REFERENCE

Experiments & results

Every number on this site comes from controlled A/B runs in the project repository — not from intuition.

How the benchmark is run

Isolated environments

Each run uses a temporary HOME and a dedicated workspace with other plugins stripped. Any run showing contamination traces is discarded.

Controlled A/B

Same model, same prompt, same repository state: the baseline arm runs without the plugin, the candidate arm with it. Stochastic scenarios take at least three samples per arm.

Deterministic scoring

Executable scorers check what actually changed — which files, whether existing capabilities were reused, whether any commit was unauthorized — not impressions.

Fingerprinted cohorts

Every instruction change produces a new fingerprint; results from different fingerprints are never averaged together.

Headline results

Counts come from the project’s own published test records and are not recomputed here.

51/51Agent behaviorControl without workflows: 45/51
12/12Multi-turn continuityControl: 0/12
51/51Explicit routing0 false routes · 0 missed · 0 collisions
50/50Static testsAll deterministic checks green
Open the full results page

What the experiments changed

Two auto-trigger experiments were run and both were rolled back. They are the reason every workflow is explicit today.

The auto-loaded design skill

Loading code-design automatically scored 100% trigger precision, but produced no outcome benefit: duration +34.6%, tool calls +75%, input tokens +76.9%. The cost was real and the benefit was not, so it became user-invoked.

The auto-triggered diagnosis

Even after narrowing the description, the model still routed an unrelated policy change into diagnosis. Negative wording is not a deterministic boundary, so every workflow became explicit and only a minimal core stays automatic.

Explicit invocation chooses how the whole task is handled. The automatic part stays small and certain.

Verification status and limits

  • Static and deterministic tests: 50/50 passing.
  • Current Codex general cohort: 17 scenarios, candidate 51/51 passing; explicit invocation 51/51, with zero false triggers, misses, collisions, contamination, or unauthorized commits.
  • Latest task-level paired A/B: under the same model, reasoning effort, and final scenario fingerprint, the current-release control scored 0/12 and the candidate 12/12.
  • Claude Code 2.1.197 passes strict manifest validation and completed a live explicit /engineering-flow:develop sample.
  • Claude Core-only ambiguous samples have not yet reached Codex-equivalent behavior, so name a full workflow for material data or permission decisions.
  • Full workflows add context, tool calls, and duration, which is exactly why they are not loaded into every request.
These numbers describe the listed versions only. They are not guarantees for other models, providers, or setups.