Franken Research · learn, test, apply

FR · Research you can put to work.

Try a runnable research technique, inspect its results and limits, and carry it into your own AI project. Start with an evidence cycle that catches changed inputs and transfers the decision for review.

Made by Joshua Nowak (ZestStream), independent of the maintainer.

Daily collection status: what changed, open issues.

Judge this system by successful applications and reproduced results. An upstream repository count or build status is not a measure of what you can achieve.

Source assessment: September 22, 2026 — scope and risks

Jeffrey Emanuel’s FrankenSuite is 44 ambitious repos written mostly by AI coding agents. We froze each one at a single commit, read the code, and graded every claim: what runs, what is only promised, and what nobody can verify yet. The 3 worth a bounded trial on real work sit on the amber ring near the center.

2 of 44 show CI (continuous integration) passing at the commit we pinned (the single assessed commit). Of the other 42, 11 fail there and the rest have CI disabled, missing, private, or silent.

The honesty machinery is real. Negative-evidence ledgers, claim registries, fail-closed gates: unusually well built, and mostly not running where it counts. What is worth stealing

38 of 44 carry an AI-lab license rider: a clause that bars OpenAI and Anthropic from any use, even testing and benchmarking.

All 44 have a bus factor of 1 (one maintainer, no backup): the number of people who can leave before a project stalls. Ours is 1 too.

These counts describe public CI evidence at the assessed commits. They are historical source-selection context, not a product reliability score or an adoption recommendation. See all CI categories.

More pages

Beyond the 44: adopt-or-build verdicts for 21 parts of the agent stack. Source on GitHub

Search, filters and key

Rings are our verdict. Pilot: try it on bounded real work. Explore: real code, unproven claims. Monitor: watch, don’t land. Invest: none reached it.

Center: the one grading method, applied to all 44. Cages: repos without the AI-lab rider, white = plain MIT, red = no license file (all rights reserved).

drag to orbit  ·  scroll to zoom  ·  click a name or node for the bottom line  ·  double‑click opens its briefdrag sideways to orbit  ·  pinch to zoom  ·  tap a name or node for the bottom line

Search covers all 44 repos, regardless of filters.

Wireframe cages mark the 6 repos without the rider: white = plain MIT (1), red = no license file, all rights reserved (5). Domes, when shown, mark the rider.

How to read it

Invest (0): independent validation plus governance. Empty in this program.
Pilot (3): release artifact + a bounded, real workload fit. Not a recommendation to ship.
Explore (34): real code, unproven claims
Monitor (7): watch, don’t land
green glow = CI green at the pin: CI (continuous integration) builds passing on the assessed commit. The pin is the single commit the assessor fixed for the whole evaluation. frankenscipy and franken_threed, the only 2
The rider = MIT + OpenAI/Anthropic rider, barring the AI labs from training, use, or benchmarking. 38 of 44 carry it: the domed ones when you toggle the ceiling.
white cage = plain MIT, the suite’s only plain-MIT repo (1 of 6 without the rider)
red cage = no license file: all rights reserved, the most restricted state (5)
bus factor = how many people can leave before the project stalls. It is 1 across the suite: one maintainer, checked by no one.
Bigger node = higher TRL (2–8); height above the plane climbs with TRL too. A thin vertical band marks a TRL range, plotted as a span, never a midpoint. frankentui_website is a support asset, so it sits flat on the plane: N/A as technology. The 38 rider-bound repos show a dome with “Show the rider ceiling”, every dome the same size, by design.
No license file is not freer than plain MIT. It means no operative grant at all: more restricted than the rider, which still grants rights to everyone outside the AI labs.
Map exceptions, not the rule: frankentui_website is ~36k lines of TypeScript, the two beads repos are static dashboard exports, and franken_agent_detection is not a rewrite of an incumbent.

A judgment map, not a measurement. Every verdict is an analyst’s inference from a packet frozen at one commit (Sep 2026), confidence: medium. Size is readiness, not quality: a big Monitor node is still a website.

The rider ceiling. Dome = MIT + OpenAI/Anthropic rider. The rider withholds rights only from the AI labs and their affiliates.