Skill Upgrade Evidence Check

Draft awaiting hosted acceptance and platform review. The offer is not yet active.

Review completed Skill evaluation records before adopting a change. Check planned coverage, inspect reported regressions and keep useful observations when the overall comparison is blocked. The maintainer makes the release decision.

Subscription offer

First 3 days free, then US$2 per week before applicable taxes. The target trial is the first 3 consecutive days of the first eligible subscription; it is not a new free period at every weekly renewal. Actual eligibility and trial timing remain to be accepted in the enabled platform offer. The subscription provides access to this checker through Capafy's cloud runtime. It does not sell a recurring download of the source package or include a separately charged Apify report.

Review the displayed trial start/end, first charge, renewal dates, usage conditions and cancellation deadline before subscribing. Paid access renews weekly unless cancelled. Use Capafy's cancellation controls; its published subscription terms require cancellation at least24 hours before the next paid renewal. Trial-end cancellation and billing details must be displayed by the enabled offer. Do not treat a post-payment refund window as a free trial.

Proposed request allowance, not activated: The first eligible 3-day trial includes 3 requests in total. Each paid weekly billing period includes 20 requests. Each accepted checker invocation counts as one request, including a blocked result, invalid input or a failed accepted request. Starting another invocation, including a retry, uses another request. Downloading an existing result without starting a new invocation does not use a request. Each request handles one completed evaluation snapshot. Unused requests do not roll over; no automatic paid overage is included. These proposed limits must be enforced by the platform before the offer is activated.

Bring completed evidence

Provide the original evals.json, completed iteration evidence, explicit candidate and baseline names, and original planned repeat count. The free MIT input-only exporter is available separately; preparing compatible evidence does not require a paid Capafy purchase. Two SKILL.md files or an aggregate score are insufficient.

The checker checks records already produced. It does not run a Skill, re-grade responses, call an external model for the evidence comparison, verify a grader's truthfulness, or approve an upgrade automatically. Capafy routes conversational messages and relevant attachments through its sandbox agent and configured external model provider. The deterministic checker itself makes no model or network calls; this does not mean the complete platform workflow is model-free.

Reports and limits

The intended hosted output includes JSON, a Markdown report and a portable evidence ZIP with relative links. Complete evidence yields READY_FOR_HUMAN_REVIEW. Missing or inconsistent evidence can yield BLOCKED_COMPARISON with partial observations retained. Invalid input yields INPUT_ERROR and no evidence ZIP. A blocked comparison is a useful delivered result, not a promise that an upgrade is safe.

The authored example has8/8 slots, means87.5% and75%, a12.5 percentage-point difference and one regression. These are synthetic records, not customer results. Core limits remain8MiB JSON,1000 selected files,3000 directories,2MiB/file,10000 inspected entries,200 cases and1–20 planned repeats; this subscription does not promise unlimited runtime or model use.

Review evidence for private content before cloud submission. There is no automatic redaction. Subscription access processes evidence on Capafy; its sandbox agent can share messages and relevant attachments with the disclosed external model provider. Download required reports and use the platform's access and retention controls. No blanket zero-retention promise is made.

Access and support

Subscription access is for personal or authorized internal business use through the platform. No complete source download is included. Stopping a subscription controls future hosted access; it does not remotely delete a user's saved reports or previously delivered files. Access, cancellation and refunds follow the displayed offer and applicable platform terms.

Support: [email protected]