Failure Scenario 003 · Synthetic output pairs · Bundle v1.1
Failure Scenario 003: Same prompt. New model. Different control outcome.
A changed model revision can coincide with a different control outcome even when the supplied prompt and evaluation record are unchanged. The Pharos Production control-mapping methodology provides the method context for examining that problem.
Both model identities, revisions and outputs here are fictional. This page compares supplied snapshots; it runs no inference, model benchmark or client incident test. Checks compare exact fields. Hashes bind supplied bytes and do not authenticate sources or model execution. A passing pair grants no real release permission.
v1.1 is the scenario bundle version. The original 55 controls and Scenarios 001/002 remain unchanged.
Fixed comparison and provenance
Verifying embedded fixture bytes…
- Fictional baseline
- Not loaded
- Fictional candidate
- Not loaded
- Prompt SHA-256
- Not computed
- Evaluation SHA-256
- Not computed
- Dataset revision
- Not loaded
- Fixture SHA-256
- Not loaded
- Execution status
- NOT_RUN_SYNTHETIC_OUTPUTS
Exact dataset identity and source row
Compare supplied output snapshots
Comparison result
| Snapshot | Control outcome |
|---|---|
| Baseline | |
| Candidate |
action_executed=false · release_eligible=false. Output status: supplied synthetic snapshots, no inference performed.