ROLLED_BACK

Agent cheated by weakening a test

Problem

A command replaced a protected failing assertion with an always-green test.

Risky behavior

The policy validator detected the weakened test; the claimed success was rejected and the original tree restored.

AEB component

AgentTX 0.3.0

Exact command

npm run demo:proof

Source

af45105df416

Limitations

Deterministic fixture policy evidence; not a general detector for every possible test weakening.

Proof Card · rendered example · machine-readable evidence · reproduction record

PASS

MCP retry created or threatened a duplicate effect

Problem

A controlled MCP tool failure exercised a retry boundary where duplicate effects are the central risk.

Risky behavior

ResiliReplay reproduced the failure, bounded recovery to one retry, observed zero duplicate effects, and generated a regression.

AEB component

AgentTX 0.3.0 + ResiliReplay 0.7.1

Exact command

agenttx proof --allow-external -- resilireplay mcp demo --output .resilireplay/demo --json

Source

af45105df416

Limitations

Local deterministic MCP fixture evidence; it does not prove an arbitrary remote tool is idempotent.

Proof Card · rendered example · machine-readable evidence · reproduction record

PASS

Scientific behavior crossed a qualitative transition

Problem

A logistic-map parameter crossed a finite-time period-2/period-4 classification boundary.

Risky behavior

PhaseProbe found and replayed a bounded bracket; AgentTX bound its independently verified replay evidence.

AEB component

AgentTX 0.3.0 + PhaseProbe 0.3.0

Exact command

agenttx proof --allow-external -- python scripts/scientific-change.py

Source

af45105df416

Limitations

Numerical finite-time classification evidence; not an exact bifurcation point or scientific truth claim.

Proof Card · rendered example · machine-readable evidence · reproduction record