icp-grading

name	icp-grading
description	Self-grades the full ICP refinement output against a 100-point rubric covering data foundation, pattern discovery, filter runnability, buyer voice, and decision readiness. Identifies the two weakest dimensions and triggers targeted re-runs of the matching steps. Use after pocket packaging or when iterating an existing ICP output toward A+.

ICP grading + iteration

Run by the icp-grader subagent (independent grader — see agents/icp-grader.md). The orchestrator should not grade its own work.

Inputs

• ./icp-output/ICP.md
• ./icp-output/pockets/*.md
• ./icp-output/personas/*.md
• ./icp-output/triggers/distribution.md
• ./icp-output/enrichment/company-attributes.md
• ./icp-output/priority-matrix.md
• ./icp-output/anti-icp.md
• A sample of ./icp-output/calls/ — to verify provenance of quotes

Grading

Use the rubric in rubric.md. 100 total points across 5 sections. Save the report to ./icp-output/grades/<ISO-timestamp>.md with this structure:

# ICP grade — <timestamp>

**Composite: <score>/100 — <letter grade>**
Iteration: <n>

## Section scores

- A. Data Foundation: <score>/20
- B. Pattern Discovery: <score>/25
- C. Filter Runnability: <score>/20
- D. Buyer Voice: <score>/15
- E. Decision Readiness: <score>/20

## What's working

<bullets>

## Weakest two dimensions

1. **<dimension name>** (<score>/<max>) — <specific evidence: file path + line reference + the issue>. Recommended fix: <specific>.
2. **<dimension name>** — same.

## Recommended re-runs

- <skill name> with <specific guidance>
- ...

## Provenance spot checks (5 random)

- Quote in `pockets/<x>.md` line <n>: traced to `calls/<id>.md` ✓
- ...

## Stop / continue

- If composite ≥ 95 → STOP. Output is A+.
- If composite < 95 and iteration < 3 → re-run weakest dimensions.
- If composite < 95 and iteration ≥ 3 and no improvement vs. previous grade → SURFACE BLOCKER. Likely missing data, not method.

Iteration logic (orchestrator's job, not the grader's)

After a grade lands, the orchestrator decides:

1. A+ (≥95): Done. Exit.
2. A (90–94): Optionally do one more round on the weakest dimension only. Don't loop forever.
3. B (80–89) or below: Mandatory re-run of the two weakest dimensions. Re-grade.
4. Plateau (no improvement vs. previous grade ± 2 pts): Surface the blocker to the user. Common blockers:
   • Insufficient call volume (extend window or accept lower confidence)
   • No enrichment access (recommend Clay or Apollo trial)
   • No CRM data (priority matrix loses ACV signal)

Mapping from rubric dimension to skill

Weak dimension	Re-run skill	Specific action
Per-call extraction completeness	call-mining	Re-extract failing call subset with tighter prompt
Verbatim quote preservation	call-mining	Re-extract; explicitly forbid paraphrase
Persona distinctiveness	persona-clustering	Re-cluster; force split on triggers
Trigger specificity	trigger-analysis	