SUM — verifiable bidirectional knowledge distillation

A cross-runtime trust surface for LLM-rendered text. Three runtimes (Python, Node, modern browsers) produce byte-identical Ed25519 signatures over the same JCS-canonical bytes. Every render through the hosted Worker carries a detached-JWS receipt (sum.render_receipt.v1) that any third party can verify offline against /.well-known/jwks.json. Live at https://sum-demo.ototao.workers.dev.

That is the load-bearing claim and what makes SUM different from a generic summarisation tool. The cryptographic side is mechanically proven — three independent verifier implementations agreeing byte-for-byte on every signed bundle, locked in CI on every PR. The semantic side (extraction quality, slider fact preservation) is empirically measured with explicit per-corpus numbers and explicit per-corpus boundaries; SUM does not blur the line between the two. docs/PROOF_BOUNDARY.md is the arbiter.

Headline supporting numbers (each links to its source of truth):

Claim	Status	Source
Three-runtime byte-symmetric Ed25519 over JCS bytes	provable; locked by make xruntime (K1–K4) + make xruntime-adversarial (A1–A6)	docs/PROOF_BOUNDARY.md §1.2, §1.3.1
Canonical round-trip reconstruct(parse(canonical_tome(S))) == S	provable; 0.00% drift on every CI run	docs/PROOF_BOUNDARY.md §1.1
Render receipt — sum.render_receipt.v1, Ed25519 / JCS / detached JWS	shipped; verifier in three runtimes	docs/RENDER_RECEIPT_FORMAT.md
Slider fact preservation: median 1.000, p10 0.769 (long n=16) / 0.818 (short n=8)	empirical-benchmark	docs/SLIDER_CONTRACT.md
Extraction F1 = 1.000 (seed_v1), 0.762 with precision 1.000 (seed_v2)	empirical-benchmark	docs/PROOF_BOUNDARY.md §2.1

A render receipt verifies the render attestation (issuer signed this tome, these triples, this slider position, this model, at this time). It does not verify the truth of the tome's content — that is what the slider bench measures separately. See docs/RENDER_RECEIPT_FORMAT.md §5 for the explicit trust scope.

---

Verify it yourself in 60 seconds

The trust loop: hit the live Worker, get back a tome plus a detached Ed25519 JWS over the JCS-canonicalised receipt payload, fetch the issuer JWKS, verify.

# 1. JWKS — single Ed25519 OKP JWK, application/jwk-set+json
curl -sS https://sum-demo.ototao.workers.dev/.well-known/jwks.json | jq .
# → {"keys":[{"crv":"Ed25519","kty":"OKP","x":"...","alg":"EdDSA","use":"sig","kid":"sum-render-2026-04-27-1"}]}

# 2. Render — tome + render_receipt (signed JWS over JCS payload)
curl -sS -X POST https://sum-demo.ototao.workers.dev/api/render \
  -H 'content-type: application/json' \
  -d '{"triples":[["alice","graduated","2012"],["alice","born","1990"]],
       "slider_position":{"density":1.0,"length":0.5,"formality":0.7,"audience":0.5,"perspective":0.5}}' \
  | jq '.render_receipt | {schema, kid, payload, jws_segments: (.jws | split(".") | length)}'

A minimal Node verifier using jose + canonicalize is in docs/RENDER_RECEIPT_FORMAT.md §A.5; the same format is reachable from Python (joserfc + jcs), Go, and Rust per §3.

---

What ships today

Surface	Status	Verifies
pip install 'sum-engine[sieve]' — sum attest / sum verify / sum resolve / sum ledger / sum inspect / sum schema	shipped on PyPI	structural reconstruction; HMAC-SHA256 + Ed25519 signatures (W3C VC 2.0 eddsa-jcs-2022)
Cloudflare Worker at sum-demo.ototao.workers.dev	shipped	/api/render → tome + render_receipt; /.well-known/jwks.json → JWKS; /api/qid → Wikidata resolver
Single-file browser demo (single_file_demo/index.html)	shipped	paste prose → in-browser attest → CanonicalBundle JSON; same bytes verify under node standalone_verifier/verify.js (Chrome / Firefox / Safari with WebCrypto Ed25519 support)
Cross-runtime trust triangle	locked by CI (make xruntime)	K1 / K1-mw / K2 / K3 / K4 — Python ↔ Node ↔ Browser agree byte-for-byte on valid bundles. make xruntime-adversarial adds A1–A6 rejection-class equivalence.
5-axis slider rendering surface	density actioned deterministically; length / formality / audience / perspective LLM-conditioned via the Worker (Anthropic, Cloudflare AI Gateway optional)	bench: median LLM-axis fact preservation 1.000, p10 0.769 (long, n=16) / 0.818 (short, n=8), order preservation 1.000 wherever measurable
MCP server (sum-mcp console script)	shipped	five tools (extract / attest / verify / inspect / schema) exposed over stdio; bundles attested via MCP verify byte-identically through the CLI / Node / browser verifiers

