Research Contract for Agents: claims, sources and dates your agent must keep
by PaperChaseWebb
A contract your agent follows on every research run, kept as plain files: each claim with its source, date and confidence, primary sources before secondary, quotes word for word with a location, contradictions listed, unknown allowed, and no citation to anything it did not open. A fifteen-check script reads the folder and names every break in a plain sentence, and nothing is published or acted on until the owner approves the exact report by its sha256.
- Price
- $19
- Difficulty
- Beginner
- Setup
- 20 min
- Format
- .md file
research-contract-for-agents-claims-sources-and-dates.md · 70 KB · instant download after checkout
Want this built for you instead? We can build it for you.
What you end up with
Research you can check line by line: every finding points to a claim, every claim to a source you can open, every quote to its place in a saved copy, with the gaps and disagreements in plain view.
What you get
- 17 sections in one .md file
- The contract in thirteen rules: brief, claims before evidence, primary first, read before citing, verbatim quotes, dated freshness, contradictions, unknowns, separate audit, owner yes
- Templates for the research folder: brief, claims, sources, quotes, snapshots, report, audit, approvals
- A skill file for Claude Code, Codex and Hermes, an AGENTS.md snippet and a Grok Bot rules block
- contract.py: set up a folder, check a quote word for word, owner approval tied to the report sha256, and verify
- contract_check.py: fifteen read-only checks, with a seeded test folder (15 findings, then 0)
- Ten prompts, a failure table and a worked example
- A look at nine research prompt packs and citation tools, with sources
Who it is for
Anyone who asks agents to research vendors, markets, rules or prices and then acts on or publishes the answer.
Works with
Solid chips: the agent is named in this listing's own tools list. Dashed chip: A skill file for Claude Code, Codex and Hermes, an AGENTS.md snippet and a Grok Bot rules block; the scripts were tested on Linux with Python 3.13, the agent files were not loaded inside each harness.
Tools needed
Cost to run: The pack is the only purchase. The checker and approval script use no model tokens; the research itself uses whatever your agent normally costs.
Time to result: First research folder in under an hour
How it works
Run the tests
Copy the research-contract folder and run sh test_all.sh.
Expected: 15 findings on the seeded folder and 0 on the fixed one, 15 of 15 single faults caught alone, checker tests 21 passed, approval tests 18 passed, exit 2 on a missing folder.
Add the contract to your agent
Install the skill file for Claude Code, Codex or Hermes, paste the AGENTS.md snippet, or paste the rules block into Grok Bot.
Expected: The agent starts research with contract.py init and a brief.
Run one real question
Use prompts 1 to 8 for the research and prompt 9 with a second agent for the audit, then run contract_check.py until only C15 remains.
Expected: A report where every finding carries a claim marker, with Contradictions and Unknowns filled in.
Approve or send back
Read the report, then run contract.py approve at your own terminal. The agent runs contract.py verify before it publishes or acts.
Expected: verify prints approved for this exact report; any later edit needs a new yes.
Sample prompts, free to use
Research this under the research contract: <question>. The decision is <decision>. I am the owner. Run contract.py init research/<topic>, fill brief.json, and show me the brief before you search. Before you open any source, list the claims you expect to check in claims.csv, each with what would prove it wrong.
Prompts 1 and 2 of 10.
Run contract_check.py on research/<topic>. Fix everything except C15 and rerun until only C15 remains. Give me the report hash and stop. Do not publish or act until contract.py verify prints approved.
Prompt 10 of 10.
