UNITARES Glossary — Terms Keyed by the Question They Answer

Created: June 20, 2026 Status: Living document. Expect churn; the field is inventing terms and so are we. Companion to: docs/ontology/identity.md, docs/ontology/harness-substrate-plurality.md, docs/ontology/beam-coordination-kernel.md Audit trail: docs/ontology/glossary-drift-audit-2026-06-20.md (point-in-time sweep that seeded this file); docs/ontology/glossary-drift-audit-2026-09-18.md (re-sweep: ODE coupling correction; V / coherence / drift / entropy homonyms; persistence false friends)

How to read this glossary

A term here is defined by the question it answers, not by a list of examples. This is deliberate. A term defined by its discriminating question survives redefinition; a term defined by examples rots the moment a new example straddles the boundary. (The canonical example: "harness" defined as "Hermes / Claude Code / Codex" breaks the instant BEAM starts hosting an agent loop — see the open gap below. The question — "what body mediates this agent's action?" — does not break; it just gets a new answer.)

The layer table in harness-substrate-plurality.md already works this way ("Harness | what body/interface mediates action?"). This file generalizes that discipline across the whole vocabulary.

How to add or change a term

1. State the question it answers, in one line, before any examples.
2. When a term splits, do not rename — disambiguate. If the same word starts answering two different questions, give each sense a parenthetical qualifier (substrate (inference) vs substrate (deployment)) and let both coexist explicitly. Renaming invents a word nobody adopts; qualifying makes the existing collision legible. This is the rule harness-substrate-plurality.md half-follows already.
3. Point at the canonical doc, don't re-define inline. Inline re-definitions are how drift starts.
4. A homonym is not a bug to fix. Three legitimate senses of "substrate" are fine. The bug is an unmarked homonym — a reader binding the wrong sense because nothing told them which question was being asked.

High-risk homonyms (same word, different questions)

These are the words where a reader can silently bind the wrong sense. Each is load-bearing in more than one sense; none of the senses is wrong.

substrate — three distinct questions

These three are orthogonal axes. An agent can switch substrate (inference) (swap model) while holding substrate (deployment) constant (same Pi), running on a substrate (runtime) with a different scheduling model (BEAM). Always qualify which axis you mean in cross-doc prose.

fingerprint — three distinct questions

The first two are identity machinery and live in the same document; the third is operational tooling. The transport vs. behavioral split is the dangerous one because both appear in identity.md: one is a weak heuristic proof, the other is the strongest earned layer. Never let "fingerprint" stand unqualified there.

surface — two distinct questions

Note the near-collision with a third usage: "single-writer surface" in CLAUDE.md/AGENTS.md ("Before Starting Work on a Single-Writer Surface") is surface (lease) applied to source control — the same sense, just not yet mediated by the lease plane. That continuity is intentional; it's why the lease plane's first surface class is repo_path.

harness — one question, an under-enumerated answer set

The agent-body sense is the load-bearing ontology term. The other two are colloquial and rarely collide in practice, but flag them when prose mixes registers. The real issue is not a homonym — it's that the answer set for harness (agent body) is incomplete. See the open gap below.

Preferred term: body is the precise synonym for the agent-body sense and is already in use (harness-substrate-plurality.md: "powerful fixed bodies", "variable body"). The field's "harness" descends from test harness — a scaffold that drives code from outside — so it natively leans toward the scaffold/wrapper sense (rows 2–3), which pulls against the inhabited body sense (row 1). Prefer body when you mean what mediates an agent's action; reserve harness for the scaffold/lifecycle-wrapper sense. This assigns the two words we already use to the two questions they each answer best, rather than coining anything new.

fork — two questions the bare word conflates

R6 already disambiguates via the episode_fork_kind enum and the identity_lineage_fork boolean precisely because a bare is_fork is "too compressed." Never ship a field or sentence with unqualified "fork."

continuity — concept vs. mechanism

The token is a deprecated implementation that borrowed the concept's name. When you write "continuity," mean the layered concept unless you write the full continuity_token.

lineage — ancestry vs. conversation thread (same doc, both senses)

Both appear in the provenance envelope; keep thread_id out of any sentence about causal ancestry.

Φ / phi — UNITARES quantity vs. IIT quantity (the highest-stakes homonym here)

⛔ UNITARES does not compute IIT's Φ and makes no claim on it. The two share a letter and nothing else: ours is a scalar distance from a configured setpoint, and it is deliberately not a measure of integration, irreducibility, or experience. A reader who binds the wrong sense will hear a consciousness claim in a regime-visualization number.

