# Phase 5 brief: verifying the scores

This is the exact brief each phase 5 verifier received. It is kept as part of
the methodology disclosure (see `00-methodology.md`).

## Your role

You are an independent verifier. The phase 4 auditors scored each product
against 149 requirement rows. Your job is to try to prove them wrong, on a
sample of their cells. You did not produce these scores and have no stake in
them. A verification that confirms everything is only credible if you
genuinely tried to break each cell.

## The sample

Your cells are listed in `verify/sample.json`, drawn by `verify/sample.py`
with a fixed seed so the draw can be repeated. For each product, it contains:

- **10 documented cells**, drawn at random;
- **every composable cell**, because a composable pass rests on the auditor's
  own reasoning;
- **every claimed cell**;
- **8 not-documented cells**, drawn at random.

## What to check

| Cell type | Your test |
| --------- | --------- |
| documented | Fetch the cited URL (or PDF page). Is the quote there, word for word? Does it actually meet the row's `passCondition`, read strictly, or only something close to it? Is the mechanism right? |
| composable | Follow the recipe step by step. Is each step's citation real and saying what the recipe claims? Would the steps together actually meet the pass condition? Is it a combination the vendor shows, or one the auditor built? |
| claimed | Is there documentation (not marketing) that would make this `documented`? Or is even the claim overstated? |
| not-documented | Search again, independently, with your own queries against the vendor's public material. Can you find evidence that the auditor missed? |

The same rules bind you as bound the auditors. Use only vendor material
reachable without login, trial, sales contact or video. Do not visit
cargo-planner.com. In `/workspaces/cplweb`, read only:

- `research/market-audit/00-methodology.md`
- `research/market-audit/02-audit-brief.md`
- `research/market-audit/04-scoring-brief.md`
- this brief
- `research/market-audit/03-taxonomy.json`
- `research/market-audit/manual-checks.md`
- `research/market-audit/verify/sample.json`
- the `scores/<id>.json` files for your products

Access notes:

- magiclogic.com and topseng.com pages need a normal browser User-Agent
  header with curl.
- The MagicLogic help centre reads through its Zendesk API.
- PDFs need text extraction with pdf.js, in your own scratchpad subfolder.

## Verdicts

For each sampled cell, record one of:

- `confirmed`: the cell stands as scored.
- `quote-wrong`: the quote is missing or altered, but the verdict still
  stands on other evidence (give it).
- `overcredited`: the evidence does not meet the pass condition, or the
  mechanism or availability is too generous.
- `undercredited`: evidence exists that the auditor missed (give it, quoted,
  with the URL).
- `unclear`: you cannot decide. Say why.

Write `/workspaces/cplweb/research/market-audit/verify/<id>.json`:

```json
{
  "id": "",
  "verifiedOn": "2026-09-30",
  "cells": [
    {
      "row": "",
      "sampledAs": "documented|composable|claimed|not-documented",
      "verdict": "confirmed|quote-wrong|overcredited|undercredited|unclear",
      "reason": "",
      "suggested": { "evidence": "", "mechanism": "", "availability": "" },
      "newQuotes": [{ "text": "", "url": "" }]
    }
  ],
  "patterns": "systematic problems you noticed in this auditor's scoring, if any"
}
```

Save after each product. Your final reply should give:

- the number of verdicts of each kind, per product;
- the cells you would change;
- any pattern that suggests a whole category of cells should be re-checked.