The foundation of Verified Completions (tracking issue #89), shipped as one verified unit: design, engine-validated rule pack, executable benchmark with per-scenario verification, tests at every level, and the docs that explain it.

The proven core loop

Conversation text becomes atomic claims with quoted evidence, claims become facts in a per-conversation knowledge graph, and the consistency rule pack derives findings with proof trees. poc/poc_verify.py runs that loop end to end (extract -> validate -> load -> judge); the scorer and ingestion validator are negative-tested (fabricated quotes get dropped, impossible expectations fail rows).

The rule pack, validated on the real engine (#81)

• Restructured to one clause per detection relation with unions only in reporting views - works around engine bug Engine: multi-clause rules silently return empty or malformed results when clauses share a derived relation #91 (multi-clause rules silently evaluating empty), external query interface unchanged
• eq identity helper symmetrized at the base so .why proof queries terminate
• Ontology seeds hardened by review: shareable identifiers (email, order_number, booking_reference) removed from inverse_functional; attr_domain seeded so the domain rule can fire
• Enforced by CI: the fixture runs as an e2e snapshot test (examples/iql/42_consistency_pack/), full suite green

The benchmark (corpus v2.1)

poc/corpus.py deterministically generates 1,628 scenarios across 16 families (~100 each, all 48 sub-variants at n>=33) covering value, negation, temporal, spatial, causal, structural, numeric, counting, identity, classification, and instruction corruption, plus 102 controls; 449 scenarios plant the conflict 12+ turns deep. Every sample carries multi-valued category labels, exact conflicting spans, a clean twin, a natural task, and extractor-truth facts. The corpus is documented for consumers in poc/DATASET.md (design principles, scenario anatomy, per-scenario validation, how to run and extend it) and regenerates byte for byte.

Per-scenario verification (poc/verify_each.py, ledger committed): spans/clean-twins/labels 1,628/1,628; engine pass 1,526/1,526 with EXACT finding-kind match - the engine fires precisely the expected kinds and nothing else, on every corrupted scenario, deterministically.

Completed behavioral study (corpus v1, 12 families at n=30, Wilson CIs, double-graded): corruption drops sound outputs from 95% to 87% [83-90] - 38 silent commits to arbitrary values, 4 contradictions written into deliverables; attaching the InputLayer finding restores 98% [96-99]. Corrupted system prompts are the standout: 17% sound alone -> 93% with the finding. Raw data archived in poc/results/archive_v1_full_bench_390.json. The v2.1 LLM re-baseline continues in #95 (resume-until-verified tooling is in this PR; 100/1,628 done).

Also in this unit

• Two-regime harness (poc/full_bench.py): asked-to-check vs asked-a-real-task, double grading with a different-model arbiter (Opus) on disagreements, tiered runs (smoke/standard/full), --resume, subscription-OAuth-first auth
• Snapshot-harness fix (scripts/run_snapshot_tests.sh): test server starts from a generated config with WS rate limits off; client key read from the credentials file (--config mode rejects the old env var - Config: documented auth env vars break --config startup; default path ignores env overrides; stale config.toml template #92)
• Coverage audit (docs/COVERAGE-AUDIT.md) mapping the full space of logic-resolvable inconsistency; gaps tracked in Benchmark coverage: close the gaps found by the consistency-type audit #94
• Repo hygiene pass: session-handoff notes, the superseded extraction prompt, and pilot harnesses removed; all references fixed
• Feature one-pager (docs/internals/verified-completions/README.md, mermaid diagrams) and researcher-grade ontology spec (docs/ontology.md: extraction contract, type system, every verification rule verbatim)
• Beginner-facing blog article (docs/blog/consistency-ontology.md): the ontology as a build-it-yourself journey
• PLAN.md tracking all of the above against issues Verified Completions Phase 1: run the consistency rule pack on the real engine #81-Benchmark: finish the v2.1 LLM re-baseline (1,528 scenarios remaining) on subscription auth #95

Verified on this exact tree

cargo fmt, clippy (-D warnings), full cargo test suite, e2e snapshots (1,142 passed / 0 failed), IQL lint on pack + fixtures, corpus regeneration byte-identical, per-scenario ledger, secret scan clean.

Engine bugs found and filed along the way: #91 (silent empty multi-clause unions), #92 (config/env-var conflicts), #93 (inserts ack while storing nothing after data-dir loss).

Closes #81.