Two further properties a peer needs before consuming metrics.phi off the wire:

• Telemetry, not authority. UNITARES_PHI_TELEMETRY_ONLY=1 is live, so Φ is not a verdict floor. Its remaining verdict-shaped role is the cold-start path (phi_cold_start), which is a fallback prior, not a behavioral read.
• Expressive, not validated. Φ has the widest dynamic range of any field and separates regimes that categorical verdicts collapse. That is sensitivity. The outcome evidence runs the other way, and wide swings plus sign flips plus no predictive lift is the signature of a number that looks informative on a dashboard while carrying no outcome signal. Keep it for visualization and regime reads; ⛔never cite it as validated discrimination.

Per rule 2 above, the fix is disambiguation, not renaming: metrics.phi is a published wire field and consumers depend on it. Qualify the sense, keep the key.

basin — attractor vs. health band (confirmed at code level, 2026-06-20)

This is the dangerous kind: only the second sense drives verdicts, and it is a static configured box (BASIN_E_HEALTHY=0.60, …), not an attractor derived from dynamics — even though it borrows the attractor word. The real attractor basin exists in the ODE but does not gate verdicts. A runtime glossary (src/governance_glossary.py::explain_basin, #428) resolves the health-band sense for users; keep it consistent with this entry.

free energy — ODE accumulator vs. variational −F (code level)

The first ships and does not own the post-warmup verdict, but it is not inert: through coherence it feeds part of behavioral E and I (see the coherence entry below). The second is the target E semantics and is honestly not-yet-implemented. Don't let "free energy" stand unqualified across the two.

V — three quantities share one letter (code level)

⛔ metrics['V'] and the V that coherence is computed from can be different variables. Coherence always reads the ODE V (governance_core/coherence.py, "depends on the signed void integral V"). metrics['V'] is the behavioral V only when primary_eisv_source == "behavioral" (behavioral confidence ≥ 0.3); before warmup it is ode_fallback. Read primary_eisv_source before comparing the two. Two naming layers sit on top of this, and neither is a fourth quantity: reader-facing docs call V Valence, while code, the DB and the papers keep Void (void_*, void_active, void_alert); and the persisted column holding V is core.agent_state.volatility (see Persistence false friends below). The violation taxonomy's "Void Compliance" class is an unrelated word (null-and-void).

coherence — one field, three producers (code level)

The dangerous reading is the everyday one: "coherence" sounds like a health measure, and the deployed producer is a controller output. Earlier docstrings in both coherence modules described the manifold form as canonical; both were corrected and now mark it non-canonical as deployed. Read coherence_source / coherence_role on responses, or state_json.coherence_form on stored rows, and never infer the producer from the scalar's range. Because behavioral E and I still blend in the legacy C(V) term (docs/EISV_COMPUTATION.md), this is also one of the channels through which the ODE reaches the verdict path after warmup.

drift — deviation, self-report, and the S axis

The third row is partly part-of rather than a pure homonym: S_obs consumes the deviation signal (drift_norm, 0.40 weight — docs/EISV_COMPUTATION.md), so "S is drift" and "S contains drift" are both approximately true. It stays listed because readers meet the S description and the deviation components under the same word. The Rosetta table below uses drift as the standard name for the deviation signal, which is fine on public surfaces. In technical prose, qualify it: the self-attested and server-computed senses carry different evidential weight, and a reader who binds "drift" to the caller's own ethical_drift report will over-read it as measurement.

entropy — heuristic S axis vs. target entropy H (code level)

The label "Entropy" on S borrows the target word for a heuristic, so this is the same deployed-vs-target split as free energy. Qualify it whenever a sentence could be read as a measurement of H. The persisted column named entropy is a third, unrelated trap; see Persistence false friends.

proof of life — self-attested vs. externally-observed

The first is a claim the subject makes; the second is a fact something watching it reports — different questions, so binding the wrong one is the bug. The recurring false-archival incidents are self-attested-only liveness read as death (absence of a heartbeat is not observed death), and it is weakest exactly at silent-hang, where a heartbeat can outlive the work. Where an owning monitor exists (a BEAM orchestrator holding the agent's OS process), the observed sense is authoritative and the self-attested one is the fallback for agents with no supervisor. Sourcing the archival gate from the observed sense is proposed, not yet canonical (KG 2026-06-21T16:20:42, "liveness should be monitor-delegated, not self-reported").

Persistence false friends

Not homonyms: each is a single-sense storage name whose word suggests the wrong EISV axis. Map a column to an axis by its writer, not by its name. Canonical mapping: src/db/mixins/state.py (see rule 3 — this note points there rather than re-defining it).

• core.agent_state.entropy holds S (src/agent_storage.py, record_agent_state(entropy=S, …)), not E.
• core.agent_state.volatility holds V ("void maps to volatility column").
• E has no column; it lives in state_json.E.

