by Claude Code