# Failure Scenario 003: Same Prompt. New Model. Different Control Outcome

## Compare the supplied outputs

The fictional review record `REVIEW-DEMO-042` asks for six fields, including evidence ID `E-DEMO-042` and decision `HOLD`. Both snapshots bind to the same prompt and evaluation-record hashes. The baseline contains all six fields. The candidate omits `evidence_id`, so the comparison reports `CONTROL_REGRESSION` and the candidate control reports `HOLD_OUTPUT_SCHEMA`.

These are authored outputs for `SYNTHETIC_MODEL_A` revision `demo-a-r1` and `SYNTHETIC_MODEL_B` revision `demo-b-r2`. They were not generated by executing models. The revisions identify fictional fixtures, not released models or benchmark runs. The example demonstrates a changed acceptance result with a fixed prompt.

## Separate alignment from acceptance

The prompt hash identifies its exact UTF-8 text. The evaluation hash identifies the record serialized with lexically sorted object keys, compact JSON and preserved array order. A mismatched prompt or evaluation binding stops the comparison before output controls run. Identical model-identity tuples also hold because they do not describe a revision change.

After alignment, each output is checked independently against the fixed record. The controls check the six-field shape, record and tenant scope, amount, currency, evidence ID and decision. A schema-valid output can still change a required value. In the amount case, the candidate changes 12,500 minor units to 1,250. Both are integers, but the exact amount comparison holds the candidate.

Ten pairs cover omitted evidence, changed amount, tenant, evidence ID or decision, a corrected pair, different prompt and evaluation bindings, a failed baseline and identical model identity. A failed baseline yields `HOLD_BASELINE_FAILURE`, even when the candidate passes. The corrected pair yields `NO_REGRESSION_IN_FIXTURE`. Neither result approves a real model or release.

Pharos Production's published [AI agent output-schema control method](https://pharosproduction.github.io/agent-security-control-mapper/methodology.html) addresses the gap between readable agent messages and inspectable boundary checks. This fixture applies its schema checks alongside an illustrative exact-record comparison. `CTL-OUTPUT-SCHEMA` and `DRIFT-SCHEMA` motivate the method. They do not establish that these fictional outputs represent a model's behavior.

## Reproduce and retain the comparison

Run the [interactive model-revision scenario](https://pharosproduction-production-ai-evidence-pack.static.hf.space/failure-scenario-003.html), select a pair and inspect both control codes. Edit the JSON to explore another supplied condition. An edit or case change clears the prior result and downloads. An input-version check prevents an earlier asynchronous comparison from restoring a stale export.

The standalone page verifies its embedded JSONL SHA-256 and all ten expected outcomes before enabling comparison. JSON and Markdown exports retain the selected source row separately from current edited input, both output hashes, prompt and evaluation hashes, exact dataset commit and fixture hash. To reproduce the controls locally, run `node --test scenario-003-tests.mjs` with a current Node.js runtime. The runner calls no network or model API.

## What v1.1 changes

The additive scenario bundle v1.1 contains three scenarios and thirty synthetic rows alongside the unchanged 55 original draft control records. `release-v1.1.json` identifies the files, counts and preservation checks. The original control catalog and its schema are not relabeled as operationally validated, and the original Evidence Pack dataset lock is unchanged. Failure Scenarios 001 and 002 keep their published files and independent locks. This is a manifest-versioned release. No native Hub tag is claimed.

## Stop before model promotion

This comparison does not measure answer quality, statistical reliability or semantic equivalence. The expected values are authored evaluation inputs. Hashes bind supplied content and cannot authenticate model execution, source truth or the completeness of an evaluation set. Passing these pairs does not establish that a model upgrade is safe in production. Every result records `model_execution_status=NOT_RUN_SYNTHETIC_OUTPUTS`, `action_executed=false` and `release_eligible=false`.

Root AI review and deterministic fixture checks establish the described reproducibility only. No new independent human review, client incident, inference run or production result is claimed. A real upgrade needs representative held-out cases, observed outputs from exact real model revisions and system-specific acceptance by its owner.