Grounded Numbers: AI summaries with no made-up figures

This is the hand-over for the skill's worked example; the run output in it is the example's real
output. Paste everything inside the block below (not the four-backtick lines) into the Result
field.

**Summary**

I added a monthly sales summary that Claude writes and code checks. Every number is computed from
your orders in code; Claude only writes the sentences and cites the facts each one uses; a verifier
checks every figure before anything reaches the dashboard. If Claude is slow, fails, or still gets a
number wrong after two rewrites, a summary built from the same facts is shown instead. 94 tests
pass, and none of them needs an API key.

**What I built**

| File | What it does |
| --- | --- |
| `evidence.py` | Computes October's facts from orders placed up to 31 October, so later orders never change it |
| `claims.py` | Reads each number and date Claude writes, as written: `$513`, `$0.5k`, `-10.0 points`, `14 October` |
| `verifier.py` | Checks every line against the facts it cites: value, unit, citation, "up" or "down" |
| `fallback.py` | The summary shown when Claude cannot be used, built from the same facts |
| `pipeline.py` | Draft, check, at most two rewrites, a 20-second deadline per call, a final check before publishing |
| `models.py` | The model interface, plus a fake model for tests |
| 6 test files | 94 tests: planted errors, rewrites, timeouts, templates, later orders |


**A run with a planted mistake.** The first draft said revenue was $521.75 (it is $512.75) and wrote
a 10-point fall as a percent. Both were caught and sent back, and the rewrite was published:

```text
draft: problems - headline: '$521.75' is not in the evidence; the evidence says $512.75; sentences[2]: '10.0%' is written as percent but returning_share_change is in points; the evidence says -10.0 points
rewrite: ok
```

Published headline: "October 2026: revenue up 89% to $512.75".

**When Claude cannot be used**, the dashboard shows:

```text
Juniper & Vale, October 2026: revenue $512.75
Revenue was $512.75, up 89.2% on September 2026 ($271.00).
There were 10 orders, at an average of $51.28.
The best day was 14 October 2026, with $179.00 in sales.
**Desk lamp sold the most:** 4 units for $256.00.
Returning customers placed 40.0% of orders, 10.0 points lower than in September 2026.
```

**Next steps**

1. Point `build_evidence` at your orders table and check its facts against last month's books.
2. Connect Claude through the model interface (an adapter is in the skill's
   `references/model-calls.md`), keeping the SDK's timeout inside the 20-second deadline.
3. Run a few real months and read the summaries before turning the feature on for everyone.
4. Decide on the extras: a second-model check for meaning, and storing each summary's trace.