Shay 79f1678923 feat: ADR-0086 + ADR-0087 + 100-register catalog — cognition lane closure

Three load-bearing pieces:

1. ADR-0086 — UNKNOWN-intent pack-resident token surface
   New deterministic composer `pack_grounded_unknown_surface` in
   chat/pack_grounding.py.  When intent classification returns UNKNOWN
   but the prompt contains pack-resident lemmas (via cross-pack
   resolver), surface those lemmas with their semantic_domains
   instead of falling to the bare _UNKNOWN_DOMAIN_SURFACE.  Wired
   into chat/runtime.py::_maybe_pack_grounded_surface as the
   last typed-intent branch before the OOV fallback.  Null-lift
   invariant pinned: fully-OOV prompts still emit the universal
   disclosure byte-identically.  Closes four cognition-eval term
   misses: unknown_logos_019 (public), unknown_evidence_042 (dev),
   unknown_spirit_041 + unknown_word_018 (holdout).  Side effect:
   evals/results/phase2_pack_measurements.json refusal_rate drops
   from 0.25 → 0.125 across all three identity packs (no longer
   refusing on these prompts).

2. ADR-0087 — PROCEDURE selector + trailing-clause subject echo
   Two coupled changes in chat/pack_grounding.py:
   (a) Numeric-determiner downrank in _extract_procedure_topic_lemma:
       tokens whose primary semantic_domain starts with
       "quantitative.numeric." are demoted; non-numeric resident
       candidates always win.  So "compare two terms" anchors on
       `compare` not `two`.
   (b) Trailing clause echoes the full normalized subject_text
       rather than just the selected lemma, so OOV head nouns like
       "terms" reach the surface even when only the procedure verb
       is pack-resident.  Closes procedure_compare_011.

3. 100-register catalog
   New packs/register/_catalog.json — canonical machine-readable
   spec for all 100 registers (7 currently-ratified + 93 drafted)
   organized into 9 voice groups (depth/tone/stance/posture/domain/
   cultural/affective/functional/composite).  Each entry is a
   complete production input — realizer_overrides, marker palettes
   (openings/transitions/closings), depth_preference, description,
   author_notes.  All realizer_overrides use only legal keys per
   scripts/ratify_register_packs.py::_KNOWN_OVERRIDE_KEYS.
   Companion packs/register/CATALOG.md documents the production
   loop: materialize → widen REGISTER_IDS → ratify → smoke.

Cognition-eval lifts (all three splits):
  public:  term_capture 91.7% → 100.0%  (+8.3pp)
  holdout: term_capture 83.3% → 100.0%  (+16.7pp)
  dev:     term_capture 78.6% → 100.0%  (+21.4pp)
  surface_groundedness: 100% preserved on all splits
  intent_accuracy / versor_closure: 100% preserved on all splits

Tests:
  tests/test_pack_grounded_unknown.py     — 14 tests (composer
    direct + runtime engagement + null-lift invariant)
  tests/test_adr_0087_procedure_selector.py — 12 tests (selector
    numeric downrank + trailing-clause echo + regression guard)
  Existing test suites unaffected — cognition lane 120 passed / 1
  skipped both before and after.  Full lane net −3 failures vs
  pristine main (39 → 36 — none introduced).

2026-05-21 00:08:12 -07:00

4.2 KiB

Raw Blame History

Register pack catalog — production guide

packs/register/_catalog.json is the canonical, machine-readable spec for the 100-register catalog. Each entry is a complete production input — an author or content pipeline materializes packs/register/<register_id>.json from it mechanically and re-runs ratification.

Group	Count	Voice
A — Depth ladder	6	Varies disclosure_domain_count + structural booleans. No markers.
B — Tone	16	Affective marker palettes. Default knobs.
C — Stance	12	Epistemic posture — assertive, hedged, exploratory, etc.
D — Posture	10	Social role — peer, mentor, scholar, etc.
E — Domain	12	Academic, technical, legal, scientific, philosophical, etc.
F — Cultural	12	Plainspoken, cosmopolitan, classic, lyrical, etc.
G — Affective	10	Cheerful, somber, wry, gentle, etc.
H — Functional	10	Documentary, instructional, persuasive, etc.
I — Composite	12	Combine knobs from multiple groups.
Total	100	7 ratified + 93 drafted

Production loop

For each entry whose status == "drafted":

Materialize packs/register/<register_id>.json from the catalog entry. Required fields: register_id, version (set to "1.0.0"), description, schema_version ("1.0.0"), mastery_report_sha256 (leave "" — the ratify script will fill it), display_name, depth_preference, realizer_overrides, discourse_markers.
Widen scripts/ratify_register_packs.py::REGISTER_IDS to include the new register_id.
Ratify: python scripts/ratify_register_packs.py. Idempotent; re-runs produce byte-identical files.
Test: the ratification script writes the .mastery_report.json companion and updates mastery_report_sha256 in the pack file. Run python -m pytest tests/test_register_loader.py -q to confirm the loader self-seal check passes.
Smoke: core chat "Test prompt" --register <register_id> should produce a surface whose grounding_source and trace_hash are byte-identical to the same prompt under default_neutral_v1 (ADR-0072 register invariant). Only the surface text varies.

Trust boundaries

realizer_overrides is restricted to the allow-list in scripts/ratify_register_packs.py::_KNOWN_OVERRIDE_KEYS: disclosure_domain_count (1, 2, or 3), compress_gloss, drop_articles, drop_provenance_tag, append_semantic_domain_clause, plus per_intent nested block keyed by IntentTag names. Any other key fails ratification.
discourse_markers has exactly three buckets: openings, transitions, closings. Each is a list of strings. An empty string "" inside a bucket is a legal entry — the seeded selector (ADR-0071) treats it as "no marker this turn", enabling natural variation.

Author guidance

Keep each marker palette small (3–5 entries per non-empty bucket). Larger palettes are not better — the seeded selector wants a tight bounded space so per-turn variation is felt without being noisy.
Include "" in openings and closings for ~20–30% of registers so the system has a "land plainly" option per turn.
Markers may contain Unicode em-dash (—), ASCII punctuation, and standard English contractions. Avoid emoji and non-English script unless the register's purpose is explicitly multilingual.
For composite registers (group I), the description field should name the two constituent voices it combines.

Invariants pinned by ADR-0072 / ADR-0071

Every register in this catalog must satisfy, on every prompt:

grounding_source is byte-identical to default_neutral_v1.
trace_hash is byte-identical to default_neutral_v1.
aggregate metrics (cognition eval lane) are byte-identical.
Only the surface text varies.

The core demo register-tour runner exists to assert this. New registers should be added to its sweep once authored.

Next steps after authoring

Update the cognition eval's register-invariance test fixtures to include the new register ids (sweep, no per-register code change).
Optional follow-up: add a register_distribution_lift benchmark (ADR-0072 variant) that measures surface-variance across the full 100-register sweep for a fixed prompt corpus.

4.2 KiB Raw Blame History Unescape Escape