terrylica Emergent Resurrection — negative-knowledge archive
Self-Evolving Skill: If the failure taxonomy misses a mode, or a resurrection trigger type recurs, update the relevant section AND append to
references/evolution-log.md. Don't defer.
Failed hypotheses are not waste. They are negative knowledge that:
- Documents the investigation's boundaries
- Prevents re-exploring known dead ends
- Can come back when conditions change
This skill is the genetic-evolutionary mechanism for "don't rule out anything we have tried, as long as it can emerge through iterative process".
Why NOT just delete failed ideas
The session accumulated 17 null campaigns. Without negative-knowledge capture, we would:
- Re-explore the same dead ends in future sessions (wasted compute)
- Lose the meta-lesson each failure taught (wrong null, wrong scaling, wrong scope)
- Be unable to detect when conditions have shifted enough to retry
Failed attempts are cheaper to preserve than to re-run.
Failure-mode taxonomy
Each failed campaign fits one or more modes. Classification determines resurrection conditions.
| Failure mode | Description | Resurrection trigger |
|---|---|---|
null-insignificant |
Signal exists but below noise floor | Noise floor drops (longer backtest, better data, reduced cost) |
overfit-in-sample |
Strong IS, poor OOS | Cross-validation success on a different data regime or architectural redesign |
wrong-null-applied |
Correct signal, wrong null type broke it | Correct null method discovered (see Skill B §5 orthogonal cascade) |
cross-asset-failed |
Works on one asset, breaks on others | Same config trial on a DIFFERENT asset passing, OR regime detector enabling per-asset selection |
regime-conditional |
Full-history fails; edge concentrated in specific regime | Causal walk-forward regime classification + train/test both positive within regime + null z>3 |
label-leaked |
Features/labels violated causality | Causality re-verified (bars[:i] exclusive) and retest OK |
agent-overestimation |
Agent's estimate > reality | External oracle or per-bar replay validates a scaled-down version |
grid-degenerate |
Parameter grid too narrow | Widen grid, OR sensitivity analysis reveals dormant dimension |
dormant-low-precedence |
Valid but lower priority | Higher-priority frontier cleared, revisit |
falsified-comprehensive |
Tested exhaustively; impossible under stated assumptions | Assumptions change (new asset class, new microstructure, regime shift) |
Each archived file tags one or more of these.
Archive structure
plugins/crucible/skills/d-emergent-resurrection/references/
├── archive/
│ ├── <YYYY-MM-DD>-<slug>.md Dormant/failed campaign record
│ └── ...
├── falsified/
│ └── <YYYY-MM-DD>-<slug>.md Permanently falsified (highest bar for resurrection)
└── exhumation-log.jsonl Append-only: when ideas moved, why
The distinction:
archive/= dormant, not dead. Awaiting the right conditions.falsified/= tested exhaustively under clear assumptions. Resurrection requires assumption change.
Archive file template
Use this frontmatter structure when moving a failed idea to archive:
---
id: "UNIQUE-ID"
title: "Short descriptive title"
status: "dormant" # or "falsified"
failure_mode: ["null-insignificant"] # list; can be multiple
confidence: 0.95 # how sure we are it's dead NOW
date_dormant: YYYY-MM-DD
verdict: |
One-paragraph summary of what was tested and why it failed.
linked_ledger_entries:
- "FINDING-ID-IN-EVOLUTION-JSONL"
linked_audits:
- "findings/evolution/audits/YYYY-MM-DD-slug/verdict.md"
resurrect_if:
- condition: "noise_floor_drops"
trigger_metric: "trade_level_volatility"
threshold: "< 1.2 bps (was 2.1 bps at dormancy)"
requires: "6+ month new backtest reducing noise > 30%"
- condition: "new_asset_regime"
trigger_metric: "asset_correlation_cluster"
threshold: "Jaccard(newasset_patterns, eur_patterns) > 0.65"
requires: "Trial on new asset with fold-shuffle p < 0.05"
whats_blocked:
- "List of specific reasons it was shelved"
next_agent_checklist:
- [ ] If noise_floor drops > 30%: re-run config on latest data
- [ ] Do NOT assume dead — it's 'not now, but perhaps next regime'
---
# Title
<detailed description of why it failed, what was tested, what would be needed to resurrect>
The 3-layer exhumation process
Resurrection is deliberate. Never auto-resurrect.
