Glossary Drift Audit — 2026-06-20

Method: Swept docs/ontology/, docs/proposals/, docs/operations/, CLAUDE.md, and AGENTS.md for a fixed list of load-bearing terms. For each term, recorded every distinct sense (the question it answers) and which file pins that sense. A term is flagged DRIFT when the same word answers a different question in different places without a marker telling the reader which sense is meant.

Outcome: This audit seeded docs/ontology/glossary.md (the durable, question-keyed glossary). This file is the point-in-time evidence and the recommended-fix list. Re-run and re-date when the vocabulary moves.

Addendum (2026-06-20b — code-grounded pass). The first sweep read only docs. A follow-up read of the code (governance_core/dynamics.py, src/grounding/free_energy.py, src/behavioral_assessment.py, docs/EISV_COMPUTATION.md) found two collisions the docs-only sweep missed and one wrong claim to retract:

NEW basin (DRIFT, confirmed in code): governance_core/dynamics.py::check_basin (bistable attractor basin, research-lens ODE) vs. src/behavioral_assessment.py::_basin_health_gate / config.governance_config.BASIN_HIGH (a static health band — the verdict-driving gate). Only the second drives decisions, and it is a configured box, not an attractor. Previously listed as single-sense; promoted to high-risk homonym. A runtime glossary (src/governance_glossary.py::explain_basin, #428) resolves the health-band sense — keep it consistent with the doc glossary.
NEW free energy (DRIFT, code): ODE V ("like Helmholtz free energy", a signed integrator) vs. src/grounding/free_energy.py target E=−F (variational, Tier-1 explicitly stubbed).
RETRACTED claim: the Rosetta skeleton's "physics register is aspirational, all candidate, zero earned" was wrong. Three rows are ◐ research-lens (implemented in the parallel ODE / grounding tiers, honestly provenance-tagged, but not wired to the verdict path). The real gap is wiring built physics onto the verdict path, not building it. Notably grounding/free_energy.py already enforces this audit's own "name nothing more rigorous than it is" guardrail at runtime (tiered estimator, source= tags, FEP stubbed).

Lesson for future sweeps: read the code, not only the proposal docs. Drift between a term's doc sense and its code symbol is invisible to a docs-only pass.

Headline: Three words carry the most collision risk — substrate (3 senses), fingerprint (3 senses), surface (2 senses) — because in each case more than one sense is genuinely load-bearing and at least two senses live close together (sometimes in the same document). These are not errors to "fix" by renaming; they are unmarked homonyms to make legible.

Collision summary (highest risk first)

Term	Senses	Co-located?	Verdict	Recommended fix
`substrate`	inference / deployment-identity / runtime-scheduler	Across 3 docs	DRIFT	Qualify in cross-doc prose: `substrate (inference\\|deployment\\|runtime)`. Glossary high-risk entry added.
`fingerprint`	transport ip:ua / behavioral-lineage / finding-dedup	First two in same doc (`identity.md`)	DRIFT	Always qualify in `identity.md`; the weak-proof vs strongest-earned-layer split is dangerous unmarked.
`surface`	lease target / tool set	Across 2 docs (+ a CLAUDE.md single-writer usage that is the lease sense)	DRIFT	Use `surface (lease)` vs `surface (tool)`; note the CLAUDE.md "single-writer surface" is the lease sense pre-plane.
`fork`	sibling-locus / identity-lineage	Same doc (`r6-…`) — already disambiguated by enum	DRIFT (managed)	Keep the `episode_fork_kind` enum discipline; never ship bare `is_fork`. Already R6's stated position.
`harness`	agent-body / lifecycle-wrapper / test-rig	Across docs; registers rarely collide	SOFT	Agent-body is canonical; flag the informal senses only when prose mixes them. Real issue is an incomplete answer set (see gaps).
`continuity`	layered concept / `continuity_token` mechanism	Same doc	SOFT	Write the full `continuity_token` for the mechanism; bare "continuity" = the concept.
`lineage`	causal ancestry / conversation-thread	Same doc, already distinct	SOFT	Keep `thread_id` out of ancestry sentences.

