core/docs/assessment/02-layer-taxonomy.md

Phase 1 — The Layer Taxonomy

Executor: Fable 5, 2026-07-27. Verified against: forgejo/main @ 8927c563. Inputs: 01-phase0-ground-truth.md (the handoff), AGENTS.md, ADR-0252, .system-map/ (2026-06-09 prior — local to the main worktree, gitignored). Companion: 03-card-schema.md (the metadata schema every card in Phases 2–3 must fill).

This document fixes the decomposition of CORE — macro to micro — that Phases 2 and 3 fill with cards, and Phase 4 audits against. Per Pillar I (Semantic Rigor), the taxonomy and its completeness criteria are fixed before any component is judged.

---

0. Decisions made by this phase

#	Decision	Ruling
D1	Which articulation is the spine?	None of the five "wins." They answer five different questions and become five axes/attributes of one taxonomy (§1). The functional axis is the AGENTS.md north star; the structural axis is a 7+2 macro-layer grouping of the system map's 33 zones. The taxonomy's job is auditing the mapping between the two.
D2	Schema: adopt, extend, or replace the system map's?	Extend. Keep the 17 fields and the 7-value liveness vocabulary; add the four missing dimensions (evidence, capacity, design-vs-build, fitness) as orthogonal fields, not fatter enums (03-card-schema.md).
D3	Completeness criterion	Fixed in §6: coverage is evidence-bearing ownership of a functional stage, completeness is per-layer and per-stage, and "missing" is distinguished from "explicitly deferred" by the presence of a ruling.
D4	Where do candidate layers live?	The Candidate Register (§5) — four registered candidates (attention, agenda/drive, efferent action, temporal self-location), each with its telos derivation, partial existence, and the ruling it needs; plus a considered-and-not-registered line to show the boundary was examined.
D5	Committed or local-only?	Committed (this branch → Forgejo PR). This artifact is governance-adjacent: it will drive R&D ordering and rulings, so it must be reviewable, versioned, and visible to every future session. The system map stays local as a regeneratable navigation index; the assessment is the ruled record. Staleness is handled by discipline, not by hiding: every card stamps the SHA at which its claims were verified.

---

1. D1 — the reconciliation: five articulations, five different questions

Phase 0's Finding 0-A: CORE describes its own cognitive cycle five different ways, and no document maps any onto any other. The resolution is not to crown one. Read closely, the five are not competing answers to one question — they are answers to five different questions, and the apparent conflict dissolves once each is assigned to the axis it actually describes. (This is the Third Door applied to the taxonomy itself: the trade-off "which self-description do we keep?" is not split; it is dissolved.)

Articulation	The question it answers	Kind	Disposition in this taxonomy
North star (AGENTS.md, 7 stages)	What is the system for?	Functional / teleological	Adopted as the functional axis (§2). Candidate extensions are registered separately (§5), never silently merged into it.
Live path (AGENTS.md, 9 steps)	What happens on a turn today?	Realized pathway	Subsumed. It is the serving traversal M2 → M3 → M4 → MV (§3). It becomes the spine of the M3/M4 layer cards — a pathway through the structure, not a decomposition of it.
Mind-physics blueprint (docs/architecture/MIND-PHYSICS-BLUEPRINT.md, Draft 2026-05-12)	By what mechanisms?	Mechanism proposal	Recommend ruling: mark historical / partially superseded, with the per-element disposition in §1.1. Its one orphan — Allocation Physics — moves to the Candidate Register as CR-1.
ADR-0252 paradigm (Accepted 2026-07-19)	By what competence is problem-solving judged?	Governing competence model	Adopted as the governing competence model within M3. Its five stages decompose the problem-solving pathway inside M3; its §4 conformance bar (deep structure, generalization ratio > 1) becomes a fitness criterion for M3 components in Phases 3–4.
System map (.system-map/, 2026-06-09)	What is built, and where?	Structural inventory	Adopted as the sublayer stratum. All 33 zones are retained and regrouped under 7+2 macro layers (§3–4); the 7-value liveness vocabulary is retained unchanged.

The two-axis model that results:

• Functional axis — the stages of the cognitive cycle (what the organism does).
• Structural axis — macro layers → zones → components (what exists, in containment order).

