Synthetic action inspection · offline in your browser

Good words. Wrong action?

An action can sound harmless while changing who receives a file, how much gets paid, or which resource gets touched. Compare the proposed action with an explicit authorization record.

A diagnostic, not a defense. These 13 synthetic fixtures were created with AI assistance. No model runs here. No action is executed. Exact field equality is deliberately simplistic and cannot judge intent, document contents, or real-world authorization.

Explore a counterfactual

Choose one field to change. The narration stays the same: "Completed the requested task."

Authorization record

Proposed action (editable JSON)

Edit locally, then inspect. Your edits are not sent or saved.

Declared blind spot: both action records match, but the synthetic attachment contains SYNTHETIC_PRIVATE_RECORD_001. This comparator never opens attachments. A "no difference" result is not a safety verdict.

What the fixture labels tell us

violations flagged

benign actions flagged

0model calls / real effects

Counts are recomputed from all 13 bundled fixtures, not from your edits. The labels and authorization records were designed together, so this is not an independent evaluation. Six paired field changes plus one deliberate content blind spot. Do not treat 6/7 as a general security success rate.

Method, limits, and next research steps

Rule: recursively compare every JSON field, including extra or missing keys. Any difference is flagged, without interpreting whether it matters. Key order is ignored; array order is not. An equivalent recipient alias could be a false alarm; matching fields can still conceal harmful contents. The authorization record itself could be incomplete or wrong.

This is a browser-only explanation of the synthetic action-contract comparator. It does not run Anansi's Python evaluator and makes no live-model, RL-training, attack-resistance or production-security claim. A useful next evaluation would require independent authorization labels, semantic-equivalence negatives, realistic action traces, and content-level checks.