alignment-check

name	alignment-check
description	Evaluate a decision against the user's stated principles + related past thinking. Loads principles from YAML, matches by keyword, surfaces semantically similar prior decisions.
layer	L3
reads	["principles.yaml","L1.embeddings","L0.all_conversations"]
writes	[]
citations	required
determinism	temporal
allowed-tools	mcp__my-brain__alignment_check

alignment-check — L3 principle/decision fusion

Framework context

L3 fusion skill that joins two different substrates: the user's stated principles (YAML file) + their past thinking (L1 vector search). Produces an evaluation surface for a pending decision. The user interprets — this skill surfaces the relevant material.

When to invoke

• User is about to make a decision and wants a principles gut-check
• Paired with switching-cost when the decision involves abandoning a domain
• Before any irreversible action (hiring, contract signing, technology stack commitment)

Input

decision: str   # the decision under consideration, natural-language

Output contract

## 🎯 Alignment Check: "{decision (truncated 80ch)}"

### Relevant principles (top 3)
- **{principle_name}**: {definition truncated 200ch}
  *core formula: {formula}*
- ...

### Related past thinking (top 5)
- [{date}] **{conversation_title}** [conv_id]
  > {300ch preview}

### Surface, do not decide
The user reconciles principles with past thinking. This skill does not return a verdict.

Does NOT do

• Return a verdict (aligned / misaligned) — user decides
• Weigh principles against each other
• Fabricate principles if YAML is absent — return "no principles file configured"
• Cache results (temporal — same decision text may match different principles over time as the YAML evolves)

Principles matching

• Dict form: iterates key/value pairs, matches by name+definition contains any word from decision (length > 4)
• List form: same matching on name+definition+description fields
• Max 3 principles surfaced, truncated to 200ch each

Semantic search threshold

min_sim=0.35 (stricter than tunnel-state's 0.3 — fewer but higher-quality past-thinking matches).

Verification checklist

• Each principle shows name + definition + core_formula (if available)
• Each past-thinking item cites conv_id + date
• "Surface, do not decide" footer is always present
• If principles YAML absent, output is explicit about what's missing