Determinism Tier 3: capture LLM reasoning and evidence trail in attestations, logs, and reports

[mlieberman85]

Summary

When a control's verdict is decided by an LLM consultation
(llm_eval handler resolving via the harness driver), the LLM's
reasoning is captured in the intermediate CheckResult but is
not serialized into attestations, is not logged, and is
not surfaced in markdown / SARIF report output (JSON contains it
if the caller opts into the full evidence dump).

An attestation without an evidence trail is a bare claim.
Reproducing an LLM verdict later requires re-running the LLM against
the same prompt, which the current system doesn't preserve.

This is Determinism Tier 3: nondeterministic results (LLM outputs
have unavoidable model-side variance) are legitimate, but the
reasoning that produced them must be recoverable.

Concrete gaps

Attestation predicate

• packages/darnit-baseline/src/darnit_baseline/attestation/predicate.py:82-102
  -- per-control attestation records: id, level, category,
  status, message, evidence, source, and (post feature 025)
  authority. It does NOT record:
  • Which sieve pass fired (file_must_exist, exec, pattern,
    llm_eval, manual).
  • For exec: the command, stdout, exit code.
  • For llm_eval: the prompt sent, the model's reasoning, the
    model version, token usage.
  • pass_history -- the cascade of attempts across phases.

LLM consultation response shape

• packages/darnit/src/darnit/sieve/models.py:81-88 --
  LLMConsultationResponse carries status, confidence, a single
  reasoning: str, and evidence_cited: list[str]. No structured
  evidence trace, no model version, no request-id for correlation.

Report formatters

• Markdown / SARIF do NOT surface LLM reasoning. Full JSON does
  (via the evidence dict) but the summary JSON strips it.
• Markdown surfaces resolving_pass_handler and pass_history (per
  tools/audit.py:918-931) so operators can see WHICH pass resolved
  the control, but not the LLM's reasoning for the resolution.

Logging

• packages/darnit/src/darnit/sieve/orchestrator.py -- no per-pass
  disposition logging (PASS/FAIL/INCONCLUSIVE transitions across the
  cascade are not emitted).
• packages/darnit/src/darnit/sieve/builtin_handlers.py -- no
  logging of exec command invocation or LLM reasoning at INFO.

LLM-based remediation

• packages/darnit-baseline/src/darnit_baseline/remediation/enhancer.py:56-133
  -- LLM enhancement is invoked via an llm_fn() callback with no
  capture of the request/response, no operator sign-off checkpoint
  before applying LLM-suggested changes. This is a Principle IV
  concern in addition to a determinism concern.

Scope

• Extend the attestation predicate per-control fields to include:
  • resolving_pass_handler and resolving_pass_index.
  • pass_history (already in CheckResult; propagate through).
  • For llm_eval resolutions: llm_reasoning, llm_model, and
    llm_confidence. Prompt inclusion is opt-in per attestation
    profile (large; may be redacted for privacy).
  • error_class when present (from Determinism Tier 2: network / side-effect failures must be distinguishable from real verdicts #419).
• Extend LLMConsultationResponse to carry model and (optionally)
  request_id.
• Extend markdown and SARIF formatters to surface LLM reasoning
  when present.
• Add INFO-level logs at: LLM request dispatch, LLM response receipt,
  per-control final disposition.
• For LLM-based remediation: log request/response at INFO; add an
  operator-review checkpoint in the remediation orchestrator before
  applying LLM-suggested changes (interactive prompt when TTY;
  fail-loud when non-interactive without explicit
  --accept-llm-suggestions).

Non-goals for v0

• Persisting the full LLM prompt in every attestation by default
  (size concerns; opt-in per profile).
• Storing LLM raw token streams / chain-of-thought (privacy and
  size; final reasoning string is sufficient for audit trail).
• Alternative LLM providers or a model-selection UI.

Constitution alignment

• II. Conservative-by-Default: an LLM PASS without recorded
  reasoning is not a defensible audit result.
• IV. Never Guess User Values: the LLM-remediation review
  checkpoint is a hard implementation of this principle -- LLM
  output alone MUST NOT apply changes.
• V. Sieve Pipeline Integrity: the sieve records which pass
  resolved a control; the attestation should record the same.

Related

• Determinism Tier 1: normally deterministic operations must be deterministic across runs #418 (Determinism Tier 1) and Determinism Tier 2: network / side-effect failures must be distinguishable from real verdicts #419 / Determinism Tier 2: remediation must be atomic and idempotent under partial failure #420 (Tier 2) -- companions.
• Feature 025 / RFC-0001 Stage 1 -- introduced authority;
  attestation already carries it. This issue extends the same file
  with additional per-control fields.
• Feature 026 / harness driver -- LLM dispatch lives there; the
  reasoning-capture wiring goes on the harness->attestation path.