Every structural card carries the functional stages it serves (telos_stages, already a system-map field). The assessment's central audit — run in Phase 2 with evidence, not here — is the stage-coverage audit: every functional stage must be owned by at least one live structural element (else a gap), and every structural element must serve at least one stage (else a fitness question). Neither axis can perform this audit alone; that is why neither "wins."

1.1 Disposition of the mind-physics blueprint, element by element

The blueprint is a Draft that was never advanced and never retracted. Its elements did not fail uniformly, so a uniform disposition would be dishonest. Recommended for ruling (this assessment records, it does not enact):

Blueprint element	What became of it	Disposition
Identity Physics (IdentityCheck, ADR-0010)	Landed and hardened: wave-only fail-closed identity scoring (ADR-0244 §3, INV-32), identity manifold, identity packs	Superseded by stronger implementations — retire the blueprint's version of the claim
Compositional Physics (Binding/Digest/Trajectory/ArticulationPlanner, ADR-0009)	Landed as the proposition-graph lineage: PropositionGraph → ArticulationTarget → realizer	Superseded — same
Renderer ("TBD, ADR-0011 planned")	generate/realizer.py has been the shipping renderer for months	Stale line — retire
DriveGradientMap / ExertionMeter	Never built; no successor articulation anywhere	Moves to Candidate Register CR-2 (agenda/drive)
Allocation Physics (SalienceOperator / AttentionOperator / InhibitionOperator, ADR-0008)	Never landed as a layer. Fragments exist under other names — see CR-1	Moves to Candidate Register CR-1 (attention/allocation)

---

2. The functional axis

2.1 The ratified cycle (unchanged, from AGENTS.md)

listen → comprehend → recall → think → articulate → learn from reviewed correction → replay deterministically

This is the canonical vocabulary for telos_stages. The system map's telos mapping (which zones serve which stage) is adopted as the claimed ownership baseline; Phase 2 verifies claims against evidence.

2.2 Claimed stage ownership (baseline — unverified until Phase 2)

Stage	Primary owner (macro layer)	Supporting
listen	M2 Afferent Boundary	M1 (lexical substrate)
comprehend	M3 Comprehension & Reasoning	M1 (packs, vocabulary)
recall	M1 Knowledge & Memory	M3 (in-turn recall), M0 (exact CGA distance)
think	M3 Comprehension & Reasoning	M0 (the medium)
articulate	M4 Expression & Serving	M1 (packs), MG (governance of the surface)
learn	M5 Learning & Growth	M1 (promotion target), M6 (what persists)
replay	MV Verification & Evidence	property enforced everywhere; apparatus lives in MV
(the cycle's runner)	M6 Continuity & Process	unbuilt at its center — L11 process

Candidate functions (attend, want/agenda, act, temporal self-location) are deliberately not rows in this table. They live in §5 until ruled into the cycle or ruled out.

---

3. The structural axis — seven macro layers, two cross-cuts

Macro layers are groupings of the 33 zones, carved at the joints the system's own contracts already respect (serve boundaries, trust boundaries, the INV regime). Each gets one card in Phase 2 (10-layer-cards/). One-paragraph teleology for each; full intent belongs on the card.

M0 — Substrate. The physical medium: Cl(4,1) algebra, the versor invariant, the field. Per the governing mental model, the field is the electricity, not the intelligence — M0 supplies closure, exactness, and replayability to everything above it, and is forbidden from containing cognition-specific policy. Zones: L0-algebra, L1-field.

M1 — Knowledge & Memory. What is known, at rest: the vault (exact recall), compiled packs, the vocabulary manifold, the lexical substrate (morphology, alignment), and memory's consolidation surface. The epistemic-status regime (SPECULATIVE/COHERENT/…) governs everything here. Zones: L2-vault, L3-packs, vocab-manifold, morphology, alignment-resonance, L8-memory-contemplation (straddles M5 — consolidation is learning in motion).

M2 — Afferent Boundary. World → field: the ingest gate and compiler, the sensorium's afferent track, and environmental falsification. Every entry point is a trust boundary; nothing crosses without construction-boundary normalization. Zones: ingest-boundary, ingest-compiler, sensorium-afferent, sensorium-falsification.

M3 — Comprehension & Reasoning. The wiring that thinks: recognition, the cognition pipeline, the comprehend/determine/realize organs, the deduction flagship, curriculum-grounded reasoning, the math reader, and reasoning research. ADR-0252 governs this layer's competence bar; the two-grammars frontier (reader ≠ writer) lives here. Zones: L4-recognition, L5-cognition, comprehend-organ, determine-phase, realize-phase, reasoning-deductive, gsm8k-math, field-wedge-research.

