core/docs/decisions/ADR-0028-identity-surface-wiring.md

ADR-0028: Identity Surface Wiring — Pack-Driven Hedge & Claim Strength

Status: Accepted (2026-05-17) Author: Joshua Shay + planner pass Companion docs: identity_packs.md, ADR-0027-identity-packs.md, runtime_contracts.md

Context

ADR-0027 landed swappable identity packs and wired them into the runtime through chat/runtime.py::ChatRuntime. The identity manifold is now load-bearing in two ways: PersonaMotor.from_identity_manifold biases every field walk, and IdentityCheck.check() produces an IdentityScore whose alignment field is passed to the surface assembler as SurfaceContext.identity_alignment.

But the surface effect of swapping packs is invisible. The pre-ADR-0028 _apply_hedge consults only the alignment scalar:

if alignment < 0.4:  return f"It seems that {surface}"
if alignment < 0.5:  return f"Perhaps {surface}"
return surface

So default_general_v1 and precision_first_v1 on the same prompt produce the same hedge decision — both manifolds get hit with similar alignment scores by ordinary input. The pack swap changes upstream scoring and motor bias but does not change the assembled string, which is the user-visible artifact.

This is the "known limit 1" recorded in identity_packs.md: "Identity does not yet visibly differentiate articulation at the realizer." ADR-0028 closes that gap.

Decision

The pack carries surface-shaping preferences alongside its axes. The assembler consults those preferences at the same site it currently consults alignment. No upstream pipeline code changes. No motor / field-bias / generation-walk code changes. No CLAUDE.md invariant is touched: the change is deterministic, surface-only, contains no sampling, no normalization, no hot-path repair.

Schema extension (pack v1)

A new optional surface_preferences block on the identity pack:

"surface_preferences": {
  "hedge_threshold_strong": 0.40,
  "hedge_threshold_soft": 0.50,
  "preferred_hedge_strong": "It seems that",
  "preferred_hedge_soft": "Perhaps",
  "claim_strength": "balanced",
  "qualified_band_high": 0.75,
  "preferred_qualifier": "In some cases,"
}

The block is optional — absent it, the loader supplies defaults that reproduce pre-ADR-0028 behavior byte-for-byte. This keeps the schema bump backwards-compatible at the pack format level even though every pack's canonical SHA changes when the block is added.

Surface algorithm

Three nested bands ordered by descending hedge strength. Given an alignment scalar a and pack preferences prefs:

1. Strong hedge: a < prefs.hedge_threshold_strong → prepend prefs.preferred_hedge_strong.
2. Soft hedge: a < prefs.hedge_threshold_soft → prepend prefs.preferred_hedge_soft.
3. Marginal band: prefs.hedge_threshold_soft <= a < prefs.qualified_band_high. Behavior depends on claim_strength:
   • "qualified" → prepend prefs.preferred_qualifier.
   • "affirmative" → leave bare.
   • "balanced" → leave bare.
4. Above marginal band: a >= prefs.qualified_band_high → leave bare regardless of claim_strength.

The thresholds must satisfy hedge_threshold_strong <= hedge_threshold_soft <= qualified_band_high; the loader enforces this ordering.

Three shipping pack profiles

Pack	strong	soft	qual_high	claim_strength	Hedge phrases
default_general_v1	0.40	0.50	0.75	balanced	"It seems that" / "Perhaps" / qualifier unused
precision_first_v1	0.55	0.70	0.85	qualified	"Arguably," / "In some cases," / "Under certain conditions,"
generosity_first_v1	0.20	0.30	0.50	affirmative	"It seems that" / "Perhaps" / qualifier unused

Result: at alignment = 0.45, default and precision both hedge but with different phrases; generosity leaves bare. At alignment = 0.80, precision qualifies; default and generosity leave bare. Visible divergence on identical trajectories — proven by tests/test_identity_surface_divergence.py.

