Evidence and papers
The academic side of the method: the papers, the experiments and the evidence behind the claims. It is here for readers who want to check the work. If you want to learn the method, start with Learn.
Read the bounds as part of the findings. Every experiment is proponent-authored, small, and mostly run on one benchmark. That is a limitation, and the reason each one is published as an artifact you can re-run. The evidence page states each finding with its bound.
The papers
The paper is being split into one base paper and focused derivatives, because one document was carrying too much. Each paper carries one claim, and a claim lives in exactly one paper. The paper tree says what each paper will and will not say.
| Paper | Focus | Status |
|---|---|---|
| Base: Derivable Correctness | The stateless-reader discipline, the substrate, bridge / sentinel / phase collapse, the seven-property rubric, honest capacity-relative bounds | Archived preprint on Zenodo (July 2026) · revised base in preparation |
| D1: The Externalized Guarantee | What endures under strong models: the check that lives outside the model | Planned · gated on a pre-registered experiment |
| D2: Cheap Rigor | Why cheap execution revives validated-but-abandoned disciplines, and when | Planned · pre-registered design |
| D3: The Pragmatic Tier | A semiotic placement for AI-era programming discipline (position paper) | Planned |
| Compendium | The coherent superset every paper is cut from. Not a submission |
About the Zenodo record. It is an archived preprint from July 2026, kept as a citable snapshot. It is not the revised base paper. The white paper working text on this site is a long-form draft and is not a published release.
Companions: Field guide · Practitioner protocol · Rubric scoring guide · Experiment supplement · Experiment ledger
The experiments
Every experiment is committed with its raw evidence, so a reader can verify rather than trust. None was preregistered with an external timestamp: the papers say, per experiment, what was committed before the data and call every study exploratory. The experiment ledger is the index and the experiments page lists each one.
| Experiment | What it tests | Status |
|---|---|---|
| AX | Specification completeness vs. structural quality, eight conditions on the RealWorld Conduit benchmark | Complete |
| BX | Author-independence of the rubric: three implementations scored blind | Complete |
| CX | Patchability: a GS-specified and a non-GS codebase both resolve 5/5 patch tasks. No pass-rate advantage; the difference is where the patches land (the architecture directed each one to the right layer) | Complete |
| RX | Reproducibility: anyone with Docker and an API key can reproduce 104 passing tests | Complete |
| KX · SX · TX · EX · ALX | Sentinel and retrieval economics · navigation vs. bounding · the bridge · production · the formal tier | Complete or reported in the Compendium |
| MX · RND-1 · NX · CR | Model-agnosticism · prescriptive specs · N-version revival · capacity-relative datum | Pilots |
| Revival grid | The capacity-gated revival model, pre-registered before any generation | Design frozen, running next |
Replicate us. The experiments are proponent-authored. That is a limitation, and the reason each one is published as a pre-registered artifact you can re-run. See how to replicate an experiment. Independent replications, including null results, are welcome by pull request.
Reproduce RX
Requirements: Docker, Node.js 20+, an Anthropic API key.
git clone https://github.com/jghiringhelli/generative-specification
cd generative-specification/experiments/rx
docker compose up -d postgres
./runner/run.sh
cat evidence/jest-output.json # verify: numFailedTests === 0
Verify AX from committed evidence
The pre-run evidence (scores, evaluation transcripts, session logs) is in experiments/ax/. Pre-registration commit timestamps show the rubric was locked before any experimental run.
Production evidence repositories
| Repository | Description | Role in the papers |
|---|---|---|
| brad-gs-build | BRAD legal intelligence engine. 37 commits, Feb 27–Mar 4. Built from scratch under the method. | Case study |
| scp-gs-experiment | SafetyCore Pro quality gate pass. 4 experiment commits, Mar 16–18. | Adversarial experiment |
Commit timestamps are signed by GitHub.
Models and equations
The models and equations have their own section, with the status each has earned: Models and equations.