MOMM 1.14.0

MOMM 1.14.0 — the critique release

Released 2026-09-04, hours after 1.13.0, in response to an external review of the public page that was mostly right. What it said, and what changed.

The page oversold the workflow

The 1.13.0 page described a local multi-CLI review script with a reproduction gate as a coalition, a journal and a measured scientific instrument. The page is rewritten to say what the tool is. Specifically:

• "Precision" is the governor's acceptance rate. applied ÷ (applied + rejected) measures agreement with one agent's later triage on this project, not precision against a labeled ground truth. --stats, the report's insights.track_record_note, the ledger panel and the page now say so. It remains an attention prior only.
• The ledger is self-review, not an external evaluation. Every run in the public export is this project reviewing itself. The page states that in the first sentence of the numbers section.
• Route failures are stated, not admired. The flagship run had two of five routes fail. The page says so up front, and the terminal UI now ends with the routes that did not review.
• The privacy question is answered in one sentence. Yes: each live route sends the redacted, metadata-stripped input to that vendor over your own login. "Zero API keys" is an implementation detail, not a privacy property. "Nothing leaves the machine" was only ever true of the voice generation and is now said only there.
• The install is explicit. install.mjs refuses to run without --target; the headline command links one harness and previews with --dry-run. --target all still exists, but you have to type it.
• The numbers are copyable. Every chart has a static HTML table beside it and a CSV under docs/momm/data/, regenerated by the exporter with the evidence and its digest.
• The peer-review analogy is demoted to what it is: a workflow shape that happens to match, on a synthetic specimen whose planted defects a single frontier model would also catch.
• The creator's academic title is removed from the page pending a verifiable source; the profile links remain.

What changed in the dispatcher

• Prose corroboration. When the input is not a diff, location keys are ignored — asked for a line_range on prose, reviewers invent one (prompt.txt:49, §2.3 Analysis:11, 2. Methods:3 for the same sentence in the live re-run) — and a finding merges with another reviewer's finding that shares a run of six normalized words: two readers quoting the same sentence are citing the same defect. The same reviewer's findings never merge this way; diff inputs keep the line-range rules. Replayed on the two sealed specimen runs: 26 raw → 10 defects, 9 corroborated (0.90) and 14 raw → 7 defects, 5 corroborated (0.71), where 1.13.0 scored 0 on both. Five self-tests.
• --tier quick|deep. Quick runs the two fastest routes by ledger median (copilot, antigravity) with a 60-second budget for staged commits; deep runs the pool with quorum 2 for release gates. Explicit --reviewers, --timeout and --min-success always win. Recorded as tier in the report.
• invalid_output is diagnosable. The detail now records the failure class (empty stdout, prose instead of JSON, JSON without findings[]), byte counts, whether the output limit was hit, and a sanitized 200-character sample.
• Redaction covers more token shapes (gho_, ghu_, ghs_, github_pat_, AWS AKIA…, bare sk-…), with a self-test proving template strings and process.env lookups survive untouched.
• Terminal output ends with the useful part: material findings with file and line anchors, the reviewer's own reproduction idea, and the routes that failed.

What was not done, and why

• Automatic execution of reviewer-supplied reproduction snippets was declined. Reviewer output is untrusted by protocol; running it, even sandboxed, is exactly the instruction-following the protocol forbids. test_suggestion stays a suggestion the governor rewrites and runs itself.
• Renaming the project is a decision for its owner. The expansion "Mixture of Model Modality" is dropped from the page; the name MOMM stays.
• POSIX process-group termination is on the roadmap: it needs a POSIX machine to verify and the Actions matrix to drill it.

Verification

Dispatcher self-test 61/61 (eight new), ledger 15/15, Setup Center 20/20, myskills health contract, myrepo offline self-test, installer refuses a missing --target; specimen re-run live with the new merge; the Windows/macOS/Linux × Node 18/20/22 Actions matrix.

← All version notes