Mission
Prune semantic search space without turning stale failures into permanent dogma.
The structural boundary is the passive definition:
${CODEX_HOME:-$HOME/.codex}/skills/negative-ledger/definitions/ledger/negative-evidence-protocol.json
Its events logical slot in $ledger-resolved custody is canonical. The Ledger
skill alone owns physical storage location, worktree-family registration,
initialization, and migration policy. Do not derive a separate history from each
checkout. Worktrees share evidence, not automatic applicability or exclusions.
The derived memory-admission channel remains:
~/.codex/memories/extensions/negative-ledger/notes/*.md
$negative-ledger owns the meaning and lifecycle of negative-evidence state.
Ledger replays/projects the declared protocol without semantic authority.
memory-note transports an immutable projection to Phase 2, which decides
whether to compile a route constraint, routing trigger, or reusable memory skill.
Never use memory notes as the operational route gate. For accepted admission,
load $memory-source-notes before invoking run_memory_note_tool.
Trigger Cues
$negative-ledger;- failed/no-effect attempts, reverts, benchmark/test regressions;
- repeated semantic routes or same-cluster retries;
- strategy pivots abandoning a concrete route future work might repeat;
- "what have we already tried?" or "do not retry this route";
- route reopening after artifact-state changes;
- fixed-point or review-governor negative evidence;
- memory admission of active/stale/reopened/superseded
NEG-*projections.
Activation Policy
Activation is broad; capture is narrow.
Invoke implicitly when work may change route selection because a concrete route failed, had no effect, regressed a signal, was reverted, was rejected by current proof/review evidence, or is about to be retried under the same cluster. Do not wait for the user to name the skill.
Before selecting a route resembling a prior failure, project canonical
route-gate. A recalled learning may trigger the check, not block directly.
After a material strategy pivot, regression-confirmed revert, or closeout leaving
a failed route likely to recur, evaluate capture. A transient red test, syntax
error, first incomplete implementation, or discarded typo is no-op unless it
exposes a durable failed hypothesis that changes future routing.
Retain exactly one internal disposition for each material activation:
mapped current ledger checked; no write required
captured witnessed negative evidence appended
transitioned existing NEG record changed lifecycle state
no-op activation evaluated; evidence was not durable or route-shaping
blocked context/store unavailable or invalid, or an active applicable exclusion matched
A material closeout without failed-route semantics does not activate this source. Do not resolve storage, query, doctor, or capture to manufacture a no-op receipt.
Valid Statuses
capture_candidate
need-evidence
unknown
active
accepted_risk
stale
reopened
superseded
Only active may block: witness evidence, valid exclusion scope, current artifact
applicability, and exact-enough route/cluster matching are all required. Fuzzy or
lexical overlap is suggest-only.
Selected guidance
Evaluate from task evidence before resolving storage. A transient failure without
a durable route-shaping hypothesis needs only this entrypoint: retain no-op and
return. Do not read operation manuals merely because the skill activated.
| Selected work | Required guidance before that decision or effect |
|---|---|
| Canonical recall, projection, or any native operation | store.md; $ledger owns bootstrap and custody. |
| Select a retry resembling prior failure | Store guidance and route-gate.md; read and interpret the canonical gate before selection. |
| Capture or promote negative evidence, or challenge an exclusion's scope | capture.md; scope challenge precedes capture/promotion. Canonical operations also require store guidance. |
| Transition, reopen, narrow, or supersede a record | lifecycle.md, with its selected capture/scope guidance; canonical operations also require store guidance. |
| Admit or update a derived memory snapshot | memory-admission.md and store guidance; load $memory-source-notes before transport. |
A selected reference is mandatory, not optional background. Several operations may apply to one task; load each before its first governed decision and reuse unchanged guidance. Selection does not grant mutation, retry, admission, or publication authority. Explicit read-only/no-file limits remain binding.
An active exact applicable exclusion stays binding until its legal source-owned lifecycle changes it. Do not bypass it by calling a retry a test, sandbox, or experiment. A failed context or gate lookup is unavailable evidence, not an empty history; continue only independent authorized work.
Canonical Store and CLI
Read store.md before any canonical operation.
Route-Gate Workflow
Read route-gate.md before selecting a resembling retry.
Capture Workflow
Read capture.md before capture or promotion.
Exclusion Scope Challenge
Read the scope challenge before capture/promotion or deciding a challenged exclusion's breadth. A failed implementation alone does not justify excluding a strategy.
Lifecycle Transitions
Read lifecycle.md before transitions or reopening.
Memory Admission Gate
Read memory-admission.md before deciding admission; canonical capture does not automatically qualify.
Admission Workflow
After accepted admission, follow the admission workflow.
Proof Lines
Store guidance owns the existing reporting forms. Report canonical writes and derived admissions separately; a later admission failure never rolls back a successful canonical write.
Learnings Relationship
Learnings are historical candidate evidence, not route exclusions. Retired source
stores are addressed only during owner-authorized recovery. Verify current
applicability and promote qualifying evidence through canonical capture.
Guardrails
- Do not record vibes or turn one failed implementation into a broad strategy ban.
- Do not block from fuzzy matches or use stale benchmarks without applicability.
- Do not treat absence of an entry as novelty proof.
- Do not bypass Ledger, hand-edit records, or use a checkout-local fallback.
- Do not let memory notes outrank canonical custody or write compiled memory.
- Do not publish incomplete projections or capture every transient red test.
- Failed context, route-gate, memory-note, or doctor boundaries fail closed.
- Do not invoke sibling sources merely because Negative Ledger activated.
For exclusion-scope changes, use Negative Ledger cases in validation-probes.md. These are package validation, not a runtime dependency or activation of Learnings.