Cost

The cost of proof

trace pays for three things: model API calls, compute on GitHub Actions and GPU training on Hugging Face. Each has its own section below. The deterministic roles, the resolver, Builder and Steward, make no model call and cost nothing in API fees. They still run in CI, and CI time costs money: trace was not developed in the open. It started as a private repository, and GitHub bills Actions minutes against a monthly allowance on a private repository. A public repository gets unlimited standard Actions minutes for free, so every dollar in the GitHub Actions section below was spent specifically because the repository was private while that work ran.

Model API spend

Every live model call made by the pipeline's own roles writes an invocation sidecar with its role, model id, prompt hash, token counts and billed cost. The figures below come straight from those sidecars, current as of the 2026-09-09 throughput report. They do not include live spend from the evaluation scripts that measure the pipeline from outside it, including the 39% accuracy run described on the findings page, which wrote no cost field at all; that gap is named directly in the report this page cites.

Three invocations carry no recorded cost and are left out of every figure above.

Cost per bundle over time

Cost per merge-ready bundle has been measured seven times since May 2026, each time on a larger sample. It climbed from $0.224 on three bundles to about $0.37 by mid-June on twenty-four, then held there through September on a sample that kept growing. Because the figure held steady as the sample kept growing, it is more trustworthy than any one of the seven measurements by itself.

Human review costs more than auto-merge

The merge-ready count had not grown since June. Total spend on bundles rose $12.88 between 21 July and 9 September, and all of it went to human-review-required bundles, whose mean cost ($0.4419) already exceeded the auto-merge mean. Recent spend bought evidence and contract findings on cases the policy routes to a person, not new merge-ready bundles. That is deliberate: the policy routes a case to a person specifically when it isn't safe to decide by cost alone.

Where the money went

The Patcher takes 81 percent of model spend on 43 percent of the calls. Opus bills output tokens at five times its input rate, and a diff is mostly output. The Reviewer costs an order of magnitude less because it returns a verdict and a handful of citations. That gap is the case for every free gate the pipeline has added. A preflight check, a rationale-binding check and a build probe each cost nothing, and each can stop a run before a paid role reads a patch that was never going to survive.

GitHub Actions compute

Every cert, scheduled harvest and test run executes on GitHub Actions. Across 2,985 workflow runs between 28 April and 9 September 2026, trace used 265 wall-clock hours and 21,076 billable minutes. Billable minutes run higher than wall-clock time because parallel jobs bill separately and GitHub rounds each job up to the whole minute. At list price that comes to approximately $168.61 gross.

The proof itself is paid for here. The cert workflow builds each package twice, once without the fix and once with it, and an uncached nixpkgs build can run for hours. A model call finishes in minutes. bundle-linux-cert and the plain test suite together take up most of the bill.

June alone is 42 percent of the total minutes. That was a cert campaign: 155 of the project's 241 cert runs ran that month, and each cert-driven commit also triggered the full test suite.

Per-job durations were read directly through the Actions API for 1,816 of the 2,985 runs before a rate limit stopped the fetch early; the remaining runs are extrapolated from those. The read runs carry 76 percent of the total minutes, so most of the $168.61 figure is a direct read, not an estimate. We also checked GitHub's own org-level billing summary against this figure before publishing it. That summary attributes most months to the wrong repository and shows trace at close to zero in June, which the Actions API itself contradicts: trace ran 852 workflow runs that month alone. We trust the per-run computation over that summary endpoint.

GPU training on Hugging Face

The fine-tuning experiments ran as GPU jobs on Hugging Face. They trained the open-weight Reviewer and Patcher candidates described on the findings page. Across 105 costed jobs (of 138 attempted) and 72.74 GPU hours, training cost approximately $173.78.

Almost all of it runs on one GPU type. An 80GB A100 at $2.50 an hour carries 86 runs and 61.22 hours, or $153.04. A cheaper L40S ran out of memory on long source diffs at a 16K sequence length, which forced the A100. July is the busiest month by run count. August, with only four runs, has the longest average run at nearly two hours, against roughly one hour per run in June; that traces (no pun intended) an arc from a few long training runs to many short scored inference passes over the adapters those runs produced.

That money bought a negative result. The fine-tuned models did not come close to the frontier models they were meant to replace, and the frontier models still do the work. The training is still part of what this project cost, and it is counted here with everything else.

What those jobs actually produced, on Hugging Face:

• reviewer-lora-v1.10: the fine-tuned Reviewer candidate itself. Given a patch and the upstream fix it claims to match, it returns accept or reject with a grounded rationale. Public.
• reviewer-lora-data: the instruction-tuning data that adapter was trained on. Public.
• authoring-lora-lossmask: the equivalent fine-tuned candidate for the Patcher role, trained with a loss-mask technique over the model's own authored diffs. Private as of 20 October 2026.
• authoring-lora-data-v13-diversefold: the training corpus behind it, kept for reproducibility now that this research line is closed. Private as of 20 October 2026.

Total recorded cost

Model calls, CI compute and GPU training together come to roughly $416 gross in recorded spend, or $307 to $328 after a typical monthly Actions allowance; this excludes the eval-script model spend and the 33 GPU jobs with no cost record, noted above. Model calls are about one-sixth of the $416. Anyone who quotes only the model bill as the cost of this project understates even the recorded figure by more than five times.

Model spend has no official monthly breakdown the way CI and GPU training do, so this compares the three category totals rather than a timeline. Full methodology, including what this report does not count, is in COST_REPORT_v1.1.0.md.