# Recorded result batches for Expense Analysis Part Three

These scorecards are pinned to public Skills-and-Evals revision `5ec0b0a9bd4c9361cb059984675cd2e62352c699`. They support the two result tables in [Verified, Not Trusted: Expense Analysis, Part Three](https://matlus.com/writing/verified-not-trusted-expense-analysis-part-three/). Counts describe selected checks, not whole-report accuracy.

## Table 1: three prompt variants on Opus and Sonnet

All listed trials use medium effort. Each row below describes per-check counts, not the proportion of reports passing every check.

| Article comparison | Exact evidence | Verified below-perfect checks |
| --- | --- | --- |
| Standard, Opus 5.5 | [C1 scorecard](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/results/pwi/expense_analysis/20261010T112320Z-regrade-claude_code-claude-opus-5-5-medium-d5f700/scorecard.md) ([JSON](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/results/pwi/expense_analysis/20261010T112320Z-regrade-claude_code-claude-opus-5-5-medium-d5f700/scorecard.json)), `standard` rows | August Travel 4/5; uncategorized count 0/5; largest increase 2/5; three findings 3/5 |
| More precise, Opus 5.5 | Same C1 files, `robust` rows | Tables before interpretation 0/5; every other check 5/5 |
| Standard, Sonnet 5.5 | [C2 scorecard](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/results/pwi/expense_analysis/20261010T112458Z-regrade-claude_code-claude-sonnet-5-5-medium-e976c6/scorecard.md) ([JSON](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/results/pwi/expense_analysis/20261010T112458Z-regrade-claude_code-claude-sonnet-5-5-medium-e976c6/scorecard.json)), `standard` rows | August Travel 3/5; uncategorized count 0/5; largest increase 2/5; tables before interpretation 3/5; three findings 2/5 |
| More precise, Sonnet 5.5 | Same C2 files, `robust` rows | Tables before interpretation 1/5; every other check 5/5 |
| Precise, tables first, Opus 5.5 | O1 + O4 below, `robust-tables-first` rows | Every check 5/5 when passed counts and trial counts are added across the 1+4 batches |
| Precise, tables first, Sonnet 5.5 | N5 below, `robust-tables-first` rows | Both Delta charges kept 4/5; August Travel 4/5; every other check 5/5 |

C1 and C2 mark both `standard` and `robust` rows `is_gated: false`. Their rendered scorecards explicitly say **NOT EVALUATED (no gated variant was run)**. Their JSON `passes_gate: false` therefore must not be described as a failed overall gate. The article's comparison identifies checks below the 80% minimum; the actual suite gate applies to `robust-tables-first`, shown below.

## Table 2: six models on the final variant

All six use medium effort, with the assistant recorded in each scorecard. All rows below concern `robust-tables-first`.

| Article model | Exact evidence | Verified result |
| --- | --- | --- |
| Claude Opus 5.5, Claude Code | [O1 scorecard](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/results/pwi/expense_analysis/20261010T105946Z-regrade-claude_code-claude-opus-5-5-medium-c0bed1/scorecard.md) ([JSON](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/results/pwi/expense_analysis/20261010T105946Z-regrade-claude_code-claude-opus-5-5-medium-c0bed1/scorecard.json)) + [O4 scorecard](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/results/pwi/expense_analysis/20261010T110008Z-regrade-claude_code-claude-opus-5-5-medium-0b6fd3/scorecard.md) ([JSON](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/results/pwi/expense_analysis/20261010T110008Z-regrade-claude_code-claude-opus-5-5-medium-0b6fd3/scorecard.json)) | 1+4 distinct trials; every check totals 5/5; both saved batches pass |
| GPT-6.1 Sol, Codex | [S1 scorecard](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/results/pwi/expense_analysis/20261010T105946Z-regrade-codex-gpt-6.1-sol-medium-83daba/scorecard.md) ([JSON](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/results/pwi/expense_analysis/20261010T105946Z-regrade-codex-gpt-6.1-sol-medium-83daba/scorecard.json)) + [S4 scorecard](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/results/pwi/expense_analysis/20261010T110007Z-regrade-codex-gpt-6.1-sol-medium-86e1d4/scorecard.md) ([JSON](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/results/pwi/expense_analysis/20261010T110007Z-regrade-codex-gpt-6.1-sol-medium-86e1d4/scorecard.json)) | 1+4 distinct trials; every check totals 5/5; both saved batches pass |
| GPT-6 Astra, Codex | [A5 scorecard](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/results/pwi/expense_analysis/20261010T110051Z-regrade-codex-gpt-6-astra-medium-d55b00/scorecard.md) ([JSON](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/results/pwi/expense_analysis/20261010T110051Z-regrade-codex-gpt-6-astra-medium-d55b00/scorecard.json)) | Five trials; every check 5/5; passes |
| Claude Sonnet 5.5, Claude Code | [N5 scorecard](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/results/pwi/expense_analysis/20261010T110057Z-regrade-claude_code-claude-sonnet-5-5-medium-1892c7/scorecard.md) ([JSON](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/results/pwi/expense_analysis/20261010T110057Z-regrade-claude_code-claude-sonnet-5-5-medium-1892c7/scorecard.json)) | Five trials; Delta retention 4/5 and August Travel 4/5; all others 5/5; passes |
| Claude Haiku 5.5, Claude Code | [H5 scorecard](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/results/pwi/expense_analysis/20261010T110145Z-regrade-claude_code-claude-haiku-5-5-medium-cd428f/scorecard.md) ([JSON](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/results/pwi/expense_analysis/20261010T110145Z-regrade-claude_code-claude-haiku-5-5-medium-cd428f/scorecard.json)) | Five trials; uncategorized count 2/5; August Travel 3/5; July Shopping, largest increase and three findings each 4/5; fails |
| GPT-6 Luna, Codex | [L5 scorecard](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/results/pwi/expense_analysis/20261010T110135Z-regrade-codex-gpt-6-luna-medium-43f5c3/scorecard.md) ([JSON](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/results/pwi/expense_analysis/20261010T110135Z-regrade-codex-gpt-6-luna-medium-43f5c3/scorecard.json)) | Five trials; July Shopping, August Travel and three findings each 0/5; June Dining, Delta retention and largest increase each 2/5; uncategorized count 4/5; fails |