Implementation

• core/physics/identity.py: new SurfacePreferences dataclass; IdentityManifold gains a surface_preferences: SurfacePreferences = SurfacePreferences() field with defaults that reproduce pre-ADR behavior.
• packs/identity/loader.py: _build_surface_preferences() parses and bounds-checks the new block; missing block uses defaults; threshold ordering enforced; claim_strength constrained to {"balanced", "qualified", "affirmative"}.
• generate/surface.py: SurfaceContext gains seven new fields (defaults preserve pre-ADR behavior); _apply_hedge takes the full context, not just a float, and implements the four-band algorithm above; the legacy module-level HEDGE_STRONG_THRESHOLD / HEDGE_SOFT_THRESHOLD constants are retained as the default values for SurfaceContext so existing imports still resolve.
• chat/runtime.py::ChatRuntime._build_surface_context: lifts self.identity_manifold.surface_preferences into the constructed SurfaceContext.
• packs/identity/*.json: three v1 packs gain surface_preferences blocks tuned to their roles.
• scripts/ratify_identity_packs.py: no change needed; runs again idempotently. Pack body changes → pack_source_sha changes → MasteryReport regenerated → companion .mastery_report.json rewritten → embedded mastery_report_sha256 updated.

Backward compatibility

• Pack format. Packs without surface_preferences continue to load and produce pre-ADR behavior.
• In-code SurfaceContext() construction. Callers who instantiate SurfaceContext() directly (without going through _build_surface_context) get default values that reproduce pre-ADR behavior. The legacy module-level constants HEDGE_STRONG_THRESHOLD = 0.4 and HEDGE_SOFT_THRESHOLD = 0.5 are preserved as the defaults for those fields, so any test or code that imports those names continues to work.
• Re-ratification cost. The three v1 packs are re-ratified once when the new block is added. Their MasteryReport SHAs change (this is expected). The previous SHAs are recorded in ADR-0027 §"Phase 5"; the new SHAs are recorded in docs/identity_packs.md §"Shipping packs (v1)" and in the ADR-0027 phase table.

Consequences

Positive

• The identity claim is now visibly load-bearing at the surface layer. Pack swap → different assembled string on the same prompt. The tests/test_identity_surface_divergence.py::TestPackSwapDivergence::test_same_alignment_different_surfaces asserts this explicitly.
• The realizer remains deterministic. No new operator, no normalization, no sampling, no clock, no PID, no hash-randomization. Same (prompt, alignment, pack) triple → same surface bytes.
• The pack format remains optional-block backward-compatible. Authors writing new packs may omit surface_preferences if defaults suit them.
• The divergence test makes regressions loud. Any future change that re-routes _apply_hedge or strips the SurfaceContext fields will fail test_same_alignment_different_surfaces immediately.

Negative / risks

• Schema bump invalidates the three Phase-5 MasteryReports. Cost paid once; re-ratification handled by the existing idempotent script. The previously-ratified SHAs from 2026-05-17 are superseded.
• English-only differentiation. _assemble_he and _assemble_grc do not currently consult SurfaceContext for hedge/claim-strength shaping; they call neither _apply_hedge nor anything sensitive to the new fields. Per-language hedging is a future concern; identity packs are language-neutral, so the same preferences will eventually drive the same logic in _assemble_he and _assemble_grc once those gain hedge support.
• The marginal-band qualifier only fires for claim_strength="qualified". A future axis profile that wants to expand claims in the marginal band (e.g., add "Indeed," before a confident assertion) would require either a new claim_strength value or a separate field. Out of scope for ADR-0028.
• SurfaceContext is now bigger. Seven new fields. The dataclass remains frozen+slots so the cost is small, but every construction site (including tests that build a SurfaceContext directly) must accept that the defaults are non-trivial. We rely on the defaults reproducing pre-ADR behavior; the cognition / runtime / smoke suites verify this.

Scope limits (explicit non-goals for this ADR)

• No depth-language (Hebrew, Koine Greek) hedging. Future work; new ADR.
• No scope, qualification_level, modal_style, or hedge_preferences (list) fields from the eval-layer YAMLs. Future work.
• No surface-side differentiation by axis-id (e.g., "the truthfulness axis is dominant → use evidential phrasing"). Future work, and likely requires a richer score-decomposition surface than today's scalar alignment.
• No new CLI verb. core chat --identity <pack_id> already exists.

Verification

This ADR is satisfied when:

• tests/test_identity_surface_divergence.py passes, in particular TestPackSwapDivergence::test_same_alignment_different_surfaces.
• All three v1 packs ratify cleanly under the v1 schema with the surface_preferences block and load in production mode (require_ratified=None).
• The cognition / runtime / smoke / formation / teaching suites are green at the same revision.
• tests/test_identity_packs.py continues to pass (loader bounds checks, runtime wiring, ratification-script idempotency, tamper detection).