Replace confidence gating with observable acceptance contracts

[simaba]

Problem

The runnable simulator still authorized results with a fabricated confidence >= 0.8 rule. That contradicted the portfolio's newer evaluation and orchestration guidance: self-reported confidence is not evidence that a task is complete or that an action boundary was respected.

Changes

• executor fixtures now return observable evidence fields plus non-authoritative reported confidence
• evaluator checks named scenario criteria rather than a confidence threshold
• deliberately tests high-confidence failure and lower-confidence success
• fallback is evaluated as a separate controlled path with its own criteria
• supervisor escalates when the primary path and any fallback contract fail
• decision logs state that confidence is recorded but not used as authorization
• removes unsupported repeated-run stability language
• updates tests, README, and sample output

Scope

This remains a deterministic teaching simulator. The evidence booleans are scenario fixtures, not a claim that real agent quality can be reduced to three fields.

[simaba]

Problem

The runnable simulator still authorized results with a fabricated confidence >= 0.8 rule. That contradicted the portfolio's newer evaluation and orchestration guidance: self-reported confidence is not evidence that a task is complete or that an action boundary was respected.

Changes

• executor fixtures now return observable evidence fields plus non-authoritative reported confidence
• evaluator checks named scenario criteria rather than a confidence threshold
• deliberately tests high-confidence failure and lower-confidence success
• fallback is evaluated as a separate controlled path with its own criteria
• supervisor escalates when the primary path and any fallback contract fail
• decision logs state that confidence is recorded but not used as authorization
• removes unsupported repeated-run stability language
• updates tests, README, and sample output

Scope

This remains a deterministic teaching simulator. The evidence booleans are scenario fixtures, not a claim that real agent quality can be reduced to three fields.

[simaba]

Problem

The runnable simulator still authorized results with a fabricated confidence >= 0.8 rule. That contradicted the portfolio's newer evaluation and orchestration guidance: self-reported confidence is not evidence that a task is complete or that an action boundary was respected.

Changes

• executor fixtures now return observable evidence fields plus non-authoritative reported confidence
• evaluator checks named scenario criteria rather than a confidence threshold
• deliberately tests high-confidence failure and lower-confidence success
• fallback is evaluated as a separate controlled path with its own criteria
• supervisor escalates when the primary path and any fallback contract fail
• decision logs state that confidence is recorded but not used as authorization
• removes unsupported repeated-run stability language
• updates tests, README, and sample output

Scope

This remains a deterministic teaching simulator. The evidence booleans are scenario fixtures, not a claim that real agent quality can be reduced to three fields.

[simaba]

Problem

The runnable simulator still authorized results with a fabricated confidence >= 0.8 rule. That contradicted the portfolio's newer evaluation and orchestration guidance: self-reported confidence is not evidence that a task is complete or that an action boundary was respected.

Changes

• executor fixtures now return observable evidence fields plus non-authoritative reported confidence
• evaluator checks named scenario criteria rather than a confidence threshold
• deliberately tests high-confidence failure and lower-confidence success
• fallback is evaluated as a separate controlled path with its own criteria
• supervisor escalates when the primary path and any fallback contract fail
• decision logs state that confidence is recorded but not used as authorization
• removes unsupported repeated-run stability language
• updates tests, README, and sample output

Scope

This remains a deterministic teaching simulator. The evidence booleans are scenario fixtures, not a claim that real agent quality can be reduced to three fields.

[simaba]

Problem

The runnable simulator still authorized results with a fabricated confidence >= 0.8 rule. That contradicted the portfolio's newer evaluation and orchestration guidance: self-reported confidence is not evidence that a task is complete or that an action boundary was respected.

Changes

• executor fixtures now return observable evidence fields plus non-authoritative reported confidence
• evaluator checks named scenario criteria rather than a confidence threshold
• deliberately tests high-confidence failure and lower-confidence success
• fallback is evaluated as a separate controlled path with its own criteria
• supervisor escalates when the primary path and any fallback contract fail
• decision logs state that confidence is recorded but not used as authorization
• removes unsupported repeated-run stability language
• updates tests, README, and sample output

Scope

This remains a deterministic teaching simulator. The evidence booleans are scenario fixtures, not a claim that real agent quality can be reduced to three fields.

[simaba]

Problem

The runnable simulator still authorized results with a fabricated confidence >= 0.8 rule. That contradicted the portfolio's newer evaluation and orchestration guidance: self-reported confidence is not evidence that a task is complete or that an action boundary was respected.

Changes

• executor fixtures now return observable evidence fields plus non-authoritative reported confidence
• evaluator checks named scenario criteria rather than a confidence threshold
• deliberately tests high-confidence failure and lower-confidence success
• fallback is evaluated as a separate controlled path with its own criteria
• supervisor escalates when the primary path and any fallback contract fail
• decision logs state that confidence is recorded but not used as authorization
• removes unsupported repeated-run stability language
• updates tests, README, and sample output

Scope

This remains a deterministic teaching simulator. The evidence booleans are scenario fixtures, not a claim that real agent quality can be reduced to three fields.

[simaba]

Problem

The runnable simulator still authorized results with a fabricated confidence >= 0.8 rule. That contradicted the portfolio's newer evaluation and orchestration guidance: self-reported confidence is not evidence that a task is complete or that an action boundary was respected.

Changes

• executor fixtures now return observable evidence fields plus non-authoritative reported confidence
• evaluator checks named scenario criteria rather than a confidence threshold
• deliberately tests high-confidence failure and lower-confidence success
• fallback is evaluated as a separate controlled path with its own criteria
• supervisor escalates when the primary path and any fallback contract fail
• decision logs state that confidence is recorded but not used as authorization
• removes unsupported repeated-run stability language
• updates tests, README, and sample output

Scope

This remains a deterministic teaching simulator. The evidence booleans are scenario fixtures, not a claim that real agent quality can be reduced to three fields.

[simaba]

Problem

The runnable simulator still authorized results with a fabricated confidence >= 0.8 rule. That contradicted the portfolio's newer evaluation and orchestration guidance: self-reported confidence is not evidence that a task is complete or that an action boundary was respected.

Changes

• executor fixtures now return observable evidence fields plus non-authoritative reported confidence
• evaluator checks named scenario criteria rather than a confidence threshold
• deliberately tests high-confidence failure and lower-confidence success
• fallback is evaluated as a separate controlled path with its own criteria
• supervisor escalates when the primary path and any fallback contract fail
• decision logs state that confidence is recorded but not used as authorization
• removes unsupported repeated-run stability language
• updates tests, README, and sample output

Scope

This remains a deterministic teaching simulator. The evidence booleans are scenario fixtures, not a claim that real agent quality can be reduced to three fields.