Skip to content

SKILL.md

Latest commit

 

History

History
41 lines (30 loc) · 2.31 KB

SKILL.md

File metadata and controls

41 lines (30 loc) · 2.31 KB
Raw
Error in user YAML: (<unknown>): mapping values are not allowed in this context at line 2 column 403
---
name: glossary-alignment-check
description: For each technical term in a substacker draft's claims, checks shared-context/glossary.md for a writer-specific definition and compares to field-standard. Emits a note for terms where the two diverge (aligned / diverged-safe / diverged-risky). Feeds into classify-claim (which definition governs) and write-review-artifact's Glossary Alignment section. Use on every claim. Trigger keywords: glossary alignment, term definition, writer vs field, load-bearing term.
---

Glossary Alignment Check

Workflow

Per claim:
- [ ] Step 1: Tokenize claim for candidate terms (named concepts, TLAs, math)
- [ ] Step 2: Grep shared-context/glossary.md for each term
- [ ] Step 3: If found, pull writer's definition
- [ ] Step 4: Compare to field-standard (prior knowledge + optional WebSearch)
- [ ] Step 5: Emit: aligned | diverged-safe | diverged-risky

Three outcomes

  • aligned: writer's definition matches field-standard. No note.
  • diverged-safe: definitions differ but the post's argument doesn't depend on the difference. Note but don't block.
  • diverged-risky: definitions differ AND the post's argument hinges on which definition is in force. Tier-2 note for the writer.

Worked example

Term: "attention" Writer's glossary: "the mechanism that lets tokens exchange information, including the residual bypass." Field-standard: "the weighted sum softmax(QK^T/√d)V, excluding the residual."

Status: diverged-risky if post asserts "attention is O(n²)" — true under field definition, false-or-unclear under the writer's.

Output note: "Writer's glossary includes residual in 'attention'; field-standard excludes. Post's claim 'attention is O(n²) in memory' is true under field but misleading under writer's glossary. Consider disambiguating in-post or tightening the glossary."

Guardrails

  1. Never rewrite the glossary. Only flag.
  2. Never mark aligned without actually checking the glossary file this run.
  3. If glossary is missing or empty, emit no-glossary and treat field definition as governing.
  4. When disagreement is strictly semantic (not functional), prefer diverged-safe over diverged-risky — err toward non-blocking.
  5. Diverged-risky notes appear in the Glossary Alignment section of the review artifact; don't bury in claim table.