core/docs/adr/ADR-0039-audit-completeness.md

ADR-0039: Audit Completeness — TurnVerdicts Bundle, Stub-Path TurnEvent, hedge_injected Signal

Status: Accepted (2026-05-17) Author: Joshua Shay + planner pass Companion docs: ADR-0035-turn-loop-verdict-surfacing.md, ADR-0036-safety-refusal-policy.md, ADR-0037-per-predicate-ethics-refusal.md, ADR-0038-hedge-injection.md

Context

After ADR-0035 → ADR-0038, the runtime emits three verdicts per turn (identity / safety / ethics) and two possible remediations (typed refusal, hedge injection). But auditing it had three rough edges:

1. Per-field correlation. An audit consumer had to read three separate fields (identity_score, safety_verdict, ethics_verdict) and infer the remediation by inspecting the surface text (does it start with the refusal prefix? with the hedge phrase?). Surface-text inference is brittle.
2. Stub-path TurnEvent gap. Stub turns (cold start, unknown domain) bypassed turn_log.append(). ADR-0035 already noted this as a known limit. The result: audit consumers reading the turn stream couldn't see stub turns at all, even though stub turns did carry a ChatResponse.safety_verdict and ChatResponse.ethics_verdict.
3. hedge_injected invisibility. Whether the runtime actually prepended a hedge this turn could only be detected by surface inspection (does it start with preferred_hedge_soft?). An audit consumer wanting "count hedged turns this hour" had to re-implement runtime decision logic.

The three rough edges share a root cause: the runtime knew the answers but didn't surface them in a form callers could read.

Decision

Three changes land together. Each is small; together they close the audit gap.

1. TurnVerdicts bundle type

A new frozen dataclass in chat/verdicts.py:

@dataclass(frozen=True, slots=True)
class TurnVerdicts:
    identity_score: object       # IdentityScore | None
    safety_verdict: object       # SafetyVerdict | None
    ethics_verdict: object       # EthicsVerdict | None
    refusal_emitted: bool        # ADR-0036 / ADR-0037
    hedge_injected: bool         # ADR-0038

Fields are typed object for the same reason TurnEvent.safety_verdict was: avoid coupling chat/verdicts.py to packs.* at module-resolution time. Audit consumers downcast at use site.

The bundle is attached to both ChatResponse.verdicts and TurnEvent.verdicts. The pre-existing individual fields (safety_verdict, ethics_verdict) remain — back-compat with ADR-0035 callers — but new consumers should read the bundle.

2. Stub-path TurnEvent emission

_stub_response now accepts an optional tokens kwarg. When invoked from a real turn (with tokens non-empty), it constructs and appends a TurnEvent to turn_log before returning the ChatResponse.

The stub event records:

Field	Value
turn	self._context.turn - 1 (after finalize_turn already ran)
input_tokens	tokens passed in
surface	typed refusal if it fired, else the unknown-domain marker
walk_surface	unknown-domain marker (preserved)
articulation_surface	unknown-domain marker (preserved)
identity_score	None (no trajectory ran)
cycle_cost_total	0.0
vault_hits	0
versor_condition	versor_condition(field_state.F)
flagged	False
safety_verdict / ethics_verdict / verdicts	computed verdicts

The correct() fallback path still calls _stub_response without tokens — that's a defensive call where no real "turn" happened, and appending a TurnEvent would mis-record the audit stream.

3. hedge_injected signal

The main turn path tracks whether the runtime actually mutated the surface during hedge injection:

before = response_surface
response_surface = inject_hedge(response_surface, hedge_prefix)
hedge_injected = response_surface != before

inject_hedge() is idempotent on prefix (ADR-0038) — if the surface already begins with the hedge phrase, the function returns it unchanged and hedge_injected stays False. This is the correct audit semantic: "did the runtime ADD a hedge this turn?", not "is there a hedge somewhere in the surface?"

Stub paths always report hedge_injected=False (ADR-0038 prohibits hedge on stub).

Consequences

Positive

• One field, full picture. An audit consumer reads response.verdicts (or event.verdicts) and gets identity + safety + ethics + remediation flags. No correlation across fields.
• Surface-inspection no longer needed. refusal_emitted and hedge_injected answer the runtime-decision questions directly.
• Stub turns are now first-class audit events. turn_log now covers the entire turn stream; downstream telemetry and replay systems can iterate over turn_log without missing stub paths.
• Mutual exclusion is verifiable. Existing test (test_refusal_and_hedge_never_both_true) confirms the runtime contract holds at the bundle level.
• No back-compat breakage. ADR-0035 / ADR-0036 / ADR-0037 / ADR-0038 individual fields and helpers still work.

Negative / risks

• Stub turns now have identity_score=None in the audit stream. Downstream consumers that assumed every TurnEvent carried an IdentityScore need to handle None. Mitigated: this was already true on ChatResponse.identity_score for stub paths; consumers reading the stream just had no stream entries at all before, and now they have entries with identity_score=None. The change is additive.
• The bundle duplicates per-field state. safety_verdict is reachable as both response.safety_verdict and response.verdicts.safety_verdict. Acceptable cost for back-compat; a future ADR could remove the per-field accessors if no in-tree consumer still uses them.
• hedge_injected doesn't distinguish "would have fired but idempotent" from "didn't fire at all." Two distinct cases collapse to hedge_injected=False. Audit consumers wanting that distinction can read the ethics_verdict directly.
• Hedge-injection test gate updated. The pre-ADR-0039 gate ("skip if turn_log empty") no longer discriminates stub vs main path; the test updated to gate on walk_surface == _UNKNOWN_DOMAIN_SURFACE instead. No semantic change.

Verification

• tests/test_turn_verdicts_bundle.py — 16 tests covering: bundle shape and frozen contract; ChatResponse and TurnEvent carry the bundle; stub path appends a TurnEvent with input tokens, unknown walk/articulation surfaces, identity_score=None, recorded versor condition; refusal_emitted flag toggles on forced safety violation, appears symmetrically on ChatResponse.verdicts and TurnEvent.verdicts; hedge_injected flag defaults False, stays False on stub paths even with opt-in; mutual exclusion (refusal supersedes hedge); response and event bundles agree on remediation flags and reference the same underlying verdicts.
• Combined pack-layer suite: 170 tests, all green (was 154 after ADR-0038).
• CLI suites unchanged: smoke 67, runtime 19, cognition 121.
• core eval cognition: intent 100%, versor_closure 100% — baseline preserved.

Open questions deferred to a future ADR

1. Structured-logging sink that consumes turn_log. Now that stub turns participate in the stream, a structured emitter could produce one log line per turn with pack_id, refusal_emitted, hedge_injected, and violated boundary ids.
2. core chat --show-verdicts CLI flag. Print per-turn verdict bundle summaries to stdout/stderr for manual audit.
3. Drop per-field accessors after a deprecation cycle. response.safety_verdict could become response.verdicts.safety_verdict-only once internal callers migrate.
4. identity_score on stub turns. Today stub paths skip the trajectory operator entirely and report None. A future ADR could compute a degenerate-but-valid IdentityScore for stubs so the audit stream is fully populated.
5. Bundle versioning. As more remediation tiers land (per-domain default policies, score-decomposition surfaces), the bundle may grow. Frozen-dataclass + optional defaults should scale fine, but a schema_version field could be added if downstream consumers need explicit versioning.