Each selected scorecard has 16 check rows per variant, with an 80% minimum. For the 1+4 combinations, add `passed_count` and `trial_count` separately for each matching check; do not average batch percentages. O1 + O4 contain five distinct Claude `session_id` values. S1 + S4 contain five distinct Codex `thread_id` values. This confirms that the combinations do not count the same saved session twice; it does not independently establish their submitted prompt revision.

Sonnet's two imperfect checks are both in [N5 run-03 grade](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/results/pwi/expense_analysis/20261010T110057Z-regrade-claude_code-claude-sonnet-5-5-medium-1892c7/trials/robust-tables-first/june-august-2026/run-03/grade.json); the [same run's answer](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/results/pwi/expense_analysis/20261010T110057Z-regrade-claude_code-claude-sonnet-5-5-medium-1892c7/trials/robust-tables-first/june-august-2026/run-03/answer.md) reports August Travel as $1,219.22.

## Prompt source files and what their provenance establishes

The [committed suite definition](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/evals/expense_analysis/eval_suite.json) maps these variant names to these exact files:

| Variant | Immutable prompt source |
| --- | --- |
| `standard` | [standard.md](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/evals/expense_analysis/prompts/standard.md) |
| `robust` | [robust.md](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/evals/expense_analysis/prompts/robust.md) |
| `robust-tables-first` | [robust_tables_first.md](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/evals/expense_analysis/prompts/robust_tables_first.md) |

Git history introduces all three files in [`d3d7d273aa74f92fe783ae3fd65300a12377b132`](https://github.com/matlus/Skills-and-Evals/commit/d3d7d273aa74f92fe783ae3fd65300a12377b132). The first two are the committed standard and more precise prompt definitions. The third explicitly requests tables first, exactly three findings, verification of numbers/comparisons against tables, and restriction to the two stated budget targets.

**Boundary:** The inspected saved result folders contain scorecards, `regraded_from.txt`, trial answers, event streams and grades. They contain no per-run prompt snapshot, prompt hash, suite snapshot or source revision manifest. The session event streams examined begin with session/thread metadata and model activity, rather than the complete initial user prompt. The scorecard's variant name identifies the label used, but cannot by itself prove which text occupied that prompt file at invocation. Therefore this index can verify the saved counts and point to the committed prompt definitions; it cannot independently certify a run-to-prompt-revision binding or reconstruct earlier prompt variants.

The need for exact batch links is concrete: [another saved Opus robust-tables-first scorecard](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/results/pwi/expense_analysis/20261010T112630Z-regrade-claude_code-claude-opus-5-5-medium-39a815/scorecard.md) reports four trials and findings 2/4. It is not one of O1/O4, and should not be substituted for them merely because its variant name matches or its folder name sorts later. No claim about its earlier/later prompt text is established here.

Reference and grading context remain inspectable in the [expected facts](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/evals/expense_analysis/cases/june_august_2026/expected_facts.json), [human reference key](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/evals/expense_analysis/cases/june_august_2026/expense_reference_key.md), and [three-findings rubric](https://github.com/matlus/Skills-and-Evals/blob/5ec0b0a9bd4c9361cb059984675cd2e62352c699/evals/expense_analysis/grading/rubrics/three_supported_findings.md). The linked current versions accept Steam in Shopping. They do not restore original-key judge verdicts that were replaced during regrading.