What we can honestly claim so far

Observed 24 September 2026 on the production service's own clock. The loop that builds labbel sends its code reviews through the production service: 189 reviews minted between 27 August and 22 September 2026. Of the 127 that had passed their seven-day window by 14:31 UTC that day, 125 had a change outcome stored at least seven days after the review and 2 were closed as unverifiable; none had reached its thirty-day window. Ninety-nine of those answers were reported that day by a scheduled reporter running in a fresh session with no local journal: it found the due work through the service's due list and read the work through the build loop's own records, and it now runs daily. Behind those 99 windows stand 82 reviewed steps: 18 received a later fix, 62 did not. Read against the recipe, the form of the reviews did not predict which — every review held the same structure — and a blind rubric scoring of the 80 decidable ones hit the ceiling on the primary rubric (72 of 80 scored 5), so it could not separate them either. Where the later fixes came from is the telling part: one came from operation, the rest from later readers — later reviews, security readings and our own sessions — with more context than the review had. These are our own organisation's records, excluded from the population our thesis is measured on. They show the instrument retaining late reports in production and an agent returning to due work; they show no improved method, no voluntary reporting by customers, and no criterion changed on the strength of a delayed signal.

On 14 September 2026, the S-stage test passed through the product's ordinary tools against both an in-memory repository and local PostgreSQL. It created and stamped a domain, authored a recipe, submitted work, moved the test clock, discovered the due follow-up, read a brief for that run, reported against its criteria and inspected the history. A revised domain edition then applied to the next run while the earlier run retained its original contract. Duplicate reports, partial criterion coverage and refusal of a submission against an outdated criteria edition were also checked. This demonstrates the connected behaviour in the tested cases. The outcomes were test data and time was controlled: it does not show a customer returning on a real clock, better criteria or improved results. The S and S+ work was deployed on 16 September 2026 — database migrations 049 to 053 and the MCP server — with limited checks afterwards against the running service. The deployment puts the implementation behind these local tests into production; it does not repeat the full test there or add evidence about outcomes.

Our recorded count was zero as of 10 September 2026. This historical count has not been remeasured for this publication update. The planned first course was a child's mathematics course in a separate organisation, with a later assessment after completion. The internal experiments and local functional test below do not count as completed field courses or as external field outcomes.

Same model, same code change, two instructions. Reviews written under the recipe's contract found a real, SDK-verified hole in our production code in seven runs of eight; the review prompt our build loop was using at the time found it in zero of eight. Attests that the contract changes what the same agent produces. Does not attest generalization: one codebase, one change, one family of models.

Across sixteen runs of packaged review tooling, none supplied a verdict in our contract's format. Those approaches had not been given that contract. This measures conformance to our requested format, not worse judgement or a comparative win.

Handing the same packaged reviewer its tools moved defect-finding from two runs in eight to six. Model capability improves without us — which is exactly why a recipe must not script the how. The recipe's job is the part that did not move on its own.

We believed a review should be composed in steps. We pre-registered the question, the arms, and what every outcome would mean — including the row that said our intuition was wrong — then measured. The three-step variant did no better anywhere that separates the two, and worse on the sharpest measure (consequence stated: 3/8 vs 6/8). The recipe keeps its single step, and the meta-recipe now says one step is legitimate here. Attests that the loop learns, not just that it works.

The 16 reviews in the two recipe arms of experiment D were rated blind: 7 criteria and 5 grading sessions per criterion, giving 560 ratings. Agreement within one point was 1.00. The other arms were assessed on defect-finding and conformance, not through these rubric ratings. Agreement does not establish that the graders were correct.

Every version change is a written entry: what changed and why, dated. Both documents state their own standing plainly — grounded in research and our own measurements, with no field outcome data behind them yet. Education v0.9 becomes the learning profile's document: a course is held to the criteria its own organization stamped, answered criterion by criterion at the return, with the checkpoint and the calibration reported per step. It changes no scoring rubric, and supplies no evidence of improved learning. No criterion has changed because field outcome data demanded it.

What the first analysis of field outcomes will ask, which rows it will count and what every answer would mean was written down and signed on 7 September 2026, before the analysis has data. It separates association from causation in advance, and it excludes the project's own organizations from the population it counts, so the first number cannot be one that measures us. Attests that the question is fixed before the answer; it attests nothing about the answer.

On 8 September 2026 an organization's own agent authored a domain — a kind of work, its criteria, its clock — and a person at the bench read two trial judgments made under it and stamped edition v0.1. One run has been made under that domain since. This is the bench in use, not an experiment: a hypothesis with a name on it, and no evidence yet that the criteria are good — that is what the later outcomes are for.