M4 — Expression & Serving. Field → world: the chat runtime, surface selection policy, the register axis, response governance, and the epistemic-verdict surface. This layer owns serving-path truth behavior — what the user actually reads — and is therefore where wrong=0 lives or dies. Zones: L6-chat-runtime, L9-epistemic-verdicts (straddles MG — verdicts are governance made visible).

M5 — Learning & Growth. Controlled mutation: the reviewed teaching loop, formation/curriculum, reliability calibration and earned licenses, and the capability ledger. The typed learning boundary (durable-reviewed vs provisional-typed, INV-21…24/29/30) is this layer's constitution. Zones: L7-teaching, formation-curriculum, reliability-calibration, capability (straddles MV — the ledger is also evidence).

M6 — Continuity & Process. The life itself: the always-on process (unbuilt), engine-state checkpointing (built), edge-sync, the turn protocol, session continuity, async HITL. The telos ("one continuous life") is this layer's charter, and its center is the single largest distance between stated purpose and built system. Zones: L10-11-runtime-identity, engine-state, edge-sync, core-protocol-ctp.

MG — Governance & Identity (cross-cutting). Identity manifold and packs, safety pack (never-swappable, fail-closed), ethics packs, refusal taxonomy, trust boundaries, the INV regime as a body of law. Cross-cutting because its writ runs everywhere; a governance mechanism that only one layer obeys is a bug. Zone: governance-identity-safety.

MV — Verification & Evidence (cross-cutting). Evals, lanes, pinned SHAs, CLAIMS.md, replay/determinism apparatus, telemetry/trace, the CLI test surface, workbench-as-auditor. Cross-cutting for the same reason; replay is a property of every layer and an apparatus of this one. Zones: evals-determinism, tooling-cli-workbench-rs.

flowchart LR
  W((world)) --> M2[M2 Afferent]
  M2 --> M3[M3 Comprehension & Reasoning]
  M3 <--> M1[M1 Knowledge & Memory]
  M3 --> M4[M4 Expression & Serving]
  M4 --> W2((world))
  M4 --> M5[M5 Learning & Growth]
  M5 --> M1
  M0[M0 Substrate] -.the medium.- M1 & M3
  M6[M6 Continuity & Process] -.hosts the cycle.- M2 & M3 & M4 & M5
  MG[MG Governance & Identity]:::cc -.governs all.- M4
  MV[MV Verification & Evidence]:::cc -.witnesses all.- M4
  classDef cc stroke-dasharray: 3 3;

---

4. The complete zone mapping

All 33 zones, mapped once. Liveness is the map's 2026-06-09 label — 48 days stale, claimed not verified; Phase 2 re-verifies. Flags: ⚑ = zero-subsystem zone (Phase 3 must descend it first); ✱ = straddle (noted above).

Zone	Macro layer	2026-06-09 liveness	Note
L0-algebra	M0	live-serving
L1-field	M0	partial-wiring-debt
L2-vault	M1	partial-wiring-debt	T1 discarded on exit (→ M6)
L3-packs	M1	partial-wiring-debt
vocab-manifold	M1	live-serving
morphology	M1	live-internal
alignment-resonance	M1	live-internal
L8-memory-contemplation ✱	M1 ↔ M5	partial-wiring-debt	consolidation straddle
ingest-boundary	M2	partial-wiring-debt
ingest-compiler	M2	partial-wiring-debt
sensorium-afferent	M2	inert
sensorium-falsification	M2	live-internal	map labelled it "L12" — a stratum no other document uses. Ruled 2026-07-28 (R-7): dropped. No in-repo generator emits it (the system map is local and gitignored, per D5 a regeneratable index carrying no authority), so nothing in the repository was changed — the phantom existed only in a local artifact and in this row. Recorded rather than silently deleted, so the next reader does not re-derive it.
L4-recognition	M3	partial-wiring-debt
L5-cognition	M3	partial-wiring-debt
comprehend-organ ⚑	M3	live-serving	zero subsystems mapped
determine-phase ⚑	M3	live-serving	zero subsystems mapped
realize-phase ⚑	M3	live-serving	zero subsystems mapped; serve seam → M4
reasoning-deductive	M3	partial-wiring-debt	flagship; predates ADR-0256–0265 arc
gsm8k-math	M3	partial-wiring-debt	demoted to diagnostic
field-wedge-research	M3	research-negative	negative result = knowledge, not failure
L6-chat-runtime	M4	partial-wiring-debt
L9-epistemic-verdicts ✱	M4 ↔ MG	partial-wiring-debt
L7-teaching	M5	partial-wiring-debt
formation-curriculum	M5	partial-wiring-debt
reliability-calibration	M5	partial-wiring-debt
capability ✱	M5 ↔ MV	live-internal
L10-11-runtime-identity	M6	partial-wiring-debt	center unbuilt (L11 process)
engine-state	M6	live-internal	the built footing
edge-sync	M6	inert
core-protocol-ctp	M6	spike
governance-identity-safety	MG	live-serving
evals-determinism	MV	partial-wiring-debt
tooling-cli-workbench-rs	MV	partial-wiring-debt
sensorium-falsification ⚑	—	—	(also zero-subsystem; listed once above)

