Proof report
SUM — Verifiable Bidirectional Knowledge Distillation
Every verifiable claim behind this project, grouped by epistemic status, each with a source you can check — 11 claims, 11 source-linked.
Project status: Beta · last verified 2026-08-28
← SUM — Verifiable Bidirectional Knowledge Distillation overview Repository ↗ Live ↗
Proved
(2)Mechanically proven. Enforced by deterministic code and verified by tests, including cross-runtime witnesses (epistemic status: `provable`) — not a formal mathematical proof.
-
Three-runtime byte-symmetric Ed25519 over JCS bytes (Python ↔ Node ↔ browser)
-
Canonical round-trip reconstruct(parse(canonical_tome(S))) == S — 0.00% drift on every CI run
Empirical benchmark
(4)Measured under a reproducible harness; numbers cite the corpus and command.
-
Slider fact preservation median 1.000 (p10 0.769 long n=16 / 0.818 short n=8) — measured; same-commit replay receipt pending (bench-hardening T2/T3)
-
Extraction F1 1.000 on seed_v1 (50 hand-authored single-fact SVO sentences); precision 1.000, F1 0.762 on seed_v2 (20 hand-authored harder prose patterns — apposition, passive, relative clause, conjunction). Synthetic author-written corpora; no clinical, customer or third-party text has been measured. Reproduce: python -m scripts.bench.run_bench --corpus scripts/bench/corpora/seed_v1.json --no-llm
-
T1 iterated round-trip K=10 stability — STABLE on all three corpora (seed_v1, seed_v2, seed_long_paragraphs)
-
First real certified chain, over the first 32 BillSum test bills (CC0-1.0): per-hop 95%-confidence strict-NLI meaning-loss upper bounds of 0.865768 (hop 1, abstractive summarization performed by the dataset itself, not by SUM) and 0.488860 (hop 2, deterministic lead-N extractive compression), composed to a Bonferroni budget ≤ 1.354628 at joint confidence 0.90; separately measured direct end-to-end leg ≤ 0.874216. The bounds are wide at n=32 (Hoeffding), and the budget is not the end-to-end loss.
Shipped
(5)The named artifact (test, doc, code path) exists in the repo today and is runnable. Per-claim, not per-project — a claim can be shipped while the parent project is still in development.
-
Render receipt format sum.render_receipt.v1 (Ed25519 / JCS / detached JWS) — verifier in three runtimes
-
Transform substrate (sum.transform_receipt.v1 + registry: slider / extract / compose) — 20-fixture cross-runtime K-matrix locks accept + reject across Python ↔ Node ↔ browser; T4 source-chain binding, T5 ShareableRender, T6 multi-school extract shipped in v0.7.0
-
sum verify --explain layered output (sum.verify_explained.v1) — seven per-dimension checks each tagged with epistemic_status; truth-of-content always not_asserted (test-locked invariant)
-
Compliance validators across six regimes (EU AI Act Art. 12 / GDPR Art. 30 / HIPAA §164.312(b) / ISO 27001 A.8.15 / SOC 2 CC 7.2 / PCI DSS v4.0 Req 10) — sum compliance check emits sum.compliance_report.v1
-
Certified multi-hop chain receipt sum.chain_receipt.v1 (Ed25519 / JCS / detached JWS over an ordered sequence of meaning-risk hops) — issued by `sum mint-chain`, replayed offline by a third party via `python -m sum_verify`; the mandatory budget_scope field is fail-closed. Research tier for issuance and dependency-light verification.