Start with the business problem, not a model¶

You do not need to know model families, metric names, or Roboflow APIs before starting. You do need authority to use the data, representative examples, the business consequence of errors, and a human who owns the result.

Count pallets crossing this line and report the hourly total. I have 60 sample
frames. A missed pallet is worse than a duplicate count.

The expected answer is not a list of models. Sentinel should clarify the output, translate the consequence into a proposed eval, measure a baseline, and recommend the least costly next step.

What Sentinel owns¶

Sentinel does not own current Roboflow product truth. Exact MCP behavior, model IDs, Workflows, platform navigation, plans, and prices should come from installed official Roboflow skills, then exposed roboflow://skills/... resources. Read Roboflow Skills Integration.

Support tiers¶

The full route-by-route claim register is in Support & Evidence.

Install¶

codex plugin marketplace add https://github.com/Borda/vision-delivery
codex plugin add sentinel@sentinel

claude plugin marketplace add Borda/vision-delivery
claude plugin install sentinel@sentinel

Each host uses two commands. The v0.2 package has passed local clean-home marketplace simulations; the public-GitHub path remains unverified until these files are published and retested from main. No credential environment variable is required for plugin installation. When the host or account requires it, using the MCP capability requests hosted authorization; an existing authorized session may need no prompt. For Claude plugin development from a checkout, use claude plugin validate . and claude --plugin-dir .. Never paste credentials into chat.

Delivery loop¶

flowchart LR
    A[Outcome and authority] --> B[Representative samples]
    B --> C[Task and success gate]
    C --> D[Cheapest baseline]
    D --> E{Pass?}
    E -->|no| F[Investigate and improve]
    F --> G[Confirm before paid action]
    G --> D
    E -->|yes| H[Verify artifacts]
    H --> I[Economics and human decision]

Generated scripts, IDs, eval files, and ledger rows are not self-validating. Inspect and execute them on a representative fixture, compare observed outputs with the committed gate, and record failures before calling a proof complete.

Evidence today¶

• Historical routing: one pre-v0.2 Claude Sonnet run over 143 prompts reported precision 0.94 and recall 0.85, with 4 false positives and 11 false negatives. It excludes the delivery and setup routes and is not current-route, Codex, or task-completion evidence.
• Synthetic process A/B: 16 mocked runs at one repeat per cell produced 1 supported cell, 6 mixed/parity cells, and 1 loss. It is directional, developer-contaminated process evidence—not live model quality.
• B1 detection: a recorded result on 11 private test images. It lacks a controlled comparator, post-train count MAE, public independent reproduction, and verified transfer to the intended conveyor domain.
• B2-B5: benchmark specifications only; live results pending.
• Entry barrier: no independent novice study yet measures completion, recovery, or safe decisions.

See Benchmarks and Evidence for sources and promotion gates.

Safety before capability¶

Paid-action confirmation is an agent instruction, not a hard authorization control. Use host approvals, least-privilege account authorization or sessions, account budgets, and non-production workspaces. If a separately generated standalone client requires a key, follow the provider's current guidance and scope that key to the minimum permissions.

For faces, license plates, people tracking, forms, medical imagery, worker monitoring, minors, or location-linked media, stop until authority, purpose, minimization, retention, representative evaluation, named human review, and appropriate legal/security/domain checks exist. Sentinel must not be the sole decision-maker for medical, employment, law-enforcement, access-control, or physical-safety outcomes.

Read Trust and Safety and the repository security policy.

Continue¶