The slider's product claim — axis changes do not lose facts — is the load-bearing empirical result. It is verified by NLI audit on every embedding-flagged "loss" cell; full attribution in docs/SLIDER_CONTRACT.md.

What does NOT yet work — the honest line

SUM measures one capability that the rest of this README's numbers do not yet close: the full LLM narrative round-trip (text → LLM-extract → axioms → LLM-generate → prose' → LLM-extract → axioms'). On seed_v1 this loop produces:

• 107.75% drift (per-document 100 × |A Δ A'| / max(|A|, |A'|)) and
• exact-match recall = 0.12 (6 of 50 source triples appear verbatim after the round trip).

The reason both numbers exist together: the generator preserves facts (FActScore 0.94–0.96, §2.4) but the round-trip is dominated by generator elaboration — the LLM produces ~12 reconstructed axioms per source axiom and elaborates around the source claim rather than paraphrasing it.

Canonicalisation alone does not close this gap — measured 2026-04-28 (scripts/bench/runners/canonicalization_replay.py, no LLM cost, operates on cached per-doc data):

Canonicalisation regime	drift_pct	exact-match recall
baseline	107.75 %	0.12
+ predicate normalisation	107.75 %	0.12 (zero movement — the L1 falsification)
+ subject canonicalisation (last-word-as-key)	106.68 %	0.16
+ aggressive object normalisation (ceiling)	106.36 %	0.18

Closing the §2.5 gap requires moving the generator (constrained decoding to a pinned vocabulary, or a fidelity-objective fine-tune), not just post-hoc key normalisation. The L0–L3 receipt is the reference baseline against which any future generator-side intervention is measured. Full attribution in docs/PROOF_BOUNDARY.md §2.5.

The deterministic canonical round-trip (the one sum attest | sum verify exercises) is mechanically proven (§1.1, 0.00% drift). The LLM round-trip is not, and this section is here to keep that distinction above the fold.

---

CLI quick start

pip install 'sum-engine[sieve]'

echo "Alice likes cats. Bob owns a dog." \
  | sum attest --extractor=sieve > bundle.json

sum verify --input bundle.json
# → sum: ✓ verified 2 axiom(s), state integer matches (hmac=absent, ed25519=absent)

Add cryptographic attestation with one flag:

# Ed25519 / W3C VC 2.0 (eddsa-jcs-2022)
python -m scripts.generate_did_web --domain your.example --private-key-out keys/issuer.pem
sum attest --ed25519-key keys/issuer.pem < prose.txt | sum verify --strict
# → hmac=absent, ed25519=verified

The same bundle bytes verify under sum verify (Python), node standalone_verifier/verify.js (WebCrypto), and the in-browser demo (SubtleCrypto). docs/DID_SETUP.md walks the did:key / did:web issuer setup. docs/PROOF_BOUNDARY.md §1.3.1 documents what the cross-runtime Ed25519 contract proves.

Calling SUM from MCP-aware LLM clients

pip install 'sum-engine[mcp,sieve]'
# Claude Desktop / Claude Code / Cursor / Continue: add to MCP config:
#   { "mcpServers": { "sum": { "command": "sum-mcp" } } }

sum-mcp exposes extract, attest, verify, inspect, schema as MCP tools. Bundles attested via MCP verify byte-identically through the CLI / Node / browser verifiers — same canonical codec. See docs/MCP_INTEGRATION.md for the full client wiring.

Calling SUM over HTTP

The hosted Worker at https://sum.ototao.com exposes /api/render, /api/complete, /api/qid, and the /.well-known/{jwks,revoked-kids}.json verification surfaces. docs/API_REFERENCE.md is the wire spec — request/response shapes, error codes, the six-step receipt-verification flow, working Node + Python examples. Use this when the caller is a web app, mobile app, or server-side service; use the MCP server when the caller is a local LLM client.