Layer 1 — Autonomous flag: periodically (can be a cron job or session-start check), scan archive/* for files whose resurrect_if conditions might be currently met. Tag with CANDIDATE_FOR_RESURRECTION_<date>. This ONLY flags; it does NOT resurrect.
Layer 2 — Agent review: next research session or autonomous agent reviews tagged candidates. Picks ONE to trial (prevents thrashing). Runs 1-2 folds to check if signal has genuinely re-emerged. Either:
- Promotes to active campaign (
status: "active", new audit folder) - Updates archive file's frontmatter: "false positive at , conditions looked met but retry failed"
Layer 3 — Human sign-off: before deploying anything resurrected to production (or giving it resources beyond initial re-trial), human reviews exhumation report. Confirms the condition-change is real, not a data artifact.
When to archive (triggers)
During a research session, move to archive/ when:
- A serial-gate failure killed the signal (Skill B §2, Gate C — the Irrecoverable one)
- A shuffled-null test decisively rejected (z < 1 and p > 0.9)
- 20+ parameter combos all failed → the whole approach gets archived, not each combo
- Cross-asset Gate D fails → archive the "universal" interpretation; keep asset-specific as scoped-down live version (see Skill C §3)
Archive ≠ silence. Write the file; append to exhumation-log.jsonl:
{
"date": "YYYY-MM-DD",
"action": "archive",
"id": "...",
"reason": "...",
"from_audit": "..."
}
When NOT to archive
- A finding survives all gates but produces modest economics — that's
live-limited-economics, not dormant; keep active and look for ensemble partners - A finding partially fails one gate — see Skill C §3 (scope narrows, doesn't kill)
- A method used in a campaign that gave null results — the method itself isn't archived; only the specific hypothesis
- User interest waned mid-campaign — that's not a failure mode; mark as
pausedelsewhere
Population-level resurrection signals
Beyond individual-file resurrect_if, watch for population-level conditions:
- Multiple failed campaigns share a failure mode → the common cause may be addressable (e.g., all 5 used wrong null type → invent new null → resurrect all 5)
- New method/technology emerges → old failures under old methods may flip
- Asset microstructure shifts → asset-specific failures may reverse
Integration with the ledger
findings/evolution/evolution.jsonl entries gain a dormant_refs field:
{
"id": "FINDING-ID",
"status": "null",
"dormant_refs": [
{
"archive_file": "plugins/crucible/skills/d-emergent-resurrection/references/archive/<slug>.md",
"failure_mode": ["null-insignificant"],
"resurrect_conditions": ["noise_floor_drops", "new_asset_regime"]
}
]
}
Bidirectional: ledger → archive file, archive file → ledger.
Current archive (seed state)
Session ca9d7ffa-ef5a-41d0-94c8-56f113a132f2 produced:
- 17 null campaigns → each should have an archive file (SEEDING STATE: archive files not yet backfilled; do this when touching individual campaign verdicts)
- 1 positive finding (
NGRAM3FU-STRADDLE-001-FULL-STACK) → remains active, NOT archived
Backfilling 17 archive entries is a future task; initial references/archive/ is intentionally empty until each campaign's verdict is reviewed and archive-written.
Post-Execution Reflection
After invoking this skill:
- Did archiving a failed idea happen cleanly? If the file was too sparse or too detailed, update the template above.
- Did
resurrect_if:conditions fire (layer 1)? Were they appropriately specific, or too vague? Refine. - New failure mode not in the taxonomy? Add a row; log to
references/evolution-log.md. - A previously-archived idea resurrected successfully? Document the exhumation in
exhumation-log.jsonlAND note the condition pattern that worked — this is meta-knowledge.