Graded by our own method

Franken Research assessed 44 repositories from Jeffrey Emanuel's FrankenSuite and reported, among other things, how few of them could show a passing automated build, a release, a license without restrictions, or a second maintainer. A project that publishes those findings should pass the same checks, in public, or say where it does not. This page scores this repository on the same columns as the master matrix in synthesis/00-overview.md, with the same evidence tiers.

Who made thisJoshua Nowak, with AI coding agents

Made by Joshua Nowak (ZestStream) with AI coding agents. Not affiliated with or funded by Jeffrey Emanuel; Joshua uses his public tools and is a paying subscriber to jeffreys-skills.md.

How independence is handled: the agent session that writes a packet or a re-check never reviews it. A separate session does, and a human decides what ships. That separation is internal. No one outside the project has reviewed the work yet, which is why the table below rates independent validation None. If you find an error, the correction form opens a GitHub issue that asks for the claim and your evidence.

Six definitions. TRL is technology readiness level, scored 1 to 9. NODUS is the program's four-ring verdict scale (Invest, Pilot, Explore, Monitor). CI is continuous integration: automated checks that run on every push. The pin is the exact commit an assessment froze on; every claim is evaluated at the pin, not at a later HEAD. The rider is the license clause withholding all rights, including benchmarking and testing, from OpenAI, Anthropic, their affiliates, and anyone acting for them; most of the assessed repos carry it. Bus factor is how many people would have to leave before a project stalls.

The matrix rowFranken Research, scored like the other 44

The middle column is the suite-wide result from the synthesis. The right column is this repository. Tier labels follow RULEBOOK.md §1; the self-assessor is also the maintainer, so nothing below counts as independent.

ColumnSuite resultFranken Research
TRL Range 2–8 across the assessed software (one website scored as an artifact only). 6 while this page was written. The site is built, runs offline from a download, and passes eleven automated gates, including a headless-browser render. The live FrankenSuite website repos scored higher because they were served and observed in production; this site earns that only once it is live at fr.zeststream.ai and someone other than us has used it. [Inference, Medium]
NODUS ring Pilot 3, Explore 34, Monitor 7, Invest 0. Monitor. The ring rules put websites and plan-stage work in Monitor, and this is a website plus a research corpus. The v1.0.0 release supplies Pilot's release artifact, but Pilot also needs a bounded, real workload fit, and a publication has none; Invest would need independent validation, which does not exist at all. The rubric was written to rate software someone might build on, and a research publication is read, checked, and corrected rather than adopted, so the ring mostly says: do not treat this as a dependency. When in doubt, ring down. [Inference, Medium]
License Rider on 38; plain MIT on 1; no operative license on 5. MIT, no rider. One LICENSE file at the repository root covers the site, the packets, the synthesis, and the starter kit. Anyone may benchmark, test, index, or train on it, including the AI labs the rider excludes. [Verified, High]
Bus factor 1 in all 44. 1. One person maintains this repository. That is the same finding we reported for every assessed repo, and it is just as true here. If you want to co-maintain, open an issue. [Verified, High]
Contribution policy Explicit refusal of outside contributions in 19 repos. Open. Corrections are the contribution we most want: cite the packet file and line, the evidence, and its tier. The process is in CONTRIBUTING.md. Maintainers of assessed projects have a right of reply. [Verified, High]
CI at the pin Public CI green at the pin for 2; the other 42 cannot show it: 11 are red at the pin, and the rest have no verdict for the pin, no test CI, CI disabled or deleted, or CI that runs only on private infrastructure. See the badge; we do not assert its color. The verify workflow runs the full gate chain on every push and pull request, checks that the committed map bundle equals what its source builds, and scans for leaked secrets. verify workflow status on GitHub Actions Viewed offline, the live badge does not load; the run history is at github.com/JYeswak/franken-research/actions. A page cannot contain the hash of the commit it is part of, so the pin for this self-assessment is the commit the v1.0.0 tag points to once it exists, and until then the commit the badge reports on. [CI-observed, when the badge loads]
Release posture No release or tag in 23 repos; a release targeting an earlier commit, or a phantom one, in 6; some release artifact in 15. Tagged releases. v1.0.0 (the site) and v1.1.0 (the agent-stack layer), and v1.2.0 (the daily watch, issue forms, and automatic deploys); each release page names its commit and carries a source archive. Before the tag, this repository sat in the same bucket as the 23: no release or tag. [Verified, High]
Independent validation None found for any of the 44. None. No third party has reviewed the packets, re-derived a verdict, or checked the synthesis counts. Our own gates and our own review are not independence, by our own rule. If you check a verdict, publish what you found and tell us; we will link it here whether it agrees or not. [Verified, High]
Behavioral reproduction Analyst re-ran tests for 1 repo (franken_threed, 61 of 61 green). Anyone can run ours. The gate chain below is the whole of it: the same command CI runs. What it proves and does not prove is spelled out after the commands. [Verified, High]

