Versions

Nine versions, one narrowing question.

Each version kept less than the one before it. The arc is a record of claims being cut down until what remained could be signed.

The arc

v0.1

Directional traversal prototype. Headline result retracted as a data leak; the generator answered its own question.

v0.2

Model comparison archive, including the PyTorch Transformer baselines still exercised by the regression suite.

v0.4

Safe repair policy: explicit anomaly and confidence thresholds, repairs written to a copy, never the source.

v0.5

Suites, metrics, and falsification criteria fixed in advance across eight generated workbook distributions.

v0.6

Region and route specification.

v0.7

Semantic and lineage specification; review exchange format.

v0.8

Differential validation specification.

v0.9

External calculation oracle protocol: three-source evidence and the nine classifications.

v1.0

Bounded AST evaluator, corrected Excel semantics, MMR-CJ-1 canonical hashing, Ed25519 certificates, and exact replay.

v1.0 changelog

Added safe AST-based bounded evaluator.
Corrected Excel exponentiation semantics.
Added Excel-style ROUND semantics.
Added explicit 1900 / 1904 DATE semantics.
Added the v1.0 oracle protocol.
Added MMR-CJ-1 canonical hashing.
Added Ed25519 certificate creation and verification.
Added exact XLSX source-hash binding.
Added replay verification for normalization, evidence, and decisions.
Added the certificate CLI and key generation.
Added 90 deterministic semantic trials.
Added certificate tamper and wrong-source trials.
Preserved zero formula-write authorization from engine disagreement.
Full historical regression: 69 tests passed.

What ships

New code: value_engine_v10, differential_v10, oracle_protocol_v10, certificate_v10, certificate_cli_v10, evaluate_v10, tests/test_v10.py, and the run scripts.
Certificate examples: Four signed certificates — correct structured, stale cache, operator conflict, and 1904 date — plus the release public key and a certificate index.
Outputs: Fixture metrics; input, exponent, rounding, and date trials; the certificate index; payload-tamper trials; wrong-source replay trials; and a signed audit workbook with plots.
Not included: Any example private key.

The v0.1–v0.9 code, reports, fixtures, and tests remain in the repository for reproducibility.

What v1.0 actually claims

The v1.0 report states its own strongest proposition, and it is deliberately not the exciting one. It is not that MMR now understands Excel. It is that workbook structure, direction and semantics, multi-engine evidence, a source hash, and a replayable signature together form an audit evidence chain that does not rest on any single model's word about itself.

authorized

The external engine is an independent witness, not an authority.

After v1.0

The report closes the first cycle and names what the second one has to be. Synthetic scores stop being the primary measure of progress.

Real anonymized workbook datasets, in place of generated fixtures.
Full formula-semantic coverage rather than the subset the bounded evaluator supports today.
Cross-engine compatibility as engineering work, not as a benchmark number.

Repository status

Research release. No general production write authorization.

authorized

Licensed under the Apache License 2.0 since 25 July 2026. You may use, modify and redistribute it.

Apache-2.0 rather than anything else because the two sibling projects in the same series — MMLC Runtime and MLF — were already Apache-2.0, and one licence across the set is easier to reason about than three.

The licence changes what you may do with the code. It changes nothing on the security page: a valid certificate still does not establish complete Excel semantics, and engine disagreement still never authorizes a formula write.