Put the requirements, test results, and limitations in front of your client. Use RedCrown's supported checks and reports to make the evidence behind your recommendation inspectable.
"We tested it and it works."
Model selection is a supporting engagement, not the only one. RedCrown also runs one task across several models on the client's own cases and ranks them on the quality bar you set, priced at their volume. Two anonymized examples are below.
Every item below is something the released product produces. Nothing on the list is a signature, an approval chain, or a release authorization. RedCrown records; the client decides.
A signed attestation shows the record was issued by RedCrown and not altered since. It does not show that the supplied cases represent production, or that a declared configuration was the one that ran.
Both anonymized. Quality was measured on the client's own data in each case and monthly cost was projected from provider list prices at the same volume. The sample size and the caveats are stated on every card, and each result holds on that sample only.
Measured on 9 clips. The 0.4-point accuracy difference is within a single clip at that sample size, so the accuracy result is directional and the cost result is the finding. It holds on this sample, not in general.
Measured on 4 notes. The pass-rate gap is a single note at that sample size, so it is directional and the cost result is the finding. It holds on this sample, not in general.
You do not have to move the client's evaluation stack to get a record. The input you bring decides what evidence the record can carry. RedCrown can also run the evaluation for you, on your keys, when you want it to.
The cases, the references, and the outputs the feature produced. redcrown score scores them offline and writes per-case receipts and a private evidence file. No provider key, no model call. RedCrown did not see how the outputs were generated, and the record says so.
A journal your test runner already writes: one directory per run, one line per case, named checks that are ok, fail, or unmeasured. redcrown harness compare lists what regressed, what was fixed, and what could not be measured. Your runner executes.
Aggregate scores from another system become a run with summary scores. Per-item receipts appear only if the payload carried them, and a summary import gets no evaluation identity, so it cannot be accepted as a baseline or compared in the app.
Providers ship new models and change prices, prompts drift, and the client's cases change. Re-runs happen when you run them: on demand, or in the pipeline you configure. Nothing runs on a schedule and nothing sends an alert.
redcrown check, offlineRe-running is on demand or in your pipeline. RedCrown does not re-run on a schedule and does not send change alerts. Neither is built, and this page will not say otherwise until one is.
pip install redcrown, then follow the quickstart. It runs a synthetic regression check on 40 invented support tickets with no account and no provider key. The commands beside this are the ones the quickstart prints; the output is what the released CLI returned.
Terminal session: pip install redcrown, score the synthetic ticket-router fixture, then check candidate v2 against the reference run under the policy. The check fails with exit code 2 because the minimum quality of 0.95 and the maximum regression of 0.05 were both breached at an observed 0.9 and 0.1.
$ pip install redcrown $ redcrown score score-request.json --evidence-json evidence.json $ redcrown check evidence.json --candidate-id candidate-v2 \ --baseline evidence.json --baseline-candidate-id reference-v1 \ --policy policy.json redcrown check: fail (exit 2) - minimum_items: pass [requirement_met]; observed=40, threshold=40, count=40, denominator=40 - maximum_error_rate: pass [requirement_met]; observed=0.0, threshold=0.0, count=0, denominator=40 - maximum_ungraded_rate: pass [requirement_met]; observed=0.0, threshold=0.0, count=0, denominator=40 - require_known_config: pass [requirement_met]; observed=True, threshold=True, count=1, denominator=1 - quality.minimum: fail [threshold_breached]; observed=0.9, threshold=0.95, count=40, denominator=40 - quality.maximum_regression: fail [threshold_breached]; observed=0.1, threshold=0.05, count=40, denominator=40
A signed, shareable record with every case's receipt, and a publish gate that refuses a number the report itself calls unsound.
A page with a receipt for every case: the input, each candidate's output, and the score it received. No login to read it. Inputs, outputs, and references are visible on a public link, so choose the disclosure level before you create one.
A report can carry a signature you verify offline against the public key at /.well-known/redcrown-attestation-jwks.json. It shows the record was issued by RedCrown and not altered since.
A share link is refused while the report carries a blocking caveat, such as an unscored case or a truncated output. An owner can override the gate, and the override is recorded as an audit event.
Free for one active AI feature, in your browser or your terminal. No sales call, no credit card.