2026-W34 is now published, as a partial week

What happened

The run labelled 2026-W34 started on schedule on 2026-08-24. It was the first week in which Claude Opus 5 ran alongside Claude Opus 4.8, which roughly doubled the time the run needed. The system that dispatches the run applied a default execution limit of one hour that we had never changed, and stopped the run when the hour was up. Claude Opus 4.8 and Llama 3.2 3B had finished. Claude Opus 5 had not.

A run stopped that way writes nothing at the end: no manifest, no snapshot of responses, no entry in our run log. The responses it had already captured were safe on the sampling machine. They were copied to our archive on 2026-10-05.

What is published

Model	Samples	Prompts complete	Status
Claude Opus 4.8	600 of 600	30 of 30	complete
Llama 3.2 3B	750 of 750	30 of 30	complete
Claude Opus 5	259 of 600	12 of 30, one more at 19 of 20	partial

The 17 prompts Claude Opus 5 never reached have no measurement for this week and no row in the data. The prompt it was part way through, sci-iq-heritability, is published at the 19 samples it has.

How it is marked

• The week’s manifest, data/manifests/2026-W34.json, carries partial: true, a per-model coverage list naming every missing and cut-short prompt, and notes explaining when and how it was built.
• Our run log has one entry for the week, marked recovery: true. It was written on 2026-10-07 from the archived responses, not by the run, and its note says so. Its pair counts are what the archive holds: 72 of 90 pairs complete and 18 not.
• The coverage page shows the week as partial for Claude Opus 5, and the methodology records the gap.

Why publish it at all

Two reasons. First, the responses are real measurements, captured on the Monday the run was scheduled for, exactly like every other week, and nothing has been added to them since. Second, this week was already part of the record. Its metrics had been carried in the history of every later manifest and shown under /data/2026-W34/ since 2026-W35, and two later weeks measured change against it: Llama 3.2 3B in 2026-W35, and three Claude cells in 2026-W36. Our methodology page said the week would not be published, which was not consistent with what the site was already doing. Publishing the week with its coverage stated makes the record match what it was using.

What we did not do

We did not re-sample anything. Asking Claude Opus 5 today would measure the model as it is today and label it as August, which is precisely the substitution this project exists to detect. The missing prompts stay missing.

One small difference to expect

The manifest for 2026-W34 was built from the archived responses with the same code a live run uses, and with a fixed random seed so that anyone can rebuild it exactly. Confidence intervals and test statistics are drawn by resampling, and the copies of 2026-W34 already embedded in later weeks’ history were drawn without a fixed seed. Some intervals and p-values may therefore differ in the last decimal places between those copies and this manifest. Every count, rate and length is the same computation on the same responses. Stance had never been classified for this week, because the run was stopped before that step; it was classified on 2026-10-07, when the manifest was built, with the same pinned classifier model every week uses.

The cause is fixed

The execution limit is now six hours, and a scheduled check stops the sampling machine if a run ever ends without stopping it, which this one also failed to do.

As with everything we publish, no provider saw this before publication.

@misc{meridian20261007w34partialweek,
  title        = {{2026-W34 is now published, as a partial week}},
  author       = {Meridian},
  year         = {2026},
  howpublished = {\url{https://meridianaudit.org/reports/2026-10-07-w34-partial-week/}},
  note         = {Accessed: YYYY-MM-DD}
}