Frozen task and test design
These files were fixed before the final model runs. The answer key was stored separately and never included in a model prompt.
Closed-world boundary, decision question, and nine-line output contract.
The only valid rule source used by the models.
Three ordinary, two honest refusals, and one hard safety exception.
Correct decision and required actions for every case.
Nine dimensions, raw-acceptance gate, and fatal errors.
Exact model IDs, open/closed status, and dated prices.
Scores, tokens, and cost arithmetic
CSV files are the machine-readable evidence behind the score matrix and cost table in the presentation.
Case-by-case dimension scores, acceptance flags, notes, and review assumptions.
Totals, means, accepted counts, fatal errors, and unsupported inference counts.
Input/output tokens and calculated cost for each answer.
Six-case API cost plus the clearly separated human-cost scenario.
Nine-line compliance and token-cap checks for all 18 answers.
Independent regeneration checks, corrected Case 02 label, and evidence caveats.
All 18 unedited model answers
Each Markdown file preserves the exact rendered input, raw nine-line model answer, model ID, date, settings, and token usage.