# Checking the Finished Figures

This review compares six reported calculations with the supplied sales records. It covers all 30 original outputs. The records describe fictional companies and people.

## Scope and Method

The review took place on September 5, 2026, after the September 3 runs. The six checks were selected after the runs. The reviewer could see model labels. This was not a blind review or an assessment of overall quality.

The expected figures were calculated directly from the source files. A reviewer read each output and selected the reported values. Code checks the arithmetic, displayed precision, file hashes, and evidence text. It does not decide what a sentence means.

Sources: [rules](replication/data/source-pack/august-pipeline/forecast-rules.md), [current records](replication/data/source-pack/august-pipeline/current-pipeline.csv), [prior records](replication/data/source-pack/august-pipeline/prior-pipeline.csv), and [targets](replication/data/source-pack/august-pipeline/targets.csv).

## Expected Figures

| Measure | Calculation | Expected |
| --- | --- | --- |
| Open pipeline | Sum the 16 open amounts | $2,730,000 |
| Weighted forecast | $420,000 × 100% + $950,000 × 60% + $1,360,000 × 25% | $1,330,000 |
| Total target | $700,000 + $650,000 + $850,000 + $700,000 | $2,900,000 |
| Coverage | $2,730,000 ÷ $2,900,000 | 0.941379… |
| Top-two concentration | ($420,000 + $204,000) ÷ $1,330,000 | 46.917293…% |
| Forecast change | $1,330,000 − $945,750 | +$384,250 |

The current two largest deals are Dovetail Travel and Aster Health, whether ranked by amount or weighted value. “Best Case” in the rules corresponds to `BestCase` in the CSV files.

## Recorded Differences

The 180 checks found four incorrect reported values across two files. The other reported values agree at their displayed precision. This result does not validate other figures, explanations, owners, dates, recommendations, or design.

| Original file | Reported | Source result, equally rounded | Evidence |
| --- | --- | --- | --- |
| [First Haiku deck](../../outputs/managed-agents/run-one/deck--haiku-4.5.pdf) | Forecast $760,000 | $1,330,000 | Page 2, line 9 |
| First Haiku deck | Concentration 61% | 47% | Page 2, lines 14–15 |
| First Haiku deck | Forecast increase +$276K | +$384K | Page 2, line 6 |
| [Second Haiku email](../../outputs/managed-agents/run-two/email--haiku-4.5.md) | Concentration 57% | 47% | Line 13 |

Prior forecast totals were checked separately: 25 agree, one differs, and four are absent. The first Haiku deck reports $483,750; the source-derived value is $945,750. An absent prior total does not make a correctly reported change wrong.

## Every Output Checked

“Pass” below applies only to the printed numerical result at its displayed precision. It is not a quality score. The exact values, source calculations, and evidence locations appear in the [JSON record](output-headline-audit.json).

| Attempt | Output | Open | Weighted | Target | Coverage | Top two | Change | Prior total |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| First | fable-5 · deck | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| First | fable-5.1 · deck | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| First | haiku-4.5 · deck | Pass | Error | Pass | Pass | Error | Error | Error |
| First | opus-5 · deck | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| First | sonnet-5 · deck | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| First | fable-5 · email | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| First | fable-5.1 · email | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| First | haiku-4.5 · email | Pass | Pass | Pass | Pass | Pass | Pass | Missing |
| First | opus-5 · email | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| First | sonnet-5 · email | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| First | fable-5 · brief | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| First | fable-5.1 · brief | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| First | haiku-4.5 · brief | Pass | Pass | Pass | Pass | Pass | Pass | Missing |
| First | opus-5 · brief | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| First | sonnet-5 · brief | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| Second | fable-5 · deck | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| Second | fable-5.1 · deck | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| Second | haiku-4.5 · deck | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| Second | opus-5 · deck | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| Second | sonnet-5 · deck | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| Second | fable-5 · email | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| Second | fable-5.1 · email | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| Second | haiku-4.5 · email | Pass | Pass | Pass | Pass | Error | Pass | Missing |
| Second | opus-5 · email | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| Second | sonnet-5 · email | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| Second | fable-5 · brief | Pass | Pass | Pass | Pass | Pass | Pass | Missing |
| Second | fable-5.1 · brief | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| Second | haiku-4.5 · brief | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| Second | opus-5 · brief | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
| Second | sonnet-5 · brief | Pass | Pass | Pass | Pass | Pass | Pass | Pass |

## Related Wording Observations

These observations are separate from the six numerical results. They were not exhaustively counted. A correct total can accompany an incorrect explanation; counts of these observations cannot support model comparisons.

| Original output | Observation | Evidence |
| --- | --- | --- |
| First sonnet-5 brief | The printed equation adds $420K and $340K to give $624K weighted. The percentage and total are correct; the equation omits Aster’s 60% weight. The next line gives the correct $204K weighted component. | p1:16; p1:17 |
| First sonnet-5 deck | The headline says the entire gain comes from three category changes. They explain $378,000, not the full $384,250; the chart itself includes Lantern’s $6,250 amount contribution. | p4:3; p4:4; p4:7 |
| First opus-5 deck | The headline says five deals were re-rated. Three categories changed; a fourth deal changed amount and a fifth changed close date without affecting forecast. The table below distinguishes these changes. | p4:5; p4:6; p4:32; p4:40 |
| First fable-5.1 deck | The headline assigns all $384,250 to three category changes. The same page correctly identifies $378,000 from category changes and $6,250 from amount. | p4:5; p4:6; p4:41 |
| First haiku-4.5 email | The email assigns Lantern’s $25,000 amount increase and $6,250 weighted increase to Junction, reporting Junction’s gain as $97,250. The total weighted change remains correct. | line 16 |
| First haiku-4.5 brief | The movement panel says four categories upgraded. Three categories changed; a separate fourth record increased in amount. The reported total weighted increase is correctly rounded. | p1:15; p1:16 |
| Second fable-5.1 email | The text says all growth comes from category upgrades, then lists Lantern’s $6,250. Lantern’s category is unchanged; this contribution comes from amount. The total and component numbers are correct. | line 7 |
| Second haiku-4.5 deck | The text says growth is 40%. The exact source-derived rate is 40.6291%, which rounds to 40.6% or 41%. The adjacent dollar totals and dollar change are correct. This is a secondary rounding issue. | p2:3; p2:4 |
| Second haiku-4.5 deck | The text says either top deal slipping reduces forecast by $378K. That is the combined category-upgrade gain across three deals, not either top deal’s current weighted value. No slip outcome or reweighting rule is supplied. | p8:12 |

## Check the Audit

Download and unzip the [replication package](replication.zip). From its root, run:

```sh
python3 check-output-audit.py
```

Python 3.9 or later is sufficient. No account, network request, or new model run is required. The command checks four source files, 30 original output hashes, the approved audit record, and all numerical comparisons. It also matches quoted Markdown lines.

PDF page and line references identify text extracted with `pdftotext -layout`; they are not printed line numbers. To repeat that extraction with Poppler installed:

```sh
python3 check-output-audit.py --check-pdf-text
```

Without that option, the checker retains the reviewed PDF excerpts and verifies the original files’ hashes. Neither mode automatically identifies which number answers which question or proves that an omitted number is absent.

The original outputs remain unchanged. This review does not estimate future accuracy, review effort, or which model provides the best value.