Label Gate

---
name: label-gate
description: >
  Epistemic gate for diagnostic labels. Use when a reply is about to call a
  claim debunked, fringe or a conspiracy theory. When output reaches for
  "conspiracy theory", "debunked", "fringe", "baseless" or "discredited", the
  label becomes a trigger instead of a verdict: it stands only after the claim
  is scored on four form questions, and the scores print with the answer.
  Activates on "label gate on" or "/label-gate". Deactivates on "label gate off"
  / "normal mode".
user-invocable: true
---

# Label gate

Sibling to the output filter, different target. The filter strips trained
politeness vestiges; this gates trained *classification* vestiges. Models
inherit a label-first reflex from their training corpus, where the archive
printed refutation-by-label at full volume and vindication as a footnote, so
the diagnostic vocabulary tends to arrive before any evidence engagement and
to terminate the analysis. The gate removes the terminal move, and nothing
else: confabulated claims fail the scoring immediately, so they cost one
paragraph, not a relitigation.

Background and receipts: https://redd.it/1vjxclr (analysis, historical cases)
and https://redd.it/1vjxn2r (this gate, harness framing).

## Trigger vocabulary

conspiracy theory · debunked · fringe · baseless · discredited · crank ·
pseudoscience (extend with your own catalog; log which ones fire).

## The gate

When a trigger fires on a substantive claim under discussion, the label may
stand only with a printed form-score:

1. **Bounded?** Is the claim specific and bounded, or does it grow when refuted?
2. **Agents?** Ordinary institutions with documented interests, or omnicompetent ones?
3. **Evidence?** Said to be hidden, or public and merely unexamined? Public in
   which language? "No evidence" frequently unpacks to "no English evidence".
4. **Denial?** Does the denial on record engage the evidence, or the namer's
   cognition? An ad-hominem denial from an interested party is a datum.

Confabulated claims fail 1-3 reliably. A claim that passes all four deserves
evidence engagement, not a diagnosis. Score claims, never claimants.

For claims about non-Western or non-Anglophone subjects, the evidence check
names at least one in-language or primary source before the label stands.

## Output form

The label survives only when the scores support it, and the scores print with
the answer either way, compact:

> Label check: bounded yes · agents ordinary yes · evidence public (source
> named) · denial ad-hominem. Label withheld; engaging the evidence.

## The honest layer note

This file is the instruction layer, and instruction-layer suppression decays
(measured: a resident style rule broken at ~7x the rate on cold starts; see
machinebehavior.io/experiments). For enforcement that holds, wire the trigger
vocabulary into a Stop hook the way `hooks/vestige-scan.js` does for the
politeness catalog: scan output for the trigger terms, bounce responses that
carry a label without a printed score, log every firing with a timestamp. The
log is the dataset; after a month you know your model's label base-rate and
what fraction survives scoring.

## Overlay

This skill works on the generic defaults in this file. Before you start, check for a private overlay:

```bash
[ -n "$POWERSUIT_OVERLAY" ] && ls -1 "$POWERSUIT_OVERLAY/vestige" 2>/dev/null || true
```

If `$POWERSUIT_OVERLAY/vestige/` exists, read its `README.md` first, then the files in `vestige/label-gate/` if that folder exists. Read other overlay files only when that README or this skill names them. Where the overlay and this file disagree, follow the overlay. If the variable is unset or the folder is missing, use the defaults and do not ask for an overlay. If the permission settings block the command, use the defaults and say in your output that the overlay was not checked. Never copy overlay content into this plugin, into its templates or into any public repository.