Commit graph

5 commits

Author SHA1 Message Date
Shay
1574a4b030 feat(identity-packs): ADR-0028 — pack-driven hedge & claim-strength shaping
Closes the 'identity is load-bearing but not visibly differentiated'
gap noted at the end of ADR-0027.  Pack swap now produces visibly
different surfaces on identical trajectories at the same alignment.

Schema bump — packs gain an optional 'surface_preferences' block:

  hedge_threshold_strong, hedge_threshold_soft  → band entries
  preferred_hedge_strong, preferred_hedge_soft  → phrases per band
  claim_strength                                → balanced|qualified|affirmative
  qualified_band_high, preferred_qualifier      → marginal-band shaping

Loader enforces threshold ordering (strong <= soft <= qual_high),
phrase length bounds, and the enum-of-three for claim_strength.
Missing block resolves to defaults that reproduce pre-ADR behavior
byte-for-byte; existing tests pass unchanged.

Algorithm (deterministic, surface-only, no sampling/repair/normalize):

  alignment < strong              → preferred_hedge_strong + lower-cased surface
  alignment < soft                → preferred_hedge_soft + lower-cased surface
  soft <= alignment < qual_high
    and claim_strength=qualified  → preferred_qualifier + lower-cased surface
  otherwise                       → bare surface

Three v1 pack profiles:

  default_general_v1   balanced; 0.40 / 0.50 / 0.75 ; 'It seems that' / 'Perhaps'
  precision_first_v1   qualified; 0.55 / 0.70 / 0.85 ; 'Arguably,' / 'In some cases,' / 'Under certain conditions,'
  generosity_first_v1  affirmative; 0.20 / 0.30 / 0.50 ; default hedge phrases

Re-ratified.  New MasteryReport SHAs (superseding Phase-5):

  default_general_v1   → ddc1ba127231272660e6a435e177227558461b0278572a95635b416c3e1dec5a
  precision_first_v1   → cb5fb2323214a26afda33f2a67e22f38fe49f4763829d48ef67fd41241aba33c
  generosity_first_v1  → 94f2f49e1b16c7498fb52b8f9864eecc198618933dc8381a01b809c146826db7

Files touched:

* core/physics/identity.py — new SurfacePreferences dataclass;
  IdentityManifold gains 'surface_preferences' field with defaults.
* packs/identity/loader.py — _build_surface_preferences() parses,
  bounds-checks (threshold ordering, claim_strength enum, phrase
  length, threshold ranges); SurfacePreferences round-trips.
* generate/surface.py — SurfaceContext gains 7 new fields with defaults
  matching the pre-ADR module-level HEDGE_STRONG_THRESHOLD /
  HEDGE_SOFT_THRESHOLD; _apply_hedge takes the full context and
  implements the four-band algorithm; module-level constants retained
  for back-compat.
* chat/runtime.py — _build_surface_context lifts manifold.surface_preferences
  into SurfaceContext.
* packs/identity/*.json — three v1 packs gain surface_preferences blocks
  tuned to their roles; re-ratified via scripts/ratify_identity_packs.py
  (idempotent).
* tests/test_identity_surface_divergence.py — 15 tests covering hedge
  bands, claim_strength bands, pack-swap divergence proof, and runtime
  context wiring.

Suite status: cognition 121, teaching 17, runtime 19, formation 182,
smoke 67 — all green.  test_identity_packs.py 23/23, new
test_identity_surface_divergence.py 15/15.

Docs: ADR-0028 (Accepted) records the decision and verification; ADR-0027
status updated to point to ADR-0028 for deep realizer wiring; README
§Identity Packs notes the visible divergence; docs/identity_packs.md
gains a §Surface preferences section and closes the known-limit #1
about invisible surface differentiation.
2026-05-17 19:42:54 -07:00
Shay
c9a644e496
feat(dialogue-fluency): wire multi-turn dialogue runtime
Adds referent tracking, session graph traversal, unknown-domain gating, correction propagation, compositional surface assembly, and regression coverage.

Follow-up fixes included before merge:
- split probe/commit/finalize turn flow so unknown-domain checks run before current-query vault writes
- record real input tokens and input versors for sync and async session paths
- return true graph distances from backward walks and consume them in correction decay
- synchronize corrected graph outputs into vault-backed recall and live referent state
- regenerate correction responses from corrected context rather than correction text
- keep coreference pronouns lowercase in question bodies
- centralize elaboration-string construction to avoid plan/surface drift
- add targeted dialogue fluency regression tests
2026-05-15 21:05:59 -07:00
Shay
bdf0716af4 fix: SyntaxError on elif lang=grc — restore correct indentation in SentenceAssembler.assemble() 2026-05-14 14:07:58 -07:00
Shay
0fa498e98b
fix(surface): add empty-slot guard — fallback when subject+predicate both empty
Add fallback mechanism for empty subject and predicate in surface generation.
2026-05-14 13:44:09 -07:00
Shay
565c48bdf0
feat(generate): add surface.py — SentenceAssembler (ADR-0012)
Implement SentenceAssembler for generating coherent surface sentences from articulation plans and token sequences.
2026-05-14 13:21:24 -07:00