ReproduceRun the gate chain yourself

You need git, Bun, python3, node, and Google Chrome or Chromium. The render gate finds Chrome in the usual install locations; set CHROME_PATH if yours lives elsewhere. No network access is used after the install.

git clone https://github.com/JYeswak/franken-research.git
cd franken-research
bun install --frozen-lockfile
bun run verify
# optional: prove the committed 3D map bundle is what its source builds
bun run build:map
git diff --exit-code site/assets/app.bundle.js

The chain prints PASS or FAIL for eleven gates and exits nonzero on any failure. They check that the shipped packets and Rulebook are byte-identical to the canonical copies; that every method-page citation resolves to a real claim-table line; that every suite number on the site is computed from the data rather than typed in; that the six terms above are defined before first use on every page; that every brief pairs with exactly one packet; that every internal link and anchor resolves offline; that every referenced file ships; that pages stay keyboard-operable with no-JS fallbacks; that copy passes a slop scan; that five representative pages, this one included, render in headless Chrome at desktop and phone widths with zero console errors and zero horizontal overflow; and that no removed scaffolding is referenced. The full list, with the reason for each gate, is in BUILD-GATES.md; the script is scripts/verify-site.sh.

What the gates do not proveThey prove the site says what the packets say, consistently and reachably. They do not prove the packets are right. A verdict is only as good as the reading behind it; to test one, follow Reproduce a verdict against a real repository at its pin.

Known gapsWhat we got wrong, or cannot prove

  • Our own mechanism ran ahead of execution, for a while. The second headline finding says the suite's checks often exist without running where it matters. When this page was first written, the verify workflow existed but had never run on GitHub, because the repository had not yet been pushed. Its first run, on commit 527c227, passed every step: the secret scan of the tree and its history, the check that the map bundle equals what its source builds, and all eleven gates (run 35944707134). Commit hashes in this repository changed in a history rewrite on 2026-09-23 that removed email addresses; CI runs from before it show the old hashes, and the old-to-new map links them. One green run is a start, not a track record. [CI-observed, High]
  • Two gate summary lines printed a blank count. The statistics gate and the link gate reported "( data-stat slots checked)" and "( links)" on macOS, because the count was extracted with a GNU-only grep flag. Both gates were still checking; only the summary lied by omission. Fixed in this release, and an empty scan set now fails instead of passing.
  • Readiness and ring are judgment calls. Every TRL and NODUS placement in the corpus is labeled [Inference, Medium], ours included. Reasonable readers will place some repos a ring differently.
  • Only one verdict rests on re-executed tests. For the other repos the evidence is reading the code at the pin, CI pages, and release APIs. That is stated in each packet's method section, but it bounds how strong any single verdict can be.
  • Two CI taxonomies disagree by one repo. The site's data files frankenjax as private-only CI, while the synthesis matrix files it as having no test CI. The reclassification is disclosed on the method page; it moves no headline number.
  • The corpus is a snapshot. Everything was assessed on 2026-09-22 at each repo's pin. The FrankenSuite moves fast, and some findings will be stale by the time you read them. Staleness is a finding, not an error, but check the pin date before quoting a verdict.
  • We changed the as-shipped packets once. Maintainer commit email addresses quoted from public git metadata were redacted from the packets, their site copies, and one synthesis brief before publication. No other evidence text was edited.

Help fix it

Each opens a GitHub issue form with this page’s address filled in. You need a GitHub account.

Made by Joshua Nowak (ZestStream) with AI coding agents. Not affiliated with or funded by Jeffrey Emanuel; Joshua uses his public tools and is a paying subscriber to jeffreys-skills.md.

Source on GitHub MIT licensed