Glossary Drift Audit — 2026-09-18

Method: Re-read every factual claim in docs/ontology/glossary.md against the code and docs it cites, at master f0261eaf. Then swept the runtime glossary (src/governance_glossary.py), the coherence and state-storage modules, and docs/EISV_COMPUTATION.md for load-bearing words that answer more than one question but have no glossary entry. Same flagging rule as the 2026-06-20 audit: a term is DRIFT when a word answers different questions in different places with no marker saying which sense is meant, or when the glossary's statement of a source no longer matches that source.

Mechanical gate: scripts/dev/check_glossary_drift.py passed before and after this sweep. Every finding below is semantic, which is the part the gate cannot see.

Outcome: Findings 1–4 are fixed in glossary.md in the same change. This file is the point-in-time evidence.

Findings

1. Glossary quoted a sentence its source had retracted (DRIFT, high)

glossary.md quoted EISV_COMPUTATION.md as saying the ODE "runs in parallel and does NOT drive verdicts", and derived its headline reading from it ("two loops", "the physics is built but not wired"). EISV_COMPUTATION.md was corrected on 2026-08-29 (#1992): the ODE is not the verdict owner, but the behavioral sensor still blends the legacy C(V) term into 25–30% of E and 30–40% of I. The old line "describes authority, not causal independence." The runtime glossary had already been updated to match (EISV_DIMENSIONS["E"]: "still contains a legacy ODE control-feedback level term"); the doc glossary was the stale copy.

The first draft of this fix overcorrected: it said the C(V) term was the ODE's only verdict-path influence. Review found three more channels: the Φ cold-start prior that owns check-ins 1–2 is evaluated on the ODE state (src/monitor_phi.py); the ODE-derived regime inputs to the behavioral sensor; and the confidence fallback, 55% of whose base is C(V_ODE) (src/confidence.py).

Fix: the Rosetta correction section now carries a dated second correction: the ODE owns the cold-start verdict, does not own the post-warmup verdict, and reaches it after warmup through compatibility couplings. The ◐ research-lens definition and the free energy (ODE V) note were brought in line with it.

2. Physics-register word in a published file (DRIFT, medium)

The Rosetta correction called governance_core/dynamics.py a "thermodynamic ODE" twice. glossary.md is rendered into the public-site glossary viewer, and its own guardrail is "name nothing more rigorous than it is".

Fix: "dynamical-systems ODE"; the second occurrence was removed in the rewritten passage.

3. Unmarked homonyms missing from the glossary (DRIFT, medium)

• V — three quantities: behavioral EMA of E−I (metrics['V'] once warm), the ODE damped integral (which coherence always reads), and the embodied instantaneous imbalance. Plus two naming layers (Valence in reader docs, Void in code/DB/papers) that are not extra quantities.
• coherence — three producers (legacy_tanh_v ODE control feedback, which is deployed; manifold, canonical only under the off-by-default UNITARES_GROUNDING_APPLY; behavioral_assessment). src/coherence_provenance.py already labels them; the glossary had no entry.
• drift — server-computed deviation components, the caller-reported ethical_drift input slots, and the runtime description of S. The Rosetta table listed drift as "already standard", which is true for public naming and hides the self-attested/measured split in technical prose.
• entropy — the deployed S axis is a heuristic blend labelled "Entropy"; the paper's target is response-distribution entropy H, which is not computed on the primary path. Same deployed-vs-target split as free energy.

Fix: four new entries under "High-risk homonyms". Persisted column names (core.agent_state.entropy holds S, volatility holds V, E only in state_json.E) are single-sense names, not homonyms, so they went into a new Persistence false friends note rather than the homonym table.

4. Open gap overstated the lease schema (DRIFT, low)

The BEAM-resident-harness gap called hermes / claude_code / codex / dispatch / lumen "the lease enum". The only place that list appears is beam-coordination-kernel.md, as an open "etc." list; no closed enum enforcing it was found (the lease-plane source itself was not located during this sweep). The gap is still real: no value has been named for a BEAM-resident body.

Fix: reworded as "documented values … an open list rather than a closed enum".

5. Stale code comment on the Φ default (code drift, low)

src/governance_monitor.py, at the resolve_verdict_risk call site, said "Φ floors verdict/risk by default". The flag UNITARES_PHI_TELEMETRY_ONLY defaults to on (config/governance_config.py::phi_telemetry_only), so by default Φ is telemetry. The glossary's Φ entry was already correct.

Fix: comment corrected.

6. Stale bounds in the ODE module docstring (code drift, low)

governance_core/dynamics.py gave V as [-2,2] and S as [0,2]. governance_core/parameters.py DynamicsParams clamps V to [-1, 1] and S to [0.001, 1]; the "~2.0" in dynamics.py is the width of the V range, not a bound. Live rows agree: over 30 days no ODE V left [-1, 1]. The first draft of this change copied [-2, 2] into the glossary.

Fix: docstring bounds corrected; the glossary V rows now say both V senses share [-1, 1], so range cannot tell them apart.

Review round

Three independent reviews ran on the first draft: a conceptual review against the glossary's own rules, a source-accuracy review of every checkable claim, and a read-only check against the live database and configuration. The source-accuracy review found no factual errors. The conceptual review produced the finding-1 overcorrection, the reclassification of persisted column names, the entropy homonym, and the removal of the V (coherence input) row (a routing fact, not a separate question). The live check confirmed the storage mapping (7,155 of 7,155 recent rows), that 99.76% of stored rows over 7 days carry coherence_form = legacy_tanh_v (the rest are synthetic fixtures), and the live flag state; it also found finding 6.

Re-verified, no drift

• Φ is telemetry by default and in the live configuration; the cold-start prior (phi_cold_start) remains its verdict-shaped role.
• src/grounding/free_energy.py tier-1 still raises NotImplementedError.
• BASIN_E_HEALTHY = 0.60 == BASIN_HIGH.E_min; the runtime BASINS entry still describes the health-band sense, consistent with the basin entry.
• governance_core/dynamics.py still describes V as "like Helmholtz free energy".
• Every canonical-source document named in the glossary exists.

Not swept

Registers other than ops and fep were not re-read against their sources; the single-sense table was checked only for cited-file existence. Re-run and re-date when the vocabulary moves.