lyndonkl/skills/hedge-detector

Hedge Detector

Table of Contents

• Precision vs weakness
• Workflow
• Worked example
• Guardrails

Related skills: Called by the Editor in the voice pass. Complements voice-check (which flags "I think" as a don't-list phrase when used as primary hedge). This skill does the finer classification.

Precision vs weakness

Precision hedge (KEEP): scope-naming, sample-size-caveat, specific-uncertainty.

• "I do not know" (full sentence) — writer's signature.
• "I am not claiming…" — explicit scope.
• "On my 3B-param run…" — sample caveat.
• "n=1 may not replicate."
• "I have only tested this on 3 teams."

Epistemic-weakness hedge (FLAG): softens without adding information.

• "I think" (when followed by a claim).
• "Perhaps" (standalone).
• "Arguably" — deniability.
• "It could be argued" — auto-deniability.
• "Somewhat" — weakening adverb.
• "Seems" (when no sensing is happening).
• "It seems clear that" — worst-of-both-worlds.

Workflow

For each hedge in the draft:
- [ ] Step 1: Detect hedge markers (modal verbs + phrase list above)
- [ ] Step 2: Classify as precision or weakness
- [ ] Step 3: For weakness, suggest a commit OR a specific hedge (both, as 2 rewrite options)
- [ ] Step 4: For precision, leave alone (note in the "calibrated hedges kept" count)
- [ ] Step 5: Emit the hedge audit with both lists

Step 2: Classifier

A hedge is precision if paired with specific bounds:

• Sample size named (n=1, 3B model, last 12 weeks)
• Scope named ("in three teams I've worked with")
• Specific uncertainty named ("I have not re-derived the gradient")

Otherwise weakness. Default to weakness when unsure — the writer prefers over-flagging here.

Step 3: Rewrite options

For each weakness hedge:

• Option A (commit): remove hedge, take position. "I think batch size matters" → "Batch size matters."
• Option B (specific): name the uncertainty. "I think batch size matters" → "On this 3-run sweep, batch size moved loss by 0.08."

Both options; writer picks.

Worked example

Draft sentences:

1. "I think RAG beats fine-tuning for most teams."
2. "I do not know whether this holds at 70B — my only test was on a 3B model."
3. "Arguably the attention mask is wrong."
4. "Perhaps fine-tuning is better when you have very specific stylistic requirements."

Classification:

| # | Hedge | Class | Rewrites | |---|---|---|---| | 1 | "I think" | weakness | (a) "RAG beats fine-tuning for most teams." (b) "In the three teams I've worked with, RAG beat fine-tuning." | | 2 | "I do not know" + scope | precision | Keep as-is. | | 3 | "Arguably" | weakness | (a) "The attention mask is wrong." (b) "The attention mask looks wrong to me — I have not re-derived the gradient." | | 4 | "Perhaps" + "very specific" | weakness | (a) "Fine-tuning wins on style." (b) "Fine-tuning wins on style; I have not tested this below 7B." |

Guardrails

1. Never flag precision hedges. They are a voice feature.
2. Never replace "I do not know" — this is the writer's signature phrase.
3. Suggest 2 rewrite options (commit + specific), not 3+.
4. Hedge clusters (≥2 weakness hedges within 50 words) get flagged once, collectively — also surfaced to slop-detector signal S8.
5. Don't flag hedges in quoted text or code fences.
6. If the draft is a reflective essay openly admitting uncertainty as its subject, relax the threshold — flag only the most decorative hedges.

Quick reference

• Input: draft.
• Output: hedge audit — weakness hedges with rewrites, precision hedges kept with a count.
• Signal downstream: cluster count goes to slop-detector S8.