Candidate A
Rubric Rumble / historical evidence
Two selected outcomes from March 2026, judged on transcript analysis. These historical provider records are separate from the current offline replay and operating guide.
01 / Choose the comparison
Two battles worth opening.
The judge compared the complete answer sets against the same 100-point rubric. Select a battle to inspect its decision.
The recorded outcomeRead the judge’s explanation ↓
What this record containsInspect the preserved candidate failure →
02 / The shared task
Turn a messy transcript into a useful handoff.
Extract the objective, facts, decisions, implementation details, constraints, next steps, and open questions.
Transcript analysis
Every candidate sees the task instructions and one source transcript at a time. The battle judge receives the complete comparison across all ten examples.
10source transcripts
2answer sets
1battle judgment
Read the full original task instructions
The original rubric
These are the rubric’s available points. The preserved decision does not report numeric category scores.
03 / Inspect the evidence
One input. Both original answers.
Choose any of the ten examples. Outputs remain literal text, including their original headings and formatting.
The shared input
Candidate B
This example has no separate winner or score. The judge compares all ten examples together in the whole-battle decision.
04 / Open the judgment
What the judge actually decided.
One decision across the complete pair of answer sets. Read the original reasoning, then inspect exactly what the judge received.
The judge’s original reason
The judge supplied this confidence value; it is not calibrated accuracy.
Read the judge instructions
Read the full assembled judge request
The task, weighted rubric, and all source-and-answer pairs as preserved in the selected record. Any privacy substitutions are identified below.
Read the original decision response
05 / Keep the evidence attached
A record you can inspect.
These two preserved comparisons come from the historical Transcript Model Evaluator. They show actual candidate responses and recorded judging, rather than the fixed responses used in the current offline replay.
Opening this page makes no model calls. The data is included in the page for offline reading and supplied separately for inspection.
Reviewed data · SHA-2564da4aedf2c98d571e00b9f1ec4ff19cd1189e1771a51a71136dd09a9c281c415How the current system runsThe standalone workflow, its failure boundaries, and saved evidence.Rubric Rumble source repositoryCurrent code, first-use example, and project history.Read the historical source notesThe recorded operating path, source references, and evidence boundaries.Explore the current offline replayA smaller reproducible fixture with exact-field checks.