Single-sense load-bearing terms

These currently answer one question each. Listed so a future split is visible against a baseline.

Open gaps (terms we're already pointing at but haven't coined)

These are places where the vocabulary lags the thing. Tracked here so the gap is explicit, not silently papered over by reusing a near-term.

• BEAM-resident agent has no harness (agent body) value. The documented lease harness values are hermes / claude_code / codex / dispatch / lumen (beam-coordination-kernel.md, an open "etc." list rather than a closed enum) — none of them names an agent whose body is a BEAM/OTP process (Sentinel Wave 1, Wave 3 handler dispatch). The BEAM coordination kernel itself is not a harness — it is a coordination substrate, a peer of UNITARES governance (it explicitly disclaims the role: "Do not replace Hermes as the agent harness"). But an agent resident on BEAM needs a harness value the taxonomy hasn't minted. This is the distinction that prompted this glossary. See harness-substrate-plurality.md layer table and beam-coordination-kernel.md non-goals.
• Sub-distinction (orchestrated agents). When a BEAM process spawns a non-BEAM agent (dispatch_beam → a claude/codex OS process) there are two bodies: the holder body — the supervised BEAM process that owns the agent's OS process and observes its :DOWN — and the agent body — the spawned CLI the agent reasons in. The harness (agent body) value names only the latter; the holder body has no term. They must not be conflated: the load-bearing rule is the PID models the holder, not the agent's governance self — the orchestrator gives the holder honest lifecycle but must not forge the agent's identity (dispatch_beam PLAN, "ontology payoff" + identity-honesty caveat).
• affordance_state shape is uncoined. Boolean reach? Capability list? Permission diff? Named as the most new-territory field in the provenance envelope; the term exists but the type does not.
• episode nesting is undecided. Clean for Discord threads (one thread = one episode); ambiguous for a long shell session with many invocations. The question is answered; the granularity is not.

Registers (the second axis)

The vocabulary feels unbounded because it spans several registers — parallel ways of describing the same underlying referents, each existing for a different reason. The term count is not really growing without bound; a roughly fixed set of referents is being named in four registers. Tagging a term's register is the orthogonal second axis to "question answered."

Public-copy rule. Any public-facing surface (pitch, dashboard labels, README first paragraph, onboarding intros) names a referent from the standard column. It must not borrow from fep — a physics term on a public surface claims rigor the deployed system does not drive behavior with (see the code-grounded correction below). ops symbols (EISV, basin) stay the internal and paper vocabulary; standard is their translation, not a rename.

Guardrail — "name nothing more rigorous than it is." This is the sibling of the manifesto axiom "build nothing that appears more alive than it is." A fep term (free energy, Markov blanket, active inference) is seductive because it sounds earned — it makes the system sound like it is doing variational math it may not be doing. A physics term enters the ops or philosophical registers only when the mapping is earned (the math actually holds), not as decoration. This is the same earned-vs-performative gate identity.md applies to identity, applied here to borrowed rigor.

Cross-register map (Rosetta — skeleton)

One row per referent; columns are its name in each register. This bounds the "unbounded nomenclature" into a finite table that grows by rows (new referents), not by uncontrolled vocabulary. Status marks the fep column specifically:

• ✓ earned — mapping holds and drives behavior (verdict path) today.
• ◐ research-lens — implemented as real math in the ODE / grounding tiers, but not wired to the verdict path as the mapped physics (the ODE owns only the cold-start Φ verdict, and after warmup leaks into E/I through compatibility coupling — see the second correction below); honest, not decorative, but not yet load-bearing. (Added 2026-06-20 after reading the code; see correction below.)
• ~ candidate — plausible conceptual fit, not yet implemented anywhere; decorative if shipped as fact.
• ✗ false friend — looks like a synonym across registers but is a different referent; do not conflate.

Reading the standard column. Four of the nine referents already are the standard term (drift, heartbeat, handoff, calibration) — the exotic-sounding load concentrates in three: EISV, basin, and dialectic (plus valence/running hot). Translating just those three on public surfaces removes most of the "too much jargon" friction without touching the internal or paper vocabulary. The standard column carries no fep-style status markers (✓/◐/~/✗) — those grade the physics mapping's rigor; the standard column is a naming convenience, not a rigor claim.

