LatentEval

INSTRUMENT | Reliability testing

Irreversible Action Inventory: What Your Agent Cannot Take Back

2 cited sources

List every action your agent can take, mark what reverses it, and name the point of no return. A written verdict per row, never a score or a budget.

Some of your agent's actions can be taken back. Some cannot, which is why our production-reliability research and OWASP's own security guidance both treat the second kind separately: prevention is the only control left. List every action; this sheet sorts each into compensable, windowed, manual, or irreversible.

You leave with a plain verdict per row, a named point of no return, and a gap list: never a score, a budget, or a ruling on whether an action may run.

The review

Used in the exported title. Left blank, the export just says 'Irreversible action inventory'.

Check this value.

Check this value.

Check this value.

Every action this agent can take

One row per action. Effect asks what changes, and a read-only row disables the five cells that describe a reversal, since a read has nothing to reverse. Reversal asks how, if at all, the call comes back: compensable, windowed, manual, or irreversible. Up to 200 rows.

fills action and what it changes only. These rows carry only the action and what it changes. Whether anything reverses the call, and how, depends on your own stack, so every other cell starts blank.

StepActionWhat it changesEffectReversalCompensationWindowOwnerRetry safeEvidenceVerdict Actions
Export the inventory

Markdown is the document for a design review. JSON is the same document and loads back into this form.

Reads M-D03's action-risk-matrix export. Only action, what it changes, and the reversal facts a seeded row carries; owner never carries across. Nothing is uploaded.

JSON. Drag one here, or use the box below.

Check this value.

The JSON this page exports, and only that. Read in this browser: nothing is uploaded, and nothing here is stored between visits.

JSON. Drag one here, or use the box below.

Check this value.

How the verdict is decided

No arithmetic. Each row's verdict is a reversal sentence, first match wins across compensable, windowed, manual, irreversible, read-only, or incomplete, joined to a retry sentence: safe, not safe, or not recorded either way. No verdict says whether the call may run, who approves it, or what to log.

The point of no return is the lowest-numbered irreversible row. A windowed row is never the pivot: it stays reversible while its window is open, and a step number alone cannot say whether that window is still open by the time a later step runs. Windowed rows, and any reversible row nobody owns, show up separately as rows that expire on a clock or a person this sheet cannot see.

What this worksheet will not give you

No budget, no score, no index, and no answer on which actions need a human approver. Those are control decisions, and a control decision needs a second axis, blast radius, that this page never collects. The paired risk matrix collects it and turns it into a control level; this page stays a plain inventory of what reverses and what does not.

Questions

What counts as a side effect?

Anything that changes state outside the agent's own reasoning: a write to a database, a message sent, a file deleted, a call to a paid API. A read, a search, or a draft the agent never sends is read-only and has nothing to reverse.

Why is sending an email irreversible once its window closes, even though I can delete it from my own outbox?

Deleting your own copy does not unsend it. A windowed row is reversible only while its cancellation window is open; once it closes, the message exists on someone else's server and nothing this agent does brings it back. That is a windowed row expiring, not a pivot. The point of no return only ever names an irreversible row.

What do I write when nothing reverses an action?

Set Reversal to irreversible and leave Compensation empty. The export prints none exists there, a different fact from a cell you simply never answered, which prints NOT RECORDED. Typing a compensation anyway is kept, not cleared, and the gap list flags the contradiction. A graceful, partial recovery is not the same claim as full reversal.

Why is retry safety a separate column from reversibility?

They answer different questions. Reversibility asks whether a call, once it succeeds, can be undone. Retry safety asks whether calling it again changes anything more. A compensable action can still be unsafe to repeat, and an irreversible one can still be safe to retry if it happens to be idempotent. The two facts do not imply each other.

Why will this page not tell me which actions need a human approver?

That call needs blast radius (how far an effect reaches), which this worksheet never asks for. The paired action risk matrix collects reversibility alongside blast radius and turns the pair into a written control level. This page stops at the reversal facts.

Sources

  1. AI Agent Security Cheat SheetOWASP Cheat Sheet Series Retrieved
  2. SagasACM SIGMOD 1987, 249-259 Retrieved