Skip to content

sum-engine 0.8.1

Latest

Choose a tag to compare

@github-actions github-actions released this 20 Jun 15:10
· 15 commits to main since this release
b5c6fdb

Headline: the front door is now honest and frictionless for a first-touch user.
pip install "sum-engine[verify]" then python -m sum_verify --demo replays a real
meaning-loss bound fully offline (the binding-gate golden now ships in the wheel); the
in-browser Verify box checks the flagship meaning receipt; the verify verdict surfaces
controlled / bound / n (a vacuous small-n receipt can no longer pose as a strong one);
and sum frontier --distill makes the offline distiller demoable from one source. Plus
README/CLI doc-currency + the perspective-spec correction. All backward-compatible
additions + fixes over 0.8.0.

  • Frontend↔backend concordance correction (delivery). A 5-seam audit found the
    front door selling an older, narrower product than the v0.8.0 backend is — the
    leading explanation for ~10k clones / 9 stars / 0 observed users. Fixed: the live
    demo hero ("Every fact becomes a prime number / Math does the trust") → "Transform
    text with AI — and carry a receipt for it / Attest, don't detect" (the Gödel-prime
    encoding demoted from headline to mechanism); the README h1 ("verifiable
    bidirectional knowledge distillation") → "chain of custody for AI-transformed text"
    • a v0.8.0 feature row (meaning-loss receipts + sum_verify); a front-door signpost
      for sum meaning-diff (the #1 adoption-sim demand, previously absent) + the
      sum-engine[verify] SDK CTA; and the slider bench note de-overclaimed (corpus-level,
      not certified for your document; the live render does not re-measure). Bug fixed:
      the in-browser mint used Math.ceil for density while Python apply_density and the
      Worker use floor — so density<1.0 kept more facts than the slider implied;
      now Math.floor (concordant; cross-runtime harness green). Slider capability itself
      was found genuinely live + concordant (axis prompts byte-identical Python↔Worker).
      The hero + density fix are LIVE (deployed in #313; the 2026-06-18 concordance audit
      confirmed live == repo byte-for-byte). The last front-door deferral — wiring the
      shipped meaning_receipt_verifier.js into the visible Verify box — is now CLOSED
      (see the front-door-hardening entry below); a redeploy of single_file_demo/ makes it live.
  • Front-door hardening from the 2026-06-18 audit (verify on-ramp + honesty + browser meaning-verify).
    The audit found every concordance break clustered at the v0.8.0 flagship meaning
    receipt. Closed the closeable-now ($0, no-puller) set:
    • On-ramp (ISS-2): the documented first command silently needed a git clone —
      examples/ + fixtures/ ship zero files in the wheel. Bundled the real BillSum
      binding-gate golden (CC0) as package data; python -m sum_verify --demo now
      replays a meaning-loss bound fully offline straight from pip, no clone. README
      leads with it; THIRD_PARTY_VERIFY.md got the clone it needs.
    • Honesty (ISS-4/5): python -m sum_verify now surfaces risk_upper_bound +
      controlled + n (a vacuous small-n receipt no longer looks identical to a strong
      one); the issuance script warns when a bound is near-vacuous; RECEIPT_FAMILY_SPEC
      no longer implies the [verify] SDK covers perspective (it is [research]-tier).
    • Browser meaning-verify (ISS-3): wired the shipped meaning_receipt_verifier.js
      into the demo's Verify surface — a new "Verify a meaning receipt" box runs Stage A
      (signature + schema + disclosure) in-browser with schema-dispatch + a one-click
      BillSum sample, honestly labelled (Stage B replay stays Python-side; the proxy
      caveat travels). Closes the last front-door deferral above.
  • sum frontier --distill — the distiller demoable from one source, offline.
    sum frontier previously only scored pre-made --version files; the offline
    distiller path existed in the transform registry but was unwired (from_render_fn
    reachable only from tests). --distill now generates the faithful→compressed path
    from --source itself — deterministic sieve extraction (text→triples) →
    SliderTransform canonical path at descending density → RenderFrontier — fully
    offline, zero-$, no LLM, no network. --steps N (default 5) / --density-floor F
    (default 0.1; faithful end always 1.0); mutually exclusive with --version.
    Research surface ([research] + [sieve]). (fresh-eyes audit follow-through)
  • proxy_caveat at the point of verification (sum_verify). Every verified
    meaning-risk verdict from the dependency-light [verify] SDK (python -m sum_verify
    and the SDK docstring) now carries a corpus-agnostic caveat: a clean PASS is a
    cryptographic fact (signature + replayed bound), not proof meaning was
    preserved — the bound is over a proxy that tracks human judgment only modestly
    (Spearman ρ≈0.27–0.33 on SummEval). UNSIGNED surfaces only; no number is baked into
    a signed field (the SummEval ρ was measured on a different corpus+judge).
  • Front-door honesty (README + on-ramp). A plain-language lede that names the
    distiller ahead of the crypto machinery (it was absent before); the lead "verify
    it yourself" demo swapped from the commodity render-receipt curl to the offline
    python -m sum_verify replay of the committed BillSum binding-gate golden
    (examples/verify-meaning-offline/run.sh), keeping the render trust loop as a
    secondary block; sum inspect help example corrected (-i PATH/stdin, no
    positional).