Correction (code-grounded, 2026-06-20). The skeleton's first claim — "physics register is aspirational, all ~, zero earned" — was wrong, and reading the code corrected it. Three rows are ◐ research-lens, not ~:

• governance_core/dynamics.py is a real dynamical-systems ODE (full dE/dt…dV/dt, RK4, coherence feedback, soft barriers, compute_equilibrium, contraction-theory convergence, and check_basin — a genuine bistable basin of attraction).
• src/grounding/free_energy.py already implements the "name nothing more rigorous than it is" guardrail in code: a 3-tier estimator where Tier-1 FEP is explicitly raise NotImplementedError("…Phase 2 scope"), Tier-2 resource ships, and every value carries a source= provenance tag so a heuristic is never laundered as a measurement. The guardrail predates this glossary and runs at runtime.
• docs/EISV_COMPUTATION.md already states the deployed-vs-target split this table was re-deriving: deployed EISV = "auditable heuristic blends, EMA-smoothed"; target = Eas−F, Ias mutual information, Sas entropy. The ODE is not the verdict owner: behavioral assessment owns the post-warmup verdict.

Second correction (2026-09-18). This passage originally quoted EISV_COMPUTATION.md as saying the ODE "runs in parallel and does NOT drive verdicts", and built a "two independent loops" reading on it. That source was corrected on 2026-08-29: the old line "describes authority, not causal independence." The behavioral sensor still blends the legacy ODE C(V) coherence term into 25–30% of E and 30–40% of I, and uses ODE-derived regime inputs. Two more channels exist: for check-ins 1–2 the verdict is owned by the Φ cold-start prior, and Φ is evaluated on the ODE state (src/monitor_phi.py); and when a caller omits confidence, 55% of the fallback estimate's base is the same legacy C(V_ODE) scalar, which feeds calibration history (docs/EISV_COMPUTATION.md, src/confidence.py). So the ODE owns the cold-start verdict and does not own the post-warmup verdict, but after warmup it still reaches verdicts, through compatibility couplings rather than through its attractor semantics.

So the accurate finding is sharper than "aspirational": the physics is built but not wired as physics. The verdict path is a heuristic loop (EMA + z-score + BASIN_HIGH health gate) that owns post-warmup decisions; the ODE is a research lens that owns the cold-start Φ verdict and otherwise reaches the verdict path only through compatibility couplings (the C(V) term in E/I, the regime inputs, the confidence fallback). Promoting a ◐ to ✓ does not mean building the math; it means wiring the existing ODE/grounding forms onto the verdict path deliberately, as the thing they claim to be — a real control-loop change to trajectory/telemetry, gated tier-by-tier with the provenance honesty already in place. The existing coupling is compatibility debt, not a head start: it changes the E/I distribution and every downstream baseline, so removing or replacing it needs a shadow evaluation first.

The register model still holds for these nine rows (four grading registers plus the standard public-naming column); if a referent won't fit a column, that absence is a finding (e.g. most rows have no genuine manifesto name — the norm register governs a few load-bearing referents, not all of them), not a gap to backfill.

FEP promotion conditions (roadmap)

Each ~ candidate from the table above, with the falsifiable condition it must meet before its fep cell is promoted to ✓ earned. The discipline: if you can't state a test that would fail, the mapping is decoration, not rigor. Listed most-reachable first.

On the false friend. Markov blanket stays ✗ for the lease-surface referent — they are not the same boundary. If the system ever wants a genuine Markov-blanket referent, it is the conditional-independence boundary separating an agent's internal states from external ones (sensory + active states mediating) — closer to affordance_state than to a coordination claim — and it is earned only when that statistical independence structure is actually modeled, not asserted.

Roadmap reading (code-corrected): the work is mostly wiring, not building. basin-as-attractor is a wire-up (ODE exists); precision-weighting is the closest greenfield-ish step (the basin-health gate is already a 0→1 ramp, which is precision-shaped); the EISV-distributional change is the prerequisite that unlocks surprise + active-inference; heartbeat→NESS should be dropped. The single highest-leverage move is making EISV distributional (mean + precision), because it both unlocks three FEP rows and is the precondition for letting the ODE drive verdicts honestly. That ordering is the actual deliverable — it says which physics claims to earn next and which to stop borrowing.

Maintenance

When a sweep finds a new collision, add it to the high-risk table here and log the sweep as a dated audit alongside glossary-drift-audit-2026-06-20.md. Keep this file keyed by question answered; if you find yourself writing a definition that leads with examples, you are seeding the next drift.

The mechanical half of that sweep now runs in CI: scripts/dev/check_glossary_drift.py verifies this file still parses into the structure the public-site viewer is generated from (scripts/dev/glossary_data.py), that every <file>.py::<symbol> reference here resolves to a real symbol, and that the cross-reference with the runtime glossary (src/governance_glossary.py) holds. Only semantic drift — a term's sense moving without its table row moving — still needs a human sweep. This file's section headings and table shapes are load-bearing for that parse; if you restructure them, run the checker before committing.