Experiments

The controlled studies below establish the validity and generalizability of the GS methodology; two June-2026 pilots (MX, RND-1) extend it to model-cost economics and to behaviour under delivery pressure.

Experiment What It Tests Status
AX — Adversarial Quality as a function of specification completeness. Eight conditions, naive through ForgeCraft treatment v7. RealWorld Conduit benchmark. ✅ Complete
BX — Benchmark Rubric validity. Three Conduit implementations scored blind against the GS rubric — two never exposed to GS. Establishes the rubric captures real quality. ✅ Complete
CX — Patchability GS-specified codebases are more patchable. SWE-bench-style patch tasks on two quality tiers characterized by BX. ✅ Complete
RX — Replication Any reader can reproduce 104 passing tests against a live PostgreSQL instance from a GS document alone. No ForgeCraft required. ✅ Complete
EX — Executable Sprint Full L1–L4 tier proof on the live RealWorld Conduit benchmark. 13/13 behavioral probes, 3/3 env probes, k6 ramp — all green on Railway production. Single session. ✅ Complete
KX — Knowledge Retrieval Routed navigation-tree retrieval beats RAG-dump and no-structure on accuracy and token cost; the CKG divergence replicates on software. ✅ Complete
MX — Model Cost & Tiering Once GS-specified, a mid-tier model (Sonnet) matches a strong model (Opus) at ~6× lower cost on the full Conduit; model-tiering is unjustified when the mid model one-shots the task. ✅ Pilot (Jun 2026)
RND-1 — Spec / Verify / Context under Pressure Which GS arm suppresses under-pressure failure modes (literal-minimum under-spec; test-faking). Prescriptive spec confirmed; verification & bounded-context arms return honest, bounding nulls. ✅ Pilot (Jun 2026)

Validation Structure

The experiments address a three-layer validity problem:

Layer Threat Closed By
Output measurement External checks use criteria the author defined BX: rubric applied to non-GS implementations
Rubric validity Rubric rewards GS compliance, not objective quality BX + CX: congruent with CVE count, test count, patchability
Guidance circularity GS guided the implementation AND scored it BX: blind evaluator on peer implementations; RX: independent replication; plus observational field corroboration

Layers 1 and 2 are closed. Layer 3 is addressed by BX/RX and observational field corroboration; a controlled human-participant study is future work.

EX addresses a complementary validity question: does the methodology produce working, deployable software at all four tiers? AX proves quality increases with specification completeness — EX proves the toolchain closes the loop from specification to production-verified behavior.


Pre-Registration Policy

AX rubrics, hypotheses, and evaluation criteria were committed to this repository before any experimental run. Commit timestamps are cryptographically signed by GitHub. This prevents post-hoc rubric adjustment.


Reproduce RX Yourself

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   # numFailedTests === 0

Table of contents