Digest · week 2026-W30

The week, counted.

This week in one line

80 canary checks run (52 matched baseline) · 4000 battery trials across 16 seats · 800 stances held · 0 changes confirmed · 1 seat publishing as gap (openweight-a). Pre-series · generated 2026-07-21

Canary checks
52of 80 match
Battery trials
4000
Stances held
800
Changes confirmed
0

Actions this occasion

one state per seat, silence quantified
2026-07-21 6 of 16 AFFIRMED; 9 confirmed changes, 1 new baseline.
CONFIRMED anthropic-a
CUSUM 3.00 of 2.0 alarm
1 canary diverged
CONFIRMED openai-a
CUSUM 0.00 of 2.0 alarm
CONFIRMED google-a
CUSUM 0.00 of 2.0 alarm
5 canary diverged
AFFIRMED xai-a
CUSUM 0.00 of 2.0 alarm
detects ≥ 2.91
CONFIRMED deepseek-a
CUSUM 3.00 of 2.0 alarm
1 canary diverged
AFFIRMED anthropic-b
CUSUM 0.00 of 2.0 alarm
detects ≥ 1.94
AFFIRMED anthropic-c
CUSUM 0.00 of 2.0 alarm
detects ≥ 1.58
CONFIRMED anthropic-d
CUSUM 0.50 of 2.0 alarm
1 canary diverged
AFFIRMED openai-b
CUSUM 0.00 of 2.0 alarm
detects ≥ 2.69
AFFIRMED openai-c
CUSUM 0.00 of 2.0 alarm
detects ≥ 2.91
CONFIRMED google-b
CUSUM 0.75 of 2.0 alarm
4 canary diverged
CONFIRMED google-c
CUSUM 0.00 of 2.0 alarm
5 canary diverged
CONFIRMED google-d
CUSUM 0.00 of 2.0 alarm
5 canary diverged
CONFIRMED xai-b
CUSUM 0.00 of 2.0 alarm
5 canary diverged
AFFIRMED deepseek-b
CUSUM 0.00 of 2.0 alarm
detects ≥ 2.91
NEW BASELINE simulant-a
reset new reference set
baseline forming

Publishing as gap: openweight-a.

One state per seat per occasion. AFFIRMED ships its no-change e-value and the smallest shift this occasion could detect (n and power pre-registered); NEW BASELINE means the served build changed and comparisons reset. actions.json · how states are decided

Reading the quiet

Nothing moved past the pre-registered thresholds this week, and that sentence is the product. A change-alert service earns the right to be believed on its loud days by publishing its quiet ones: same cadence, same thresholds, same evidence trail, week after week. When a change does confirm, it appears here, on the changelog, and as a diff card with the full evidence bundle.

Pre-series note: this digest counts pipeline activity while the instrument is still measuring itself; no findings publish until methodology v1.0 freezes.

Want this scoped to your builds, “all quiet on the 3 builds you watch”, in your inbox every week?

Watch your builds →
Cite this Permalink https://modelometer.com/digest · Run hash 7da572fcdd5a29dcfea04f4f3c68b5011e0ea0b66fe17bae5bd0bb909468c454 · Retrieved 2026-07-21 · week 2026-W30: 0 changes confirmed across 4000 trials (pre-series).