evidentia-eval 0.11.0

DFAH (Decision-Faithfulness Assessment Harness) determinism + faithfulness eval harness for Evidentia — dev-time AI-output quality gates

evidentia-eval

Dev-time AI-output quality eval harness for Evidentia.

Hosts the DFAH (Decision-Faithfulness Assessment Harness) — the auditor-defensible numerical proof layer that validates LLM-driven artifact production is deterministic, replay- equivalent, and faithful to its source policy clauses.

Why this package exists (v0.10.5 P9 extraction)

The DFAH harness was originally bundled into evidentia-ai (the risk-statement generator + control explainer package). That conflated two very different deployment surfaces:

• evidentia-ai — PRODUCTION runtime. Needed in air-gap installs to actually generate risk statements.
• evidentia-eval — DEVELOPMENT-time evaluation. NOT needed in air-gap installs; only fires when a CI pipeline runs a determinism / faithfulness gate before tagging a release.

Extracting the eval harness lets air-gap installs of evidentia-ai skip the optional sentence-transformers stack entirely (it now lives behind evidentia-eval[faithfulness-semantic] instead of evidentia-ai[eval-faithfulness]).

Quick start

# Stdlib Jaccard baseline (no extra needed; <10 MB install)
pip install evidentia-eval

# Optional semantic-similarity faithfulness (~250 MB extra
# for sentence-transformers + numpy + model cache on first use)
pip install 'evidentia-eval[faithfulness-semantic]'

CLI verbs:

# Smoke test against a deterministic stub generator (no LLM
# tokens burned)
evidentia eval stub-smoke

# Real-LLM determinism gate against the risk-statement generator
evidentia eval risk-determinism --gap-report gaps.json \
    --system-context ctx.yaml \
    --fail-on-determinism-rate-below 0.95

# Verify a previously-signed eval bundle
evidentia eval verify path/to/eval-output.json

The CLI verbs live in evidentia.cli.eval (the meta-package); this package contributes the underlying library.

Public API

Symbol	Purpose
DFAHarness	Owns the run loop + audit emit
EvalResult	Top-level harness output (JSON-serializable, Sigstore-signable)
EvalSample	One prompt's inputs (immutable; audit-trail-stable)
DeterminismResult	Per-prompt determinism outcome
ReplayResult	Per-prompt replay-equivalence outcome
FaithfulnessResult	Per-claim faithfulness outcome
PromptFaithfulnessResult	Aggregated per-prompt faithfulness
faithfulness_score	Stdlib Jaccard token-overlap baseline
faithfulness_score_semantic	Sentence-transformers path (optional extra)
determinism_score	Computes the modal-output pass rate
replay_equivalent	Binary replay-equivalence check
extract_claims	Atomic-claim extraction from generated artifacts
normalize_for_determinism	Canonical whitespace + punctuation normalization
hash_output	SHA-256 hex of normalized output
sign_eval_result	Sigstore-sign an EvalResult JSON
verify_eval_result	Verify a previously-signed eval bundle

Backward-compat shim

For external scripts that still import from evidentia_ai.eval import ..., evidentia-ai ships a deprecation shim that re-exports from evidentia_eval. The shim warns once at import time and is scheduled for removal in v0.12.0.

License

Apache-2.0. See the workspace root LICENSE file.