How to choose an AI model

What to verify before integrating any model

• The model's lifecycle status — is it active, preview, deprecated, or retired? Retirement dates often appear months before they bite.
• The model's context window — does it cover your full prompt plus retrieved context plus expected output?
• The model's max output tokens — separate from context window. A 1M-token context with 8k output limits some workloads.
• The model's input modalities — text only, or images, audio, video too?
• The pricing reference — is it first-party or hosted? Different sources, different freshness, different terms.
• The verification status of every field above — does the catalogue have a primary-source citation on record?
• The data gaps the catalogue marks with the unverified-data label — those are the questions you still need to answer externally.

Inspection only. The catalogue does not rank these models on the field above.

Common mistakes

• Treating context window as a hard fit indicator alone.

  A model with a huge context window can still degrade on long inputs. Context size is a necessary condition, not a sufficient one.

• Reading hosted pricing as the model creator's pricing.

  Hosted pricing is set by the hosting platform. The same model can have very different pricing semantics across creators and hosts.

• Ignoring lifecycle status.

  Integrating a deprecated snapshot weeks before its retirement date locks the team into a migration immediately after launch.

• Treating a missing value as zero.

  An unverified field is not 'no support' — it's 'no primary-source citation on record yet'. Confirm externally before assuming either way.

Apply this workflow

• Open the selection workspace →

  Narrow a source-backed shortlist using verified catalogue fields.

• Compare verified fields side by side →

  Render up to four models against each other from the typed data layer.

• Export an evidence decision brief →

  Generate a paste-ready Markdown or JSON brief — evidence, not a recommendation.

• Inspect the citation registry →

  Open every primary source the catalogue references for a model or pricing row.

Practise this lesson

These exercises route the lesson concept through the verified-data product surfaces. Each one ends with a concrete artifact you can share.

• beginner8 min

  Build your first source-backed shortlist →

  Pick a use case, filter the catalogue by verified fields, and end with a shortlist URL you can share with the team.

• beginner7 min

  Compare context windows without ranking models →

  Use the comparison builder to render verified context window + max output tokens for 3–4 candidate models side by side.

• beginner5 min

  Inspect lifecycle before integration →

  Pull lifecycle state for a candidate model, check for retirement date, and add a migration target to your notes if one exists.

• intermediate10 min

  Create a decision evidence brief →

  Use the decision brief builder to generate a paste-ready evidence pack from your shortlist, then export it in Markdown.

• intermediate12 min

  Plan an external model test →

  Use the brief and the testing lesson to plan your own prompt, latency, rate-limit, and cost validation work for the shortlist.

Teaching example

Illustrative — not a recommendation.

Weak vs better approach

Why better: The weak approach optimises for novelty. The better approach optimises for an auditable trail your reviewer can read independently.

Example artifact

Illustrative example — not a recommendation. Substitute your own values when you run the workflow.

Decision brief excerpt — shortlist note

## Shortlist (illustrative)
Use case: support-ticket summarisation
Candidates: <model-A>, <model-B>, <model-C> (slugs from /select)

Verified fields per candidate:
- Lifecycle: active (cite source)
- Context window: <tokens> (cite source, retrievedAt)
- Hosted availability: yes / no
- Pricing reference: <unit> + retrievedAt

Open data gaps:
- Max output tokens not stated for <model-C> (unverified-data label)

Next step: open /compare/build for the three slugs and export brief.

Substitute your real values when you walk the workflow. The catalogue never generates this artifact for you.

Concept → workflow bridge

1. Step 1

   Learn the concept →

   Frame the decision workflow before opening the catalogue.

2. Step 2

   Apply in /select →

   Narrow a source-backed shortlist by use case + lifecycle.

3. Step 3

   Verify in /sources →

   Read the primary-source citation for each verified field.

4. Step 4

   Test with /lab →

   Run the prompt-testing playbook against the shortlist.

Review before moving on

• ☐I can name the use case and the acceptance rubric.
• ☐Every candidate's lifecycle field is verified active.
• ☐Every candidate has at least one verified pricing reference or an explicit gap noted.
• ☐I have a /select URL I can share with a teammate.
• ☐I have NOT named a winner — only candidates with evidence.

Caution: No persistence — the checklist resets on every visit. Capture progress in your own notes.