We measure whether a skill still helps, not which model is best

A Driftproof report asks one question: does a skill still help, and by how much, on the substrate underneath it? When that substrate is two different model families - say a Claude flagship and a GPT flagship - the report shows the skill's with-vs-without lift on each, side by side. The interesting signal is whether a skill's benefit is durable across substrates or whether it depends on the model it was written against.

This is not a model leaderboard. Driftproof does not rank GPT against Claude, and a report's baseline scores are not a "which model is smarter" table. Comparing raw baseline numbers across providers would conflate model capability, judge affinity, and prompt-format differences into one meaningless figure. The unit of evidence is the skill's delta on a fixed judge, not a model's absolute score.

• Same suites. Every provider is run on the identical eval suites at the identical pinned rubrics (Report #001 v1.2 rubrics). No provider gets a suite tuned to its strengths.
• Same judge. One fixed, cheap judge (claude-haiku-4-5) grades every generation from every provider. The judge is held constant so the target model is the only thing that varies. Only a model with judge_eligible: true in the registry may judge - and no OpenAI model is judge-eligible, by policy.
• Same n. Identical sample count per case (n = 5) and the identical band + effect-floor verdict rule across providers. See the methodology.
• No favoritism in authoring. Rubrics grade against a skill's own documented claims (the grounding policy), not against any provider's house style.

The fixed judge is a Claude-family model, and it grades generations from every provider. An LLM judge can be marginally kinder to prose that resembles its own family's - so a Claude-family judge grading a Claude generation may score it slightly higher than an equally-good GPT generation. We do not pretend this is zero.

The mitigation is structural: the headline metric is the with/without-skill delta, not the absolute score. Any constant style preference the judge holds toward a provider is present in both the with-skill and the baseline run for that provider, so taking the difference reduces it, though it is not cancelled. A report leads with each skill's delta on each substrate because the delta is less exposed to judge affinity than the absolute score; absolute cross-provider scores are reported as context, never as a ranking. The judge policy covers style affinity, the grading ceiling on hard suites, quantization, and fail-closed parsing in full.

The same model can be reached over different surfaces: a metered API or a subscription CLI. Driftproof's provider layer is two-axis - provider (anthropic / openai) × surface (api / cli) - and every receipt records the exact provider and surface it ran on.

• The API surface is preferred for published runs (temperature is settable, token usage is reported, and it is reproducible for a third party who has a key).
• A CLI / subscription surface is disclosed whenever it is used - Report #001 ran on claude-cli and said so, and that precedent holds. On the openai-cli (Codex) surface the receipt additionally carries a surface_overhead_note stating the fixed harness preamble the surface prepends.
• A drift comparison between receipts on different surfaces is flagged, because sampling control differs between them.

Report #002 goes further than disclosure: both substrate columns run on the vendor's own first-party subscription CLI - Claude via claude -p (surface claude-cli), GPT via codex exec (surface openai-cli). The type of surface (a first-party vendor CLI on a subscription) is therefore held constant across substrates, not merely recorded - so a difference between the columns is not an api-vs-cli artifact.

Per-surface harness overhead is disclosed, because it differs. The openai-cli (Codex) surface prepends a fixed ~12,000–15,000-token base-instruction preamble to every call (recorded in each receipt's surface_overhead_note); the claude-cli (claude -p) surface carries a smaller first-party harness/system context. Neither preamble is authored or controlled by Driftproof, both are first-party to the vendor, and each is noted so a reader knows the per-call input-token counts are dominated by a vendor-side constant, not by the skill.

Absolute cross-substrate scores are not a ranking. Because the two substrates sit behind different first-party harnesses (with different overhead and different sampling control), comparing one column's raw number to the other's would conflate model, harness, and judge affinity. The headline metric is the within-substrate with/without-skill delta - computed against the same baseline on the same surface - which cancels the vendor-side constant and isolates the skill's contribution. Read the delta down each column; do not read across the columns as a scoreboard.