MMR-Bench v1.0

Three readings of one cell.

MMR-Bench compares the value stored in the workbook, a bounded evaluator, and an independent engine — then signs exactly what it found. A disagreement between engines never authorizes it to rewrite your formula.

cell Budget!D14

formula =SUM(D2:D13)

Kstored cache

1,284.50

the value Excel last wrote into the file

Mmmr evaluator

1,284.50

a bounded AST-whitelisted recomputation

Eexternal engine

1,284.50

an independent Python engine, not Excel

three_way_agreement

authorizes Nothing. The cache already agrees.

never a formula rewrite

Six of the nine classifications, shown with worked values. The 2^3 case is the real v0.9 defect — the old MMR evaluator read Excel's ^ as Python's XOR. The 1904 case is the external engine's 1,462-day date-system gap. Measured fixture outcomes are on the results page.

Where it started

MMR-Bench began as a test of a representational question: should matrix position, direction, display structure, formula dependency, and semantic roles be represented together, rather than flattened into one linear sequence? Nine versions later, the question that survived was narrower and more practical — what can a spreadsheet audit actually prove about itself?

Version 1.0 closes the first research cycle by answering that in a replayable form: a signed certificate that records the evidence, the decisions, and the boundary of both.

What v1.0 adds

A safe bounded formula evaluator built on a Python AST whitelist.
Excel-compatible ^ exponentiation semantics.
Excel-style ROUND, with half-away-from-zero behaviour.
Explicit 1900 / 1904 date-system handling in the MMR evaluator.
Three-source evidence: the OOXML cache, the MMR evaluator, and an independent formulas engine.
Ed25519-signed computation certificates.
Canonical payload hashing, MMR-CJ-1.
XLSX source-hash binding.
Replay verification of normalization, formula evidence, and safety decisions.
Tamper rejection and wrong-source rejection.
Zero formula-write authorization from engine disagreement.

Verified results

Every line below is a count from the v1.0 regression run on 23 July 2026.

Trial	Result
Full historical regression	69 tests passed
Input mutation trials	30 / 30 MMR and external engine agree
Exponent trials	20 / 20 three-way agreement
`ROUND` trials	20 / 20 three-way agreement
1904 date trials	20 / 20 external-engine gap detected
Signed example certificates	4 / 4 signature verified
Signed example certificates	4 / 4 exact replay verified
Tampered payloads	10 / 10 rejected
Wrong XLSX sources	10 / 10 rejected
Formula writes authorized by certificates	0

What the numbers cover

These are results on synthetic business contents stored in real XLSX packages. The packages are genuine OOXML; the contents are generated.

does not establish

Accuracy on arbitrary enterprise workbooks, or complete Microsoft Excel compatibility.

Try it

The benchmark runs from the repository. Certificates are created and checked through one CLI.

bash

python -m pip install -r requirements.txt
./run_v10.sh

Create a signing keypair. The private key stays with you — no example private key ships with the release.

bash

python -m mmr_bench.certificate_cli_v10 keygen \
  --private-key private.pem \
  --public-key public.pem

Sign a workbook audit, verify the certificate, then replay it against the original file.

bash

python -m mmr_bench.certificate_cli_v10 create workbook.xlsx \
  --private-key private.pem \
  --output workbook.mmr-cert.json

python -m mmr_bench.certificate_cli_v10 verify workbook.mmr-cert.json

python -m mmr_bench.certificate_cli_v10 replay \
  workbook.mmr-cert.json workbook.xlsx

Research release. No general production write authorization.