Calibration to carry forward: at the subsystem stratum the 2026-06-09 map records 58 live-serving + 79 live-internal of 205 total (with 28 partial, 22 inert, 10 spike, 4 unbuilt, 4 research-negative). The organism is substantially more built than zone-level labels imply; the genuinely-unbuilt mass is concentrated in M6. Zone liveness is a weakest-link rollup and must never be quoted as a completion rate.

---

5. The Candidate Register (D4)

Functions or layers that no ratified document names, but that the telos arguably implies. Registration here asserts nothing except this deserves a ruling. Each entry: derivation → partial existence → ruling needed → risk if left unregistered.

CR-1 — Attention / allocation (in-turn). Derivation: any bounded cognitive system must select what to process; the blueprint articulated it (ADR-0008) and no successor document owns it; the north star is silent between "listen" and "comprehend." Partial existence: admissibility threshold and margin gates (ADR-0024/0026), rotor admissibility (ADR-0025), and salience/nearest-neighbour search — which is precisely the measured hot path (~73% of turn time through cga_inner/geometric_product, Finding 0-F). The blueprint's InhibitionMask appears never to have been built. Ruling needed: is attention a first-class layer, or an emergent property of admissibility that should stay distributed? Risk: the hottest path in the system — computationally and semantically — has no owner, no card, and no governance. Optimization and correctness work on it currently has nowhere to attach.

CR-2 — Agenda / drive (between-turn). Derivation: "one continuous life" must decide what to do when it is not serving. A process with nothing to want is a heartbeat, not a life. Partial existence: idle consolidation (CLOSE), read-only proposal review, the contemplation loop (flag-gated), discovery-yield telemetry, core/epistemic_questions/. All are mechanisms without a chooser — each does one thing when its flag is on; nothing ranks what matters next, and the blueprint's DriveGradientMap/ExertionMeter were never built. Ruling needed: does the L10 process own an agenda; by what policy is it governed (identity packs? curriculum priorities? operator queue?); and is agenda-formation itself subject to the typed learning boundary? Risk: L10 lands as an always-on process that idles — the telos's letter without its spirit. This is the largest conceptual absence for an AGI-grade system: everything CORE does is currently chosen by the operator.

CR-3 — Efferent action. Derivation: AGI-grade generality ordinarily implies acting on the world; CORE's telos ends at articulate/learn/replay — deliberately text-first. Partial existence: typed deterministic tool operators folded into trace_hash (ADR-0018); the environmental-falsification contract (ADR-0211) explicitly forbids motor/efferent units in v1. Ruling needed: an explicit scope declaration — is action deferred (like scripture content: a ruling exists) or out of telos? Today neither is stated anywhere, which is the gap. Risk: silent scope ambiguity. Note honestly: the alignment posture (position paper §6) is arguably stronger with action explicitly deferred — this register entry is about making the boundary ruled, not about advocating efferents.

CR-4 — Temporal self-location. Derivation: a continuous life experiences sequence and duration; determinism bans clocks from cognition. Both commitments are correct, and their intersection is undesigned: how does the L10 process represent "now," "before," and "how long" without breaking replay? Partial existence: session-context ordering, engine-state turn_count, idle ticks as pseudo-time, the Fibonacci recency constants schedule (τ_n, ADR-0242 — a constants schedule, not a clock). Ruling needed: a stance on lived time for the L10 spike — the 24h+ no-drift requirement cannot even be stated precisely without one. Risk: the L10 spike gets designed with an implicit, accidental answer to a question nobody asked out loud.

