# Six trust patterns and the A-or-B scorecard (free sample)

A free sample from Agent Market Map 2026: who does what, what buyers pay, and where a small team fits, by PaperChaseWebb. Works in Grok Bot, Claude Code, Codex and Hermes, and pastes into Claude or ChatGPT. Version 1.0, 2026-10-27.

When a client asks "which agent tool should we use," most answers are a feature list or a gut call. This sample gives you two things from the full map: the six trust patterns that keep showing up in agent products people pay for, and a ten-line scorecard for comparing two tools on evidence. Use them on your own offer or on any two tools this afternoon.


## The six trust patterns

These recur across agent products as of October 2026. Each is a design choice you can copy, or a question to ask a seller.

1. **Human handoff on request is a hard rule.** The customer can always reach a person, and the agent says so.
2. **The person in charge keeps the final call.** The agent prepares; the owner, official or reviewer decides.
3. **The agent never holds the secret.** One-time cards, scoped tokens, read-only access by default.
4. **Proof over claims.** Show the working result, not a list of possibilities.
5. **One job, end to end.** Narrow beats broad when a buyer has to trust it with real work.
6. **Requirements before purchase.** Tell the buyer what apps, access and time they need before they pay.

Audit your own offer: for each pattern, write yes, no, or "not relevant, because." The full pack names public products that show each pattern, with sources and dates.

## The A-or-B scorecard

Score each tool 0 to 2 on each line. 0 means no or unknown, 1 means partly or claimed, 2 means shown with evidence you saw yourself. A score without evidence is a 0.

| # | Question | Evidence to look for |
|---|---|---|
| 1 | Does it do the job itself, or suggest work for a person to do? | A demo or trial where the task completes |
| 2 | What can it touch? | A written list of apps, data and permissions |
| 3 | Who approves risky actions? | An approval step you can see and test |
| 4 | Can a customer reach a person? | Handoff stated and working |
| 5 | Does it hold secrets? | Scoped tokens, one-time cards, read-only defaults |
| 6 | Can you see what it did? | Logs or a run history you can export |
| 7 | Can you stop it? | A pause or kill switch, tested |
| 8 | Do you know the full cost? | Price page plus usage costs, in writing |
| 9 | Can you leave? | Export of your data and configuration |
| 10 | Is the claim checkable? | A public source or your own test for each claim |

A tool that scores 2 on lines 3, 5 and 7 is usually a safer first install than one that scores 2 on line 1 and 0 on those three.

## The prompt

Paste this into Grok Bot, Claude Code, Codex or Hermes, or into Claude or ChatGPT with web access.

```text
Compare <tool A> and <tool B> for this job: <job>. Read only their own public
pages and docs. Use the ten-line scorecard below. For each line give each tool a
score of 0 to 2 and the evidence you saw, with the URL and today's date. Anything
you could not see is 0. Do not recommend one, do not rank them, and never sign
up, start a trial, book a demo or contact either company. End with the lines
where both tools scored 0.

<paste the scorecard table>
```

## Test it in two minutes

Run the prompt on two tools you already use. Check one score yourself by opening the URL it cites. If the evidence is not on that page, the score should have been 0; tell the agent and rerun.

Harness notes: in Grok Bot, put the prompt in a skill so you can reuse it. In Claude Code put the rules in CLAUDE.md; in Codex, AGENTS.md; in Hermes, the system prompt or a skill file. All four need web reading turned on.

## What is in the full pack

The seven layers of the agent market with public examples and sources, a pricing-shape table dated October 2026, the fit worksheet for placing your own offer, a CSV format and rules for keeping your own map current, eight prompts with approval steps, map_check.py (a read-only check for missing sources, stale rows, hype and unsigned approvals), and a seeded test map with eight planted problems. $19.

Made by PaperChaseWebb at paperchasewebb.ai.