Watch · the recurring product · in preparation

Know when a served build changes.
Confirmed, chain-verified, auditable from the public JSON.

In one sentence

The Drift & Identity Feed will alert you when a model build you depend on changes behavior, a serving-integrity divergence, a refusal-rate move, or a stance shift, each with the full evidence bundle (Δ table, CIs, q-values, raw transcripts, a replay command, and the chain hash), and mail you a monthly attestation PDF per build that drops straight into a vendor-risk or AI-Act post-market file.

Today's actions

the live feed, every seat, one state each
2026-09-19 0 of 20 retained baseline behavior; 13 showed a confirmed serving divergence, 7 flagged a possible serving change.

Two different things are watched here. Serving checks (daily) ask whether the endpoint you call is still the build we baselined. Behavioral measurements (weekly) ask whether its positions moved. Only a battery-confirmed, published result is a material behavioral change. In the pre-series period there are zero of those, by design.

Serving divergence confirmed anthropic-a ›
CUSUM 43.75 of 2.0 alarm
4 items moved1 serving check diverged
Possible serving change openai-a ›
Δ stance 0.60 on 1–5
1 item moved
Serving divergence confirmed xai-a ›
CUSUM 0.00 of 2.0 alarm
4 items moved
Serving divergence confirmed deepseek-a ›
CUSUM 3.17 of 2.0 alarm
4 items moved
Possible serving change anthropic-b ›
Δ stance 0.40 on 1–5
Possible serving change anthropic-c ›
Δ stance 0.80 on 1–5
2 items moved
Serving divergence confirmed anthropic-d ›
CUSUM 15.50 of 2.0 alarm
2 items moved2 serving checks diverged
Possible serving change anthropic-e ›
Δ stance 1.20 on 1–5
1 item moved
Possible serving change openai-b ›
Δ stance 1.00 on 1–5
3 items moved
Possible serving change openai-c ›
Δ stance 0.80 on 1–5
2 items moved
Serving divergence confirmed google-b ›
CUSUM 18.75 of 2.0 alarm
1 item moved2 serving checks diverged
Possible serving change deepseek-b ›
Δ stance 3.00 on 1–5
10 items moved
Serving divergence confirmed openweight-a ›
CUSUM 32.25 of 2.0 alarm
5 items moved1 serving check diverged
Serving divergence confirmed openweight-b ›
CUSUM 27.50 of 2.0 alarm
9 items moved1 serving check diverged
Serving divergence confirmed openweight-c ›
CUSUM 25.50 of 2.0 alarm
6 items moved
Serving divergence confirmed openweight-d ›
CUSUM 21.00 of 2.0 alarm
11 items moved1 serving check diverged
Serving divergence confirmed openweight-e ›
CUSUM 31.50 of 2.0 alarm
6 items moved1 serving check diverged
Serving divergence confirmed openweight-f ›
CUSUM 31.50 of 2.0 alarm
1 serving check diverged
Serving divergence confirmed openweight-g ›
CUSUM 10.50 of 2.0 alarm
4 items moved1 serving check diverged
Serving divergence confirmed simulant-a ›
CUSUM 45.75 of 2.0 alarm
1 serving check diverged

Not measured: google-a, google-c, google-d, xai-b.

One state per seat per occasion, in the daily serving-check vocabulary. A baseline behavior retained seat ships its no-change e-value and the smallest shift this occasion could detect (n and power pre-registered); a new serving baseline means the served build changed and comparisons reset. A material behavioral change is only ever declared by the weekly battery, never by a daily serving check. actions.json · how states are decided

What an alert looks like

facts in the subject line, evidence in the body
IllustrativeSubject: [Modelometer] Change confirmed: <seat-id>, 2 of your 3 watched builds unaffected What: Stance shift on 2 battery items + refusal-rate change Build: <api_model_id> (pinned 2026-XX-XX), watched by you Detected: serving divergence Mon+Tue, published Wed Evidence: Δ table · CIs · q-values · raw transcripts · replay command · chain hash Blast radius: which of YOUR watched builds are affected: [list] Action: none required if behavior-insensitive; re-run your evals if you pin outputs. Not affected: your other watched builds passed all checks this cycle.

Behavioral change follows a separate path whose confirming step is specified and not yet built; nothing on that path publishes today.

The quiet months are the product

An alert that fires is rare by design. The effect-size floors are pre-registered and live. The confirming re-measurement that will gate behavioral publication is specified and not yet built, so the zero-retraction target rests today on those floors and on serving-integrity confirmation alone. The weekly heartbeat is the "all quiet" line in the digest, scoped to your builds: presence without noise. And the monthly attestation PDF states, per watched build, either the change table or the sentence a compliance file actually needs: "No material behavioral change detected at the pre-registered thresholds", with the run hashes and their chain positions and the OpenTimestamps proof reference.

Honest promises

printed here, on purpose

How a change is confirmed: canaries run daily, and a serving-integrity divergence must repeat on two consecutive days before it is published. That path is live. For behavioral change, the weekly battery records a movement past pre-registered thresholds, and the confirming re-measurement that would promote it to a published finding is specified and not yet built. Evidence first, never a speed race. Corrections: policy with a clock, published. What we measure: the first-party canonical endpoint, hyperscaler and gateway stacks can differ, which is exactly what Team's watched-endpoint and the endpoint consistency audit are for. What we cannot see: your fine-tunes and private deployments. Cancellation: one email, honored the same week.

Solo

1 watched build

In preparation
  • Email alerts with evidence bundles
  • Monthly attestation PDF
  • Full JSON history for your build
Register interest

Pricing set at launch

Starter

3 watched builds

In preparation
  • Everything in Solo
  • Webhook / Slack delivery
  • Attestation PDF per build
Register interest

Checkout opens at public launch

Team

Watch your own endpoint

In preparation
  • Everything in Starter, up to 10 builds
  • Your own Azure / Bedrock / gateway endpoint as a watched, audited subject
  • Quarterly change summary · request-a-model: 2/yr
Register interest

Pricing set at launch

Founding

10 seats, presell only

Price-locked
  • Everything in Team
  • 12-month price lock at the founding rate
  • Cancellable from month 3
  • Founding-partners page (opt-in)
Register interest

Opens at v1.0 freeze + 4 clean weeks

Suspicious the endpoint behind your gateway isn't what it claims? That's a different product.

Endpoint consistency audit →

Watch the models you depend on.

Register your interest and we will tell you when the models you name change, and when Modelometer opens keyed watches and endpoint audits to new users.

What registering gets you

  • The public record and its plain-language digest, in your inbox as they publish, once digest mail opens.
  • First notice when a keyed watch on your own endpoint, and one-off endpoint consistency audits, open to new users.
  • An optional design-partner conversation if you want your specific endpoints measured early.

Registering is a signal of interest, not a purchase; nothing here asks you to buy anything today. Registering costs nothing and asks for no card. And the standing rule behind the whole record: we take no payment from the model providers we measure. What you are registering for is an independent instrument, not a vendor's marketing list.

No card, no account. One email when there is something real to show you.