Status: authoritative · Version: 1.0 · Loaded by: src/episteme/_profile_audit.py
These are the five lexicons the profile-audit loop uses to compute the per-axis signatures the loop tests against the episodic record. They are deliberately modest — not exhaustive. A lexicon that enumerates every possible token invites false-positive matches and false-confidence coverage; a narrow one is honest about what it detects. (Original design rationale: the v0.11 Phase 12 spec; archived in git history pre-Event-54.)
Operator override. To extend or replace these lexicons for a specific project, create core/memory/global/phase_12_lexicon.md with the same heading structure. When the override file is present, the audit uses it instead of this default and records the fingerprint of the lexicon-in-use in every audit record — so a change in lexicon is visible in the record stream, not silent.
Match rule. Terms are matched case-insensitively against whitespace-collapsed content. Multi-word terms (e.g. "rolls back") are matched as regex word-boundary phrases. This matches the hot-path normalization pattern already used by core/hooks/reasoning_surface_guard.py.
The Goodhart limit. A sufficiently aware agent can pad its Reasoning Surfaces with failure-frame tokens to pass axis A's signature. The defense is D1 (multi-signature convergence — A's second signature is syntactic, not lexical) plus operator review, not a longer lexicon. Growing the lexicon to catch more shapes also grows the surface area for false positives; this file is version-stamped so any growth is deliberate and reviewable.
Used by: axis A · dominant_lens: failure-first — signature S1 (failure-frame ratio in unknowns + disconfirmation).
An operator with failure-first as their dominant lens names what could break before what should work. High density in this lexicon across episodic records predicts the claim; low density is a weak signal against.
- fails
- breaks
- rejects
- errors
- regresses
- blocks
- denies
- timeout
- timeouts
- invalidates
- violates
- leaks
- diverges
- corrupts
- exhausts
- stalls
- rolls back
- reverts
- aborts
- crashes
- loses data
- data loss
- hangs
- deadlocks
Used by: axis A · dominant_lens: failure-first — signature S1 (denominator of the ratio).
The counterpart. Dominance of success-frame language over failure-frame is evidence against the failure-first claim. Used as the denominator in the ratio: failure / (failure + success).
- succeeds
- works
- passes
- completes
- validates
- approves
- renders
- delivers
- ships
- finishes
- resolves
- returns ok
- returns success
- green
- healthy
Used by: axis B · noise_signature: status-pressure — signature S1 (buzzword density in core_question + knowns).
Fluent, impressive-sounding descriptors that survive by being unfalsifiable. An operator with a claimed susceptibility to status-pressure is predicted to emit these at above-baseline rates — and an operator with a claimed counter-screen against buzzword-leak (per core/memory/global/operator_profile.md) is held accountable for keeping the rate below baseline.
- robust
- seamless
- end-to-end
- enterprise-grade
- world-class
- best-in-class
- holistic
- synergy
- synergies
- cutting-edge
- battle-tested
- production-ready
- scalable
- resilient
- elegant
- idiomatic
- next-generation
- next-gen
- paradigm shift
- game-changer
- game-changing
- disruptive
Used by: axis 5 · explanation_depth: causal-chain — signature (density of causal connectives in knowns + assumptions).
Explanation-depth causal-chain predicts high density of because-chains. Reserved for checkpoint ≥ 3 implementation of this axis; lexicon available here so checkpoint 1 can fingerprint it in the audit record.
- because
- therefore
- so that
- as a result
- which causes
- due to
- hence
- consequently
- leads to
- results in
- thereby
- thus
- owing to
Used by: axis D · asymmetry_posture: loss-averse — signature S2 (rollback-path mention on irreversible-op records).
An operator with loss-averse posture on irreversible ops is predicted to mention rollback / recovery / abort paths in their Reasoning Surface's knowns or assumptions before committing. Absence is weak evidence against the claim.
- rollback
- revert
- undo
- abort
- back out
- restore from
- restore from backup
- recovery
- disaster recovery
- kill switch
- circuit breaker
- feature flag
- feature-flag
- emergency stop
Each audit record carries a lexicon_fingerprint field — the first 16 hex chars of the sha256 over the sorted, canonicalized lexicon contents. A change in any lexicon produces a different fingerprint, so readers of profile_audit.jsonl can detect lexicon drift between runs even if the filename didn't change. See src/episteme/_profile_audit.py::_lexicon_fingerprint for the canonicalization.