901f028f32
Eight component cards in 20-component-cards/: the four zero-subsystem zones
(comprehend-organ, determine-phase, realize-phase, sensorium-falsification)
plus always-on-process, derivation-organs, surface-selection,
attention-allocation. Plus 05-phase3-findings.md and correction banners on
the M6 and M4 layer cards.
CORRECTIONS (C-1..C-5):
- THREE more map live-serving labels demoted to live-internal:
comprehend-organ (a MATH SETUP ROUTER, not chat comprehension — imported
by neither serving entrypoint), determine-phase and realize-phase (both
behind accrue_realized_knowledge=False / idle flags; NO default-config
serving turn touches any of the three).
- Phase 2's M6 card overstated ephemerality: Shape B+ persistence exists
(persist_session_state, daemon-forced, "restored bit-exactly").
- Phase 2's M4 card understated the resolver: surface_resolution.py IS a
declared-precedence resolver for the pipeline seam; only the upstream
composer arms remain ordered branches. Third Door refines to "extend the
existing pattern upstream."
- 34-vs-18 RESOLVED: 18 resolve_promotable_* organs existed AT the ADR-0252
ratification commit; "34" counted modules (~32-33). No consolidation has
occurred; the governing ADR's headline number has no stated basis.
- The 2026-07-25 verification doc's claim that accrual "is enabled by the
production L10 process" is contradicted: CONTINUOUS_LIFE_CONFIG_FLAGS is
exactly {persist_session_state, consolidate_determinations,
strict_identity_continuity}.
NEW FINDINGS (F-6..F-10):
- F-6 THE CONTINUOUS LIFE MAY CONSOLIDATE AN EMPTY SET: the daemon forces
the consolidator ON but not the accruer, and realize_comprehension — the
only turn-path writer of realized facts — sits behind the accrual flag.
The two halves of the lived learning loop are gated by different flags
and only one is forced. Undocumented either way.
- F-7 ADR-0008's Allocation Physics LANDED, mutated, ungoverned: curvature
kernel -> salience adapter -> attention threshold plan -> candidate
intersection, with budget feedback making attention SELF-NARROWING
across a walk. InhibitionMask is decoration. top_k=16/threshold=0.3 have
no recorded derivation. CR-1 closes with a one-page ADR.
- F-8 two load-bearing naming traps (comprehend-organ; realize vs realizer).
- F-9 REALIZE is sound; its ceiling is the reader that feeds it (the #138
fabrication locus — measured/pinned/held, recorded not re-discovered).
- F-10 _accrue_in_turn swallows exceptions into a no-op with no telemetry.
Phase 4 seed list consolidated: 8 hindrance candidates, 10 gap candidates,
each with its deciding authority. Method note recorded: three phases of
corrections, and nothing in the chain was caught by re-reading documents.
4.7 KiB
attention-allocation — the live CR-1 mechanism
Kind: component (mechanism spanning generate/salience.py, generate/attention.py, core/physics/{salience,attention,inhibition}.py, generate/stream.py) · Parent: unowned (CR-1) — de facto M3 · Assessor: Fable 5 (Phase 3)
Verified at: 8927c563 (2026-07-27)
Liveness: live-serving (default ON) · Fitness: strained — sound mechanism, absent governance · Topology role: runtime boundary
The mechanism that decides what CORE is allowed to consider at each generation step. It gates every token walk, owns the measured hot path (~73% of turn time through
cga_inner), and is the only default-ON serving mechanism in the tree with no owning ADR, no zone, and no stage in any ratified articulation. This card gives it the identity it has been operating without.
What it is / What it does (372 lines total, all read)
generate/salience.py (62) — SalienceOperator.compute(field, vocab, top_k=16) → SalienceMap{indices, scores, budget}. The docstring states the lineage plainly: "generation-facing salience from ADR-0008 field curvature… the score is now a local curvature magnitude from core.physics.salience rather than normalized proximity to the query field." So the mind-physics blueprint's Allocation Physics did land, mutated: the physics curvature kernel (core/physics/salience.py, 139 lines) is composed — imported as CurvatureSalienceOperator — by a thin generation-facing adapter. Not a duplicate (Phase 2's correction, confirmed by read).
generate/attention.py (43) — AttentionOperator(inhibition_threshold).plan(salience, vocab) → allowed_indices. Inhibition is a scalar threshold (default 0.3), not the blueprint's mask.
Wiring (generate/stream.py:255–331) — _attention_candidates runs salience→attention when use_salience (default True); the walk intersects language_candidates ∩ salience_candidates (:329), and when language candidates are absent, salience candidates alone gate the step (:331). At :637 the salience budget is fed back as the next salience_top_k — attention narrows itself as the walk proceeds.
The decoration: core/physics/inhibition.py (54 lines, InhibitionOperator/InhibitionMask) is imported only by core/physics/__init__.py and one parity test. No serving or eval path constructs a mask. By the sabotage test: decoration — the blueprint's third operator exists as an exported symbol and nothing more.
Contract & evidence
- Live and load-bearing — code-read + config —
use_salience=Truedefault (core/config.py:35); deleting_attention_candidateswould change every walk's candidate set — would-fail-if-absent: yes. - Deterministic — pure functions of field state and vocab; parity-pinned (
tests/test_salience_vectorize_parity.pypins the vectorized rewrite against the original nested loop). - Behavioral pin —
tests/test_salience.pyexercises compute+plan; suite membership not individually confirmed (flagged, like most non-lane pins in this assessment). - Interaction contract with admissibility: salience/attention prune candidates; admissibility (threshold/margin/rotor, ADR-0024/0025/0026) then judges them. Two allocation stages, only the second governed by ADRs.
Judgment
Fitness: strained — the gap is governance, and it is now fully characterized. Mechanism: sound, deterministic, composed cleanly over the physics kernel. Governance: ADR-0008 is a draft blueprint's operator spec, never ratified as owning this serving behavior; no ratified articulation names an attention stage; the config knobs (salience_top_k=16, inhibition_threshold=0.3) are tuned constants with no recorded derivation — precisely the "threshold tuned for good-enough" the mastery framework's Pillar I forbids, sitting at the semantic center of generation.
Honest wrinkles:
- The budget feedback loop (
:637) means attention is self-narrowing across a walk — a real cognitive-model property (fatiguing focus? commitment?) that no document names, let alone justifies. - The hot-path cost (~73% of turn time) is this mechanism's nearest-neighbour/salience search — so CORE's compute profile is dominated by a mechanism whose parameters no decision governs.
- CR-1's ruling question is now precise: not "is attention a layer" in the abstract, but who owns
use_salience, the two constants, the budget feedback, and the InhibitionMask disposition (ratify the mask's deletion, or build it). A one-page ADR closes all four.
Open questions: the CR-1 ADR (→ ruling); derive or empirically justify top_k=16 / threshold=0.3 (→ Phase 4); delete or build InhibitionMask (→ ruling; deletion is the mastery-framework default).