How to Review a Pinned Model Version and Deprecation Notice

Bo, Ops at a 20-person UK health-tech, is about to treat “latest” as a SKU because Finance signed off and Legal said they always ship the current model. Knowing how to review a pinned model version and deprecation notice is a 25-minute hunt: write the named ID, treat a silent repoint and a safety-filter swap as material, demand notice plus a parallel-test window, then walk or take a credit if the successor fails your eval.

September 2026. English law; exclusive courts of England and Wales. The PDF is an MSA — master services agreement, the frame the order form hangs on. Line 4: “Model: latest (alias gpt-x).” Schedule B: changelog notice, fourteen (14) days. No named snapshot. No dual-run. The SLA exhibit — the uptime schedule stapled to the order form — is 99.9% monthly uptime; credits are the sole remedy. A “safety classifier refresh” then flipped eval’d prompts to refusals overnight. Status page: green. Friday is booked. The rush is the problem. The hidden risk is a pointer, not a product name.

A pin is not immortality. OpenAI Deprecations (checked 19 September 2026): generally available models get at least 6 months, specialised variants 3 months, preview names as little as 2 weeks — unless safety requires faster. Bo’s paper has no named ID.

Disclaimer: Checkory provides AI support, not legal advice. Consult a qualified lawyer for binding decisions.

Which model string sits on the order form — a named ID or latest?

Circle the model string on the order form. An alias is a pointer the vendor can repoint without a 404. Write the exact identifier the API will accept — a dated snapshot, or the vendor’s canonical pinned ID. Do not treat “latest” or gpt-x as a product name. Bo’s line 4 is alias — fail.

Freeze the packet — the signed MSA, order form, SLA schedule and updates appendix — dated today. Open line 4 and Schedule B. If the fight is training on prompts, files, outputs or logs, that is a different hunt: vendor AI training rights. If the fight is a posted AUP, SLA or price-list URL, that is a different hunt: unilateral online-policy updates. Stay here for the identifier. Google’s Gemini Models page (updated 17 September 2026) says gemini-flash-latest will get hot-swapped with every release. For example, Tian Pan (2 July 2026): order-form “latest model” is unbounded silent behaviour change. Put the snapshot on line 4.

What does a silent alias repoint or a safety-filter swap count as?

Write into the paper that a silent alias repoint, a successor snapshot, and a safety-classifier change that alters refusals are material, even if HTTP stays 200. A changelog “safety refresh” is not a footnote. Do define both as material. Do not shrug because the status page stayed green.

Inferbase (3 July 2026) watched mistral-ocr-latest jump a major generation between 18 June and 3 July 2026 with no deprecation notice and no error — the request still returned 200. Anthropic’s Model IDs page admits safety classifiers can move inside a pinned ID. Bo does not even have the ID. Last month’s refusal spike is the costly pattern. Do: log “safety refresh = material.” Don’t: wait for a 404.

When to demand a deprecation notice and a parallel-test window?

Circle the fourteen (14) day changelog email. That length is preview-tier on the big-provider docs pages, not generally available. Demand written notice to a named inbox plus a dual-run window long enough for her golden set. Do log notice days and dual-run days. Do not treat 14 days as enough for a clinical-adjacent eval.

OpenAI’s 6 / 3 / 2-week floors sit on a docs page the vendor can edit — not Bo’s paper. Anthropic deprecations: at least 60 days’ email before retirement; the 5 June 2026 Opus 4.1 notice retired on 5 August 2026. Azure Foundry: GA email at least 60 days, replacement overlap about 90 days. Google’s breaking latest-swap email is 2 weeks — Bo’s changelog length. In practice, dual-run is weeks. Put days in this agreement, to a named inbox.

Why does a successor that fails your eval still need a walk or a credit?

The successor must pass her frozen prompt set — format, refusal boundaries, clinical-adjacent tasks. If it fails a written threshold, she may stay on the pinned ID, take a fee credit for re-eval, or terminate that SKU. Do not accept “commercially equivalent” with no test.

Kanopy (2026) is market ask, not a rule: pin for a defined period and at least 6 months’ deprecation notice. RFE Online (Australian colour, not England-and-Wales statute): exit if a version change degrades your evaluation set. Do: freeze the golden set. Don’t: let “you refused the upgrade” become a fee.

What does a 99.9% uptime SLA actually measure?

A 99.9% monthly uptime credit is the endpoint answering. It does not cover a dumber snapshot, a tighter filter, or eval drift. Write SLA equals HTTP. Do not treat a green status page as cover for last month’s refusal spike.

Tian Pan’s SLA Illusion (7 May 2026) is the structure: 99.9% is requests accepted, not accuracy. Do: write “HTTP, not quality.” Don’t: treat a green light as last month’s eval.

How do you choose pin, notice, or walk?

Pin only if the log shows a named ID, material-change language covering repoint and filter, inbox notice that is not a 14-day changelog, a parallel-test window, and a walk or credit if her eval fails. Notice means redline those four. Walk — do not sign Friday — if latest plus 14-day changelog plus no dual-run remain as a package.

Success bar: a one-page log plus one sentence that would pause Friday. Bo already has three: alias gpt-x; fourteen (14) day changelog; no dual-run. Workflow: named ID vs latest/alias → material (repoint + safety filter) → notice + parallel-test → walk/credit on failed eval → SLA is not quality → pin / notice / walk. Spend counsel — the England-and-Wales solicitor who opens line 4 — on that package. Pause Friday if the log still says alias plus changelog-only. Escalate the AE’s “always current” line under Friday pressure. Optional: upload the same PDF to document analysis for a first-pass — a machine extract of clauses before a human reads every page — then a human still opens line 4 and Schedule B and verifies every High flag — a high-severity extract a human must still open.