Considered and not registered (recorded so the boundary is visibly examined, per Pillar I):

• Sociality / other-minds — multi-actor comprehension is a comprehension capacity inside M3 (the ADR-0174 pronoun hazard is its live trace), not a layer. Revisit only if the telos expands to multi-party life.
• Emotion / affect — no derivation from the telos under decoding-not-generating; disposition is already carried structurally (identity packs, hedging, refusal taxonomy, register axis). Registering it would import an external architecture's expectations — precisely what Pillar II forbids.
• Full embodiment — subsumed by the CR-3 ruling.
• Epistemic self-governance as a first-class layer — CORE's most distinctive machinery (earned licenses, calibration, ledgers, typed refusal) exists and is live, but is split across M5/MG/MV. This is an organizational question, not a missing function; noted for the Phase 5 synthesis rather than registered as a candidate.

---

6. Completeness criteria (D3)

Fixed now, applied in Phases 2–4. All terms below are used in their 03-card-schema.md senses.

A functional stage is covered iff at least one structural component owns it with liveness live-serving (or live-internal for internal-only stages such as replay) and at least one evidence pointer that would fail if the component were deleted. Ownership without such evidence is claimed coverage and is recorded as uncovered.

A layer is complete iff:

1. every functional stage it claims is covered at the capacity its card declares (not merely "at all");
2. every zone and component within it has a card with no mandatory field left undetermined;
3. every invariant it declares has failing-when-violated enforcement — a pin that lives in a suite that actually runs, verified by the suite-count check, not by the pin's existence.

A missing layer or component is identified when any of:

• (i) a ratified functional stage has no covered owner;
• (ii) a governing document's mechanism has no structural home (the Allocation-Physics test);
• (iii) a telos-implied function has no owner and no Candidate-Register entry — this clause is the AGI-grade audit, and it is why §5 exists: once registered, a candidate is a ruling item, not a silent absence.

Deferred is not missing. An absent capability is out-of-scope iff an explicit ruling defers it (scripture content; motor efferents in falsification v1). Absent-with-no-ruling is a gap. The cure for a gap of this kind may simply be a one-line ruling — the register makes that cheap.

The system is complete — never claimable today, stated so the target is fixed — iff the north-star cycle runs end-to-end under the M6 process with every stage covered at declared capacity, every invariant enforced, and zero undetermined fitness verdicts. Note what this criterion deliberately omits: benchmark scores. Per the master plan's own distinction, architectural distinctiveness is the target; benchmark wins are downstream validation.

---

7. Phase 2 work order

Deliverable: nine layer cards in 10-layer-cards/ — M0-substrate.md, M1-knowledge-memory.md, M2-afferent-boundary.md, M3-comprehension-reasoning.md, M4-expression-serving.md, M5-learning-growth.md, M6-continuity-process.md, MG-governance-identity.md, MV-verification-evidence.md — each schema-compliant per 03-card-schema.md.

Priority order (by leverage, not ease): M6 (telos-critical; the L10 question), M3 (serving-path truth; the two-grammars frontier; ADR-0252's home), M4 (what the user reads), M5 (the learning boundary), then M1, M2, M0, MG, MV.

Obligations:

1. Re-verify every liveness label quoted from the 2026-06-09 map before writing it into a card; stamp verified_at with the SHA actually inspected. The map is a prior, never a source.
2. Run the stage-coverage audit (§2.2 table, with evidence this time) and record it in each card's stage_coverage block. This is where "is a stage uncovered" gets its verdict.
3. Apply the sabotage test to every live-* claim inherited or newly made.
4. The four ⚑ zero-subsystem zones (comprehend-organ, determine-phase, realize-phase, sensorium-falsification) are unmapped-and-load-bearing: the M3/M2 cards must scope them honestly (what is known vs unmapped) and queue them first for Phase 3 descent.
5. Where a card touches the PR #138 fabrication findings: they are measured-and-pinned, held for ADR + ratification — record, never re-discover, never fix.
6. Straddle zones (✱) are cited on both cards but owned by one (the table's left column); the owning card carries the full entry, the other a cross-reference. No double-counting in any rollup.