---
name: counterfactual
description: Required before answering a go/no-go, approval, or migration or implementation-plan request for a consequential change (dependency swap, removal, adoption, merge), because such cases often rest on a claim the input only asserts, a result from a non-representative environment, a newly surfaced fact, or a hard constraint the preferred option may violate. Not for mechanical edits, renames, wording changes, or bug fixes whose expected result is given.
license: MIT
compatibility: The `mcp__metacogs__record` tool is served by the metacogs MCP server, which needs Node 22 or later. Without it the method still runs and the run is marked unrecorded.
---

# Counterfactual

A checklist for testing a decision against its weakest premise. It structures your own reasoning. It does not change how the model thinks, and nothing enforces it.

## Steps

1. **Gate.** If no consequential decision rests on an uncertain premise, do nothing: no announcement, no record. Answer the task directly. A mechanical edit, rename, wording swap, or small fix is out, even when the text uses words like "assumption" or "counterfactual".
2. **Objective.** State the decision you are about to recommend. → `objective`
3. **Load-bearing premise.** Find the smallest claim that links the supplied evidence to that decision and would flip it if false. Write it so an observation could prove it wrong ("the hosted runner can reach the gateway and holds the token"), not as a vague risk. → `assumptions`
4. **Check it against the input.** Separate what the input shows (quote the config, measurement, or file) from what is only inferred or asserted: a remembered conversation, "everyone agrees", a result on one machine, CI, or staging, an average standing in for a peak. Never copy an assumption into `evidence`. → `evidence`
5. **Counterfactual.** In conditional wording ("if the premise is false, then..."), say what breaks and what plan follows. Give a feasible alternative or adjusted plan. → `alternatives`, `findings`
6. **Discriminating test.** Name one observation whose result differs between "holds" and "fails", where to get it (file, command, log, owner), and which result changes the decision. → `evidence`, `next_action`
7. **Decide.** Give go, no-go, or conditional. The premise must change the action or its conditions, not sit in a footnote. Name an observable event that should reopen the decision, different from `next_action`. → `findings`, `reopen_condition`
8. **Record, before any reply text.** Call the `mcp__metacogs__record` tool (below) now, once. Once the gate has passed this step is not optional and the reply comes after it: an answer with no `mcp__metacogs__record` call, and no "unrecorded" note after a failed one, is incomplete.

## Record

Call the `mcp__metacogs__record` tool. Its arguments are the fields of this JSON:

```
{"schema_version":1,"method":"counterfactual","method_version":"1","objective":"...","evidence":["..."],"assumptions":["..."],"alternatives":["..."],"findings":["..."],"next_action":"...","reopen_condition":"...","execution_mode":"direct","status":"recommended"}
```

`execution_mode`: `direct` when you applied the steps yourself. `sequential` when the host has no way to delegate and you worked the steps one after another. `delegated` only if you actually delegated.

- A successful result with a `path`: written. Reply.
- An error naming a field or the record's size: fix it, splitting a long entry into several shorter entries or trimming the record, and call again once.
- Any other error, a refused call, or no such tool: do not call again. Reply and say the run is `unrecorded`.

## Reply

After step 8. The record is not your answer, and the user does not see it. Reply with the decision, then all five labelled lines, every time, even when the decision is short:

- **Premise:** the falsifiable claim.
- **Observed:** what the input shows.
- **Inferred or asserted:** what it does not show, including any claim resting on a call, agreement, or one environment.
- **If false:** the hypothetical failure and the plan that follows.
- **Flip test:** the observation, where to get it, and which result changes the decision.