Compare all three

Import external results

Bring results another system produced into RedCrown as a run you can rank, inspect and share. Before you import, know which class of evidence you are bringing, because the class decides what the record can show and which controls it can carry.

1. Three classes of evidence

RedCrown keeps three classes apart on the record, and never upgrades one into another.

	Reported summary	Scored supplied outputs	Recorded execution
What it is	Aggregate scores supplied by another system: a mean quality, counts, a cost, per candidate.	Case-level outputs scored by RedCrown, with the generation details declared by the supplier.	Evidence collected through an execution path RedCrown observes, limited to what it records: the request, the response, the timing, the provider's usage counts.
Schema	The `ImportResults` upload object: `name`, `objective`, `quality_metric`, `quality_bar`, `step`, `candidates[]` with `mean_quality`, `n_scored`, `n_errors`, `total_cost_usd`; optional `items[]` receipts per candidate.	A scoring request (`experiment_id`, items with references, candidates with a declared `configuration` and outputs), or CSV plus JSONL.	An experiment spec: task, items, references or incumbent, candidate models.
Data flow	Your file to the app, once, when you push or paste it. Nothing is computed from cases.	Local: nothing leaves the machine until you push. Hosted: the file goes to the app and is scored there.	RedCrown sends each item to the configured provider on your keys and stores what came back.
Cost	Whatever the payload declares; `None` means unknown, never $0.	Declared by the supplier, or unknown. No model call, so no provider charge.	Provider charges on your account, at your rates, recorded per candidate.
Prerequisites	An account. Any declared metric.	Local: the CLI, no account. Hosted: an account, one of `exact_match`, `similarity`, `json_field_accuracy`, `f1`.	An account and a connected provider key for every candidate and judge.
Comparisons available	Ranking by the declared numbers. A harness export payload also carries per-check statuses, so two pushed harness runs on one experiment can be compared check by check in the report view.	Offline `redcrown check` between two evidence files of the same feature; the changed cases from the evidence file. Hosted: add a version to an existing feature (Add to feature in the app, or `POST /experiments/score?experiment_id=`) and, with a declared provider and model on each candidate, accept and compare it against an earlier accepted version.	Ranking with receipts; for eligible runs, hosted acceptance and paired comparison against the accepted baseline.
Controls not available	No evaluation identity, so no acceptance and no paired comparison. No case-level evidence unless the payload supplied receipts.	Hosted acceptance and comparison need a declared provider and model on each candidate; without one, a version cannot be compared or accepted. A pushed run carries no evaluation identity, so it cannot be accepted.	None of the above limits, within section 1 of Accepted baselines.

This page is about the first column. The second is Score outputs produced elsewhere and the quickstart. The third is Compare models on your workload.

The record says which class it is. An imported run is labeled as an external-harness upload and its provenance withholds the methodology version, because RedCrown did not compute the numbers. Receipts appear on the record only when the payload carried them; an aggregate-only import shows the summary and nothing per case.

2. Prepare the results JSON

RedCrown expects an aggregate results object with one entry per candidate. The required top-level fields are name, objective, quality_metric, quality_bar, step, and candidates. Unknown fields are refused. A redcrown harness export payload is already in this shape.

Go to Upload (/advanced/upload). Paste your results JSON directly into the text area or load it from a file using Load file.

CLI

Prepare a results.json file in the upload schema. The required fields are name, objective (cost, quality, or latency), quality_metric, quality_bar, step (an object with an id and an incumbent_candidate_id), and candidates (an array of per-candidate aggregate scores, each with at least an id, provider, and model). A candidate may carry items, its per-case receipts; leave it out and the record has none. See the example under the MCP tab for the exact shape.

MCP

Pass the same object shape to import_results. The required fields are name, objective, quality_metric, quality_bar, step, and candidates:

{
  "name": "My external eval",
  "objective": "cost",
  "quality_metric": "exact_match",
  "quality_bar": 0.8,
  "step": { "id": "generate", "incumbent_candidate_id": "gpt-4o-mini" },
  "candidates": [
    { "id": "gpt-4o-mini", "provider": "openai", "model": "gpt-4o-mini",
      "mean_quality": 0.82, "total_cost_usd": 4.10, "is_incumbent": true }
  ]
}

3. Import

Once your results are imported, RedCrown ranks the candidates by the declared numbers, keeps the caveats the payload carried, and makes the run shareable through a share link. An import creates a new experiment, which is a new active AI feature against your plan's cap, unless you append it to an existing one.

After loading your results on Upload (/advanced/upload), click Import. The ranked run report appears immediately.

CLI

redcrown push results.json

push uploads the file to your workspace and prints the run id and the experiment id. --experiment <id> appends the run to a standing experiment instead of creating one. --no-receipts strips any per-case receipts before upload. --proof-link creates a share link in the same call. redcrown import-results results.json is the older command for the same upload.

MCP

Call import_results with the full results object. The response includes the experiment ID and run ID:

{
  "name": "My external eval",
  "objective": "cost",
  "quality_metric": "exact_match",
  "quality_bar": 0.8,
  "step": { "id": "generate", "incumbent_candidate_id": "gpt-4o-mini" },
  "candidates": [
    { "id": "gpt-4o-mini", "provider": "openai", "model": "gpt-4o-mini",
      "mean_quality": 0.82, "total_cost_usd": 4.10, "is_incumbent": true }
  ]
}

Last verified 2026-09-22.