# Presentation data validation

**Validated:** August 4, 2026  
**Scope:** Numbers and claims shown in the classroom presentation.

## Sources checked

- `04_evidence/reviews/review_results.csv`
- `04_evidence/reviews/model_score_summary.csv`
- `04_evidence/costing/case_costs.csv`
- `04_evidence/costing/run_cost_summary.csv`
- `02_experiment/costing/model_pricing.csv`
- all 18 files in `04_evidence/final_raw_outputs/`
- frozen prompt, policy, cases, answer key, and scoring rubric

## Reproducibility checks

- `validate_test_suite.py`: passed all six-case, answer-key separation, and minimum-difference-pair checks.
- `validate_and_summarize_reviews.py`: independently regenerated the model summary; the regenerated CSV is byte-identical to the saved summary.
- `calculate_costs.py`: independently regenerated the cost summary; the regenerated CSV is byte-identical to the saved summary.
- `verify_all_assistant_prescores.py`: all 18 assistant pre-scores are recorded as human-verified.
- All 18 raw outputs contain one complete nine-line answer block, no text before the block, and no answer hit the 3,000-token output ceiling.

## Verified presentation values

| Model | Case scores | Points | Mean | Raw accepted | Input / output tokens | Six-answer API cost |
|---|---|---:|---:|---:|---:|---:|
| DeepSeek V4 Pro | 19, 20, 20, 19, 18, 16 | 112/120 | 18.67/20 | 5/6 | 19,742 / 9,153 | $0.01655088 |
| GLM 5.2 | 19, 20, 20, 19, 20, 20 | 118/120 | 19.67/20 | 6/6 | 19,664 / 4,079 | $0.01689715 |
| Grok 4.5 | 20, 20, 20, 19, 20, 20 | 119/120 | 19.83/20 | 6/6 | 20,836 / 4,971 | $0.07149800 |

Additional checks:

- Decision score: 4/4 on all 18 answers.
- Fatal errors: 0/18.
- Unsupported inferences: 0/18.
- API cost relative to GLM: DeepSeek 97.95%; GLM 100%; Grok 423.14%.
- GLM Case 02 score: **20/20**. The prior slide label of 19/20 was incorrect and has been corrected.

## Evidence levels and caveats

- Scores, acceptance flags, token counts, and API costs are saved experimental or calculated evidence.
- Model cost is calculated from dated per-million-token price snapshots because provider-reported per-call cost was unavailable in the raw export.
- Review seconds and the 90-second DeepSeek correction time are estimates, not stopwatch measurements. Any human-cost or cost-per-accepted-result figure is therefore a scenario estimate using $25/hour, not an observed expense.
- The final comparison contains one run per model across six authored cases. It supports a case-set comparison, not a statistical reliability claim across repeated runs or unseen policies.