---
name: research-harness
description: Turn broad research or proof-like work into deduplicated, testable obligations and coordinated evidence-producing braids.
version: 1.0.0
updated: 2026-08-14
---

# Research Harness

Use this for open-ended technical research, formal proof work, architecture exploration, or any task where “try more ideas” could consume unbounded time.

## Normalize before spending

Write a source-obligation record containing:

- the precise input and desired output;
- the contract, theorem, or behavior shape that would connect them;
- the strongest relevant primary source or executable artifact;
- the stress model or counterexample that could kill the claim;
- the source distance: direct, one transformation away, or speculative;
- a measurable gradient showing that another run can change what is known.

Deduplicate by this structure, not by vocabulary. Different names for the same obligation share one record. If there is no measurable gradient, sharpen the obligation or route it to a bounded checker instead of spending a new wave.

## Run a braid

For hard obligations, combine complementary lanes around one shared target: generation, literature or source retrieval, decomposition, adversarial falsification, executable/formal verification, and integration. Each lane produces a durable artifact that constrains the others. The braid has one owner responsible for reconciling contradictions and promoting only supported results.

## Promotion rules

A candidate is not a result. Promote it only when its dependencies, decisive check, counterexample search, provenance, and remaining uncertainty are recorded. Check internal results before killing or reviving an avenue. A renamed obstacle is not progress.

Keep findings, failures, metrics, and corrections append-only. Preserve original metadata and link superseding entries. For long runs, checkpoint outputs, maintain a lease/heartbeat, and document safe resumption; do not reap a productive run merely because it is old.

Report the smallest honest state: drafted, ranked, decomposed, checked, falsified, verified, or promoted.

For a load-bearing obligation, use `braid-discovery`: hash the head of record; pin prompts and verdicts; check the smallest explicit instance; run differently shaped prove and refute branches; and issue a method-obstruction receipt for each abandoned approach. A reformulation must state where the hardness moved. Consolidate supported edits, recompute hashes from disk, and seek fresh review over the new bytes.

When a research receipt is used to claim a build item complete, it must conform to the one canonical
[evidence receipt schema](../tutorial-verified-done/references/evidence-receipt-schema.md). An
obstruction receipt alone never satisfies the build completion gate.

Persist the exact long-run session ID and resume point. Maintain a compact append-only tail/index so resumption does not require a destructive cold start or full-history reread.