Why I built this
I kept running into the same problem: AI can make a lot of connections, but it is much worse at telling me which connections are actually useful and which ones merely sound smart.
Braid Engine is my attempt to turn cross-domain reasoning into a disciplined research workflow. It can connect different bodies of knowledge, but every generated idea has to carry its evidence, expose its assumptions, and survive an “anti-braid” that actively searches for counterexamples and overreach.
What it does
Every corpus is translated into the same four inspectable objects:
- Evidence units with provenance
- Candidate hypotheses
- Anti-braids that attack those hypotheses
- Verifier verdicts: source-checked, model-generated, verified, falsified, or unresolved
The Build Week benchmark attacks the still-open Erdős Problem #96. The control branch uses mathematics only. The treatment branch compiles the problem into structural language, retrieves reasoning operators from a cited cross-domain corpus, reconstructs them as mathematical strategies, and then subjects them to exact checks.
The live demo also includes a rebuilt Bible Galaxy based on a pre-Build-Week corpus. It is deliberately labeled prior work. Its purpose is to show that the same evidence-to-candidate workflow can travel between very different domains.
What happened
Braid Engine did not solve Erdős #96, and the product says so prominently.
It did produce one exact reduction: in the two-core, all excess edges can be concentrated into a special double-middle subgraph D, giving m ≤ 2n + |D|. Proving |D| is linear would prove the desired linear bound.
Then the anti-braid did something I consider just as important: it killed two seductive shortcuts.
- “D is a matching” is false; a 9-point rational certificate disproves it.
- “D is a forest” is false; an 18-point exact certificate contains a supported 6-cycle.
The surviving frontier is narrower, explicit, and falsifiable. That is the point of the product: useful research progress without laundering model confidence into truth.
How I used GPT-5.6 and Codex
GPT-5.6 helped compile the formal problem into structural language, retrieve distant reasoning operators, reconstruct candidate mathematical moves, generate adversarial objections, and translate verifier outcomes into the visible research replay.
Codex was used throughout: auditing my prior systems, designing the universal contracts, implementing the Python engine and adapters, building exact certificates and fail-closed tests, running the control/treatment research, and shipping the two interactive galaxies.
I made the load-bearing product decisions: make the method modular, use Erdős #96 as the benchmark, use reverse-braiding as the retrieval architecture, and never present an unresolved theorem as solved.
What is new versus prior work
Bible Braid, Director Brain, and Empire Dashboard existed before Build Week. The universal evidence/hypothesis/anti-braid/verifier contracts, the Erdős experiment, exact counterexample workflow, tests, and unified judge experience were built during Build Week.
“Candidate new insight” means new relative to the loaded reference set and worth testing. Braid Engine cannot certify that no human has ever published the idea.
Try it
Open the live demo, run the 90-second green path, compare Math-only vs Braid, then switch to Bible Galaxy to see the same workflow on a second corpus.
Built With
- cloudflare-workers
- gpt-5.6
- next.js
- openai-codex
- python
- typescript
- vinext