Per-term sweep

substrate — DRIFT (3 senses)

inference — "What inference substrate generates behavior?" — harness-substrate-plurality.md, layer table ("Model / substrate").
deployment / identity layer — "What persistent hardware/disk/DB/config survives restart and can earn continuity?" — identity.md five-layer table ("Substrate | persistent hardware, disk, DB, configuration") and the Substrate-Earned Identity appendix (Lumen's Pi).
runtime / scheduler — "What execution model runs the work?" — CLAUDE.md / AGENTS.md Substrate-Tax section ("substrates with per-process scheduling and protocol-level connection checkout (e.g., BEAM / db_connection)").

Three orthogonal axes; an agent can vary one while holding the others. This is the worst offender because all three are load-bearing and none is wrong.

fingerprint — DRIFT (3 senses)

transport (ip:ua) — weak sticky-resume pin — identity.md (recent_onboard:<ip:ua>).
behavioral / lineage — EISV match against claimed lineage — identity.md ("behavior diverges from lineage fingerprint"); R1.
finding-dedup (sha256) — Watcher/CI dedup — CLAUDE.md Watcher section (--resolve <fingerprint>); docs/operations/ci-issue-surfacing.md.

The first two co-locate in identity.md and sit at opposite ends of the assurance scale (weakest heuristic vs strongest earned layer). Highest in-doc hazard.

surface — DRIFT (2 senses + 1 aligned usage)

lease — shared mutation target — beam-coordination-kernel.md.
tool — available actions — harness-substrate-plurality.md.
aligned: "single-writer surface" in CLAUDE.md/AGENTS.md is the lease sense applied to source control before the plane mediates it — intentional continuity, not drift.

fork — DRIFT, but managed

sibling locus — same registry UUID, fresh process-instance, no child — r6-episode-fork-response-shape.md (episode_fork_kind = sibling_locus).
identity lineage — distinct child UUID + parent_agent_id/spawn_reason — same doc (identity_lineage_fork).

R6 already treats a bare is_fork as "too compressed" and splits it. The audit endorses keeping that discipline; the risk is regression to an unqualified "fork" in new prose/fields.

harness — SOFT

agent body (canonical) — "What body/interface mediates action?" — harness-substrate-plurality.md.
lifecycle wrapper (informal) — "Claude Code runs through a plugin-style harness" — CLAUDE.md.
test rig (informal) — scripts/dev/calibration_harness/, deep-research skill.

Registers rarely collide. The substantive finding is not the homonym but the incomplete answer set for the agent-body sense (no value for a BEAM-resident agent) — recorded as an open gap in the glossary.

continuity — SOFT

layered concept — identity.md five layers.
continuity_token mechanism — advanced same-live-process rebind, largely retired for cross-process use — identity.md.

Mechanism borrowed the concept's name. Write the full token name for the mechanism.

lineage — SOFT

causal ancestry — parent_agent_id; "inherits work from, not is identical to" — identity.md.
conversation-thread — thread_id "names logical conversation/history lineage" — harness-substrate-plurality.md.

Both feed the provenance envelope; already distinct, keep them apart.

Single-sense (no drift at audit time)

process-instance, registry (identity layer), transport, lease, handoff, basin, locus, episode, affordance_state, assurance, governance_mode, typed absence, proof of life/heartbeat, provenance envelope. Each currently answers one question. episode and locus reappear as qualifiers inside the fork classification but do not change sense there. Listed in the glossary so a future split is visible against this baseline.

Recommendations

Adopt the question-keyed glossary (glossary.md) as the pointer of record; stop re-defining these terms inline in new docs — link instead.
Qualify the three high-risk homonyms in cross-doc prose. Especially fingerprint inside identity.md, where the weak/strong senses are one paragraph apart.
Do not rename to resolve drift. Qualify. Renaming a live term invents vocabulary nobody adopts; qualifying makes the existing collision legible — the rule harness-substrate-plurality.md already half-follows.
Coin a harness (agent body) value for BEAM-resident agents before Wave 3 handler-dispatch lands, so the taxonomy doesn't backfill the gap by overloading dispatch or lumen.
Re-run this sweep when a new proposal cluster matures (next likely sources of new terms: Plexus scope, lease-plane Phase B, the principal/ operator-delegate track). Date the next audit and append it beside this one.