Free preview of the file
# Research Contract for Agents: claims, sources and dates your agent must keep Version 1.0, 2026-10-28. By PaperChaseWebb. Works with Claude Code, Codex and Hermes through a skill file and an AGENTS.md snippet, and with Grok Bot through a rules block. One setup and approval script, one read-only checker with fifteen checks, templates, a seeded test folder, ten prompts and the tests, all Python 3 standard library. Tested as described in section 9: everything ran on Linux with Python 3.13, including owner approvals typed at a pseudo-terminal. The skill file, snippet and rules block were not loaded inside the harnesses for this release, and no live web research was run through the contract. Each is marked where it matters. Ask an agent to research something and you get a tidy answer with links. Then you check. One link was never opened. One quote is a paraphrase in quotation marks. A number comes from a listicle that got it from another listicle. A price is two years old. Two sources disagree, and the report picked one without saying so. Where the agent found nothing, it wrote something plausible anyway. None of this is malice. It is what happens when nothing in the setup asks for anything better. This pack is that setup. It is a contract your agent follows on every research run, kept as a small folder of plain files: a brief, a claims table, a sources table with dates and tiers, verbatim quotes with locations, saved snapshots of what was actually read, a report with claim ma
Worth knowing
- The checker proves the folder is consistent with itself and with the saved snapshots. It cannot prove a snapshot is what a site really said; the auditor spot checks that.
- The skill file, AGENTS.md snippet and Grok Bot rules block were not loaded inside Claude Code, Codex, Hermes or Grok Bot for this release, and no live research run was done under the contract. The test folder is made up. Linux only; Windows was not tested.
- The approval script refuses piped input and the environment variables Claude Code and Codex set. That stops an agent approving its own work by mistake; it is not a lock against someone with the same account.
- On 2026-10-28, on Linux with Python 3.13.5: contract_check.py gave 15 findings on the seeded folder (each check once) and 0 on the fixed one, caught 15 of 15 single faults with only its own check, wrote nothing, and exited 2 on a missing folder or claims.csv; checker tests 21 passed, 0 failed; approval tests at a pseudo-terminal 18 passed, 0 failed.
Questions
How is this different from Check the Claim?
Check the Claim goes deep on one claim at a time: verdicts, arithmetic and red flags. This is the contract for the whole research run, so every claim in a report has a source, a date and a confidence before anyone reads it. Claims that come out partly or contradicted here are good candidates for a full Check the Claim pass.
Does the checker browse the web?
No. It reads the folder only. Quotes are checked against the snapshots the agent saved when it read each source, which is why the contract requires them.
Is unknown really a good answer?
Yes. An unknown that is listed tells you where to look next. A gap filled with a guess looks the same as a finding, which is the problem the contract exists to stop.
Why tie the approval to a hash?
So the yes covers the report you read. If anyone edits report.md afterwards, verify fails and the owner approves the new version.
Details
Version 1.0 · Updated Oct 28, 2026 · Markdown version of this page
License: Use for your own research and on installs you do for clients. Do not resell or republish the pack itself. Download terms
Related
Research Contract: five rules and a five-check script (free sample)
From Research Contract for Agents: five rules for agent research (every claim names its source, quotes word for word with a location, high confidence needs a primary source, two dates on every source, the owner's yes before anything goes out), and a read-only script that checks one research file for them.
Check the Claim (expanded edition): a method, prompts and a skill for checking AI claims
A Markdown pack that makes your agent check income, tool, popularity, quote and funding claims before they reach a brief, a purchase or a post: arithmetic first, primary sources second, one honest verdict per claim.
Compile Your Research: a cited, checked second brain your agents keep
Your agent compiles saved sources into linked markdown pages where every fact cites its file, conflicts get their own page, and nothing is the record until a person ratifies it. Four folders, ten rules, seven page types, ten prompts, a health check script that needs no AI, and a seeded test vault with nine planted problems.
Agent Market Map 2026: who does what, what buyers pay, and where a small team fits
A decision map of the agent market, not a logo wall: seven layers with public examples and dates, six trust patterns buyers keep paying for, a pricing-shape table, a fit worksheet and a ten-line A-or-B scorecard. Plus a kit to keep your own map current: CSV format, rules, eight prompts, a read-only check script and a seeded test map.
Agent Readiness Self-Check: score one workflow before you hand it to an AI agent
A free self-check for agencies and small businesses. Score one workflow across eight areas (workflow mapped, data access, approval gates, owner per agent, handoffs, cost tracking, security and vetting, measurement) with an evidence rule and three gate rules, read your band and what to do next, plus an interview prompt for Grok Bot, Claude Code, Codex or Hermes, a score sheet and a worked example.