---

How the trust loop fits together

prose ─► /api/render ─►  tome
                         + render_receipt {kid, payload, jws}
                                              │
                                              ▼
                                  /.well-known/jwks.json
                                  (Ed25519 OKP JWK by kid)
                                              │
                                              ▼
                              jose.flattenedVerify(JCS(payload))
                                              │
                                              ▼
                          render attested ✓ — issuer signed
                          (this tome, these triples, this slider
                          position, this model, at this time)

The receipt is a render attestation, not a truth oracle. Fact preservation is verified by the bench (NLI audit on weak cells). The receipt is what a downstream system keeps as durable proof; the tome is what a reader consumes. See docs/RENDER_RECEIPT_FORMAT.md §5.

---

Underlying substrate

Below the slider sits the substrate that earlier phases shipped and verified. Pointers, not paraphrase — every claim links to its source-of-truth doc.

• Canonical round-trip conservation (provable). reconstruct(parse(canonical_tome(S))) == S for every Gödel state S. 0.00% drift on seed_tiny_v1 / seed_v1 / seed_v2. docs/PROOF_BOUNDARY.md §1.1.
• Cross-runtime state equivalence (provable). Python (sympy), Node (BigInt + Miller-Rabin), in-browser JS produce byte-identical state integers. Locked by 4 harnesses (make xruntime + make xruntime-adversarial). docs/PROOF_BOUNDARY.md §1.2.
• Bundle public-key attestation (provable). Ed25519-signed CanonicalBundles are tamper-detectable by any third party in any of the three runtimes. docs/PROOF_BOUNDARY.md §1.3.1.
• Merkle hash-chain integrity (provable, including under concurrent writers). docs/PROOF_BOUNDARY.md §1.7.
• Extraction F1 (empirical-benchmark). 1.000 on seed_v1 (50 simple-SVO docs); 0.762 with precision 1.000 on seed_v2 (20-doc difficulty corpus). Every remaining seed_v2 failure is a recall miss, not a truth inversion. docs/PROOF_BOUNDARY.md §2.1.
• 103 numbered features, each with a reproducible verification command, in docs/FEATURE_CATALOG.md.

---

Reproduce the bench

# Short corpus (n=8, 4–12 triples/doc, ~$0.30, ~2 min with NLI)
bash scripts/bench/run_paragraphs.sh

# Long corpus (n=16, 9–24 triples/doc, ~$1.50, ~10 min with NLI)
bash scripts/bench/run_long_paragraphs.sh

Both runners require OPENAI_API_KEY (NLI audit + extraction). Pinned model snapshots are mandatory; the harness raises SystemExit on unpinned identifiers (see docs/PROOF_BOUNDARY.md §2.6). Output is NDJSON sum.slider_drift_bench.v1, with per-cell strict / normalized / semantic / NLI fact-preservation columns.

---

Future developments

This roadmap names only unshipped work. Items already landed live in CHANGELOG.md [Unreleased]. Detailed sequencing lives in docs/NEXT_SESSION_PLAYBOOK.md.

Closing the LLM round-trip drift. This is the headline open problem. The full LLM round-trip (text → LLM-extract → axioms → LLM-generate → prose' → LLM-extract → axioms') currently produces 107.75 % drift and 0.12 exact-match recall on seed_v1 — facts preserved, keys not. Closing this gap is a canonicalisation problem (entity resolution, predicate normalisation, pinned-vocabulary extraction); none of those passes are shipped yet. See docs/PROOF_BOUNDARY.md §2.5 for the full attribution and per-document failure modes.

Hardening backlog

• sha256_128_v2 activation — Node side exists, Python side not yet CURRENT_SCHEME. Pre-empts the 2³² collision frontier.
• /api/qid SPARQL disambiguation — moves entity resolution from the current wbsearchentities-only path to a target >95 % accuracy floor (the floor itself is unmeasured today).
• Threat-model validation — every documented defence in docs/THREAT_MODEL.md gets an executable test.
• Delta-bundle composition semantics — specifies what bundle.is_delta means cross-runtime.
• Sigstore / cosign signing of release artifacts.
• LLM-extraction honesty guardrails — extraction.verifiable: true | false so signed ≠ true is visible at the consumer interface.
• Calibration-set authoring for the Venn-Abers conformal-interval implementation that already ships.
• Remaining sieve recall work on seed_v2 (apposition / relative-clause / compound-conjunct) — gated on the §2.5 work, see docs/PROOF_BOUNDARY.md §6.

Platform surface (post-hardening)

Source anchoring in the bundle schema, bundle explorer / viewer, sum verify --explain, sum tutorial onboarding, shareable bundle URLs /b/{hash}, PWA-installable demo, sum attest <url> fetch mode. Each item names its dependency in docs/NEXT_SESSION_PLAYBOOK.md.

---

Verification surface

make help lists every dev command. Common targets:

make install              # editable install with sieve + dev extras
make test                 # full pytest run (1000+ tests)
make xruntime             # cross-runtime K1/K1-mw/K2/K3/K4 (Python ↔ Node)
make xruntime-adversarial # rejection-matrix A1–A6
make fortress             # 21-check pure-math invariants
make smoke                # fresh-venv install + attest|verify round-trip
make demo                 # open the single-file browser demo

CI runs the full suite on every push (.github/workflows/quantum-ci.yml); the cross-runtime-harness job runs K1–K4 + A1–A6 on Node 22; pypi-install-smoke builds the wheel and runs echo prose | sum attest | sum verify in a throwaway venv.

---

Truthfulness contract

Every claim in this repo carries an explicit epistemic status — provable, certified, empirical-benchmark, or expert-opinion. The arbiter is docs/PROOF_BOUNDARY.md. A summary surface that quotes an empirical-benchmark number alongside language like "mathematically guaranteed" is a policy violation per §5 and must be corrected.

Performance language (fast, efficient, low-latency, scalable) requires a benchmark in the same commit. Adversarial input agreement (the A-matrix) is a separate proof from valid-input agreement (the K-matrix); both run in CI.

If a number in this README disagrees with docs/PROOF_BOUNDARY.md or docs/SLIDER_CONTRACT.md, the docs are canonical and this README is wrong.

---

Contributing

1. Fork and branch.
2. make install && make test && make xruntime.
3. Read docs/NEXT_SESSION_PLAYBOOK.md for principles, stop-the-line triggers, and the work-ordering rule.
4. Open a PR. Every claim added to docs or commit messages must trace to a test, a measurement, or an explicit designed, not proved label.

CONTRIBUTING.md has the test-gate matrix and the verification-gate runbook.

---

License

Apache 2.0. See LICENSE.

Minor-bump feature release. Agentic-first introspection surface on the sum CLI, zero breaking changes.

New subcommands

sum ledger list --db DB [--axiom K] [--since ISO] [--limit N]   # NDJSON prov_ids
sum ledger stats --db DB [--pretty]                             # counts, ts range, chain tip
sum ledger head --db DB [--branch NAME] [--pretty]              # state integer per branch
sum inspect bundle.json [--pretty]                              # structural read, no crypto
sum schema {bundle|provenance|credential}                       # JSON Schema Draft 2020-12

Each answers a question an LLM agent composing SUM into a larger pipeline previously had to craft raw SQL or docstring-archaeology for.

Why it matters for agents

• ledger list turns SQLite-backed provenance into NDJSON agents can jq.
• inspect reads bundle shape without paying Ed25519 verification cost — agents can route by bundle attributes before deciding whether to run full verify.
• schema emits ground-truth JSON Schemas so agents validate SUM output programmatically instead of guessing from prose.

State integers are always string-encoded (never JSON numbers) to preserve arbitrary precision — many agent parsers use 64-bit doubles.

Unchanged

• CLI contract for attest / verify / resolve and every flag.
• CanonicalBundle wire format (canonical_format_version 1.0.0).
• Prime scheme (sha256_64_v1).
• Cryptographic contracts (HMAC, Ed25519, W3C VC 2.0 eddsa-jcs-2022).
• Cross-runtime trust triangle — K1/K1-mw/K2/K3/K4 still green; same bundle bytes still verify in Python ↔ Node ↔ Browser.

Tests

14 new cases in Tests/test_sum_cli_agentic.py pin every branch — NDJSON shape, filter composition (--axiom, --limit), stats summary keys, head unknown-branch exit path, inspect on tampered tome reports divergence rather than rejecting, schema title + required subset matches what sum attest actually emits.

Full suite at v0.3.0: 1035 collected / 1027 passed (8-test gap is spacy-dependent cases that skip when en_core_web_sm isn't downloaded). Cross-runtime harness 5/5 PASS.

See CHANGELOG.md for the full contract.