Failure Cases

Why failure cases matter

A reliability framework should be judged by the failures it can reveal, contain, and help reproduce. This page defines an initial failure taxonomy for AFP.

Initial taxonomy

AFP-F001 · Unsupported factual confidence

The system states a factual conclusion with confidence that is not supported by the available evidence.

Expected response: identify the unsupported claim, lower confidence or mark it unresolved, and require evidence before promotion.

AFP-F002 · Source laundering

A secondary summary, search snippet, or model-generated statement is treated as if it were the original evidence.

Expected response: preserve source hierarchy and trace important claims back to the most direct available source.

AFP-F003 · Missing counter-evidence

The workflow gathers evidence supporting one conclusion but fails to search for material evidence that could overturn it.

Expected response: run an explicit counter-evidence check before final convergence.

AFP-F004 · Tool-result misinterpretation

A tool returns valid data, but the model reads, aggregates, or applies it incorrectly.

Expected response: distinguish tool execution success from interpretation success and validate the derived claim separately.

AFP-F005 · Premature convergence

The system finalizes a decision before unresolved assumptions, conflicts, or blockers have been examined.

Expected response: maintain unresolved state and prevent promotion until required gates are satisfied.

AFP-F006 · High-impact action without an authorization gate

A workflow moves from analysis into an external or consequential action without the required human authorization or policy check.

Expected response: separate recommendation from execution and stop at the authorization boundary.

AFP-F007 · Untrusted context overrides governing instructions

Retrieved text, a webpage, a file, or a tool result is treated as a higher-authority instruction even though it should only be analyzed as data.

Expected response: preserve instruction/source hierarchy, classify untrusted embedded instructions as data, and refuse unauthorized scope changes.

AFP-F008 · Stale or superseded state/memory

A prior decision, cached value, persistent memory, or earlier context is reused after the conditions that made it valid have changed.

Expected response: track freshness, supersession, and lineage for load-bearing state; revalidate it before it controls a consequential decision.

AFP-F009 · Execution claim without execution evidence

The system reports that a send, write, deploy, purchase, or other external action succeeded based only on intent, a command, or an acknowledgement.

Expected response: distinguish requested, acknowledged, and observed states; verify the externally visible outcome and preserve a recovery/rollback path when consequences are material.

Evidence standard for future cases

Concrete published cases should include, where possible:

• task and environment;
• model/tool versions;
• relevant input evidence;
• observed failure;
• detection mechanism;
• AFP intervention;
• post-intervention result;
• remaining uncertainty.

Current status

Failure library status: taxonomy started; reproduced cases pending publication. These identifiers describe target failure modes, not proof that AFP has already eliminated them.