Bootstrap Evidence App

Bootstrap evidence app

Trigger and refusal

Use for a complete application bootstrap, not merely installing coding tools. Do not publish private context, invent proof, grant access or activate a harness without separate authority. Synthetic examples prove contracts, not applications.

Workflow

Before implementation, copy templates/harness/outcome-contract.json and declare the trigger, current/candidate envelope, target, exact cases, expected before/after, invariants, owner and stop criteria. Unknown baseline means DISCOVERY. Use scripts/outcome_review.py to validate, route and record one append-only history; derive HTML/ASCII from it. Keep failures, budgets and approvals across handoffs. Profile enterprise data, model, telemetry and operator boundaries when applicable.

1. Inspect instructions, manifests, history, schemas, tests and existing services. Search only authorized local and organizational sources. Inventory provenance, scope, freshness, duplication, conflicts and incomplete searches. Decide REUSE, EXTEND, BUILD or BLOCKED for each capability before implementing it.
2. Define one persona, decision and end-to-end proof. Keep measured facts outside model judgment and human approval outside model authority.
3. Implement bounded, authenticated data sourcing with timeouts and privacy. Curate source-linked versioned cases with group-disjoint visible/heldout splits. Keep heldout contents outside candidate access, not merely labeled.
4. Reuse the runtime and add request/tool/result/score trace correlation. Start locally; configure Langfuse only with authorized credentials. Verify hosted record readback, not just SDK flush. Avoid raw private-input export.
5. Implement the actual UI journey using assistant-ui when compatible. Render validated results and citations from the same backend receipts. Exercise streaming, cancellation, errors, retry, denied access and durable reload. Capture real viewport evidence and inspect console/network failures.
6. Define descriptive boolean hard gates and five text-anchored ordinal grades, each with rationale. Calibrate on reviewed examples and negative controls. Provider failures are missing evidence, not scores. Freeze dataset, rubric, model and evaluator identities before comparing runs.
7. Distinguish current-run repair from future harness improvement. Mine trace-linked weaknesses. Propose one bounded harness change with source, scores and execution traces. Keep policy, judges and heldout cases outside candidate edit scope. Reject hard-gate failures and split regressions.
8. Numerical eligibility is not acceptance. Require a distinct human acceptance receipt and rollback path before activation. Report unresolved access gaps.

Mechanisms and receipt

Apply task/meta improvement separation, weakness/proposal/validation loops, source-plus-score rollout feedback, and task-specific capability assembly within the existing runtime. Do not create a competing orchestration platform.

Return source/reuse inventory, implemented slice, versioned evaluation inputs, baseline/candidate results, real trace/UI proof and a promotion decision. Validate the generic skill receipt with references/contract.json; passing receipt validation does not prove the application works.

Read the execution contract before starting. Validate the final receipt with its parameterized verification.argv, replacing {python}, {skill_dir}, and {receipt} with concrete local values.