mtarcure/claude-vibe-squad/blind-rediscovery

Blind rediscovery — operational checklist

When chrono runs blind rediscovery (e.g. "does v1.3 reproduce findings from a prior daemon run?") the validation signal IS voice independence. If a fan-out voice reads a Stage N dossier that names prior findings by ID/class, the experiment is invalid.

This skill operationalizes the _blind/ vs _curated/ split codified in CLAUDE.md "Dossier authoring convention" so the discipline becomes a checklist, not just prose.

When to use

• Running Stage 3 hypothesis exploration on a target with EXISTING saved findings
• Validating a v1.3 system pass against legacy daemon work from a predecessor project
• Any case where "did the voices arrive at this independently?" is the core question

When NOT to use

• Resuming a known finding for PoC drafting (Stage 4) — voices need the prior work
• New target with no prior chrono history — there's nothing to contaminate
• Bulk-mechanical refactor / lint pass — no judgment, no voice independence to preserve

Directory layout

For a target with the slug <target>. The casing is load-bearing: dispatch_context_builder. enforce_blind_floor() keys the dispatch-level floor on exactly this path, and the maintainer's volume is case-insensitive, so a lowercase chrono/ here would resolve locally and fail on a case-sensitive filesystem.

$CHRONO_VAULT_ROOT/Chrono/dossiers/<target>/
├── _blind/                          # fan-out voices CAN read this
│   ├── stage-1-scope-<date>.md
│   ├── stage-2-codebase-map-<date>.md   # NO prior-finding refs
│   └── ...
├── _curated/                        # ONLY brain + operator read this
│   ├── prior-findings.md            # F1, F2, etc — the saved work
│   ├── stage-3-synthesis-<date>.md  # comparison-with-prior happens here
│   └── methodology-notes.md
├── raw-voices/                      # raw fan-out outputs (also voice-readable)
└── _resume.md                       # session-resume index, brain-readable

What the dispatcher already guarantees

Since 2026-08-17 the _blind/ directory is not only a convention — it is enforced. When a packet declares target: <target> and that dossier has a _blind/ directory, build_context forces memory_aperture: cold for every role whose specialist brief declares blind_discovery: true, regardless of what the packet asked for. The briefs under departments/*/specialists/ and shared/specialists/ are the roster; grep them for the key rather than trusting a list copied into prose. See shared/protocol.md § The blindness floor.

target: is free text and stays free text — the floor reads dossier slugs out of whatever the field contains (examplechain/<target>@0000aaaa, <target> / example-chain-node @ 1111bbbb, Target), so it fires on the live packet convention without the packet having to change vocabulary. A target: that names no blinded dossier leaves the aperture alone; it is never a dispatch error. Naming the dossier slug plainly is still the clearest thing to write.

That covers vault recall. It does not cover the two contamination paths below — dossier prose that names prior findings, and web search for prior art. Those remain the author's job, which is what the rest of this checklist is for. Set target: on every blind-rediscovery packet; the floor is target-scoped and has nothing to key on without it.

Pre-fan-out checklist

Before assembling the fan-out brief, verify ALL of:

• [ ] Every packet in the blind stage declares target: <target> so the dispatch-level floor applies

• [ ] _blind/ exists; voice-readable dossiers live there
• [ ] _curated/ exists; prior-findings + comparison docs live there
• [ ] _blind/stage-N-*.md files contain NO references to: prior finding IDs (fnd_*), prior attack classes (e.g. "stale BRIDGE_PROVIDER_ROLE"), or prior PoC sketches
• [ ] _blind/stage-N-*.md files DO contain: scope manifest, codebase map (architectural only), program rules, in-scope vuln class taxonomy
• [ ] Brief explicitly tells voices to read ONLY _blind/ paths
• [ ] Brief explicitly forbids reading _curated/ paths (use exact Do NOT read /path/to/_curated/ framing)
• [ ] Brief forbids web search for prior art on the target (prior submissions on the target platform, GitHub issues, etc.)

Post-fan-out synthesis (where contamination is allowed)

After voices return their independent outputs, the synthesis stage IS allowed to read _curated/ for comparison. Synthesis dossier lives in _curated/, not _blind/. Synthesis includes:

• Convergence table: which targets did N voices independently surface?
• Comparison to prior findings: did voices rediscover F1 / F2 etc.?
• Minority report: single-voice insights worth preserving
• Contamination notes: any _blind/ content that accidentally hinted at priors

If contamination is discovered mid-pass

If a voice's output reveals it had access to prior-finding hints (the "GPT-5.5 explicit caveat" pattern from 2026-04-30 Dexalot Stage 3):

1. Flag in the synthesis dossier — that voice's contaminated targets cannot count as independent rediscovery
2. The voice's INDEPENDENT targets (those not in priors) still count
3. Score system-validation against the clean independent targets only
4. Patch the _blind/ dossier authoring for next pass; update this checklist if the leak path was novel

Authoring rule for new stage dossiers

Every new dossier in _blind/ should pass this self-check:

"If a voice reading this dossier had no other knowledge of the target, would they end up with the same priors as a voice that has access to my full prior-findings file?"

If yes — the dossier is contaminated. Move the prior-finding refs to _curated/.

Cross-references

• CLAUDE.md "Dossier authoring convention — _blind/ vs _curated/" section
• feedback_chrono_blind_rediscovery_methodology (auto-memory) — durable rule
• docs/phase9-cleanup-2026-04-30.md — Dexalot Stage 3 contamination case study
• plugins/chrono-vault/skills/fan-out/SKILL.md — when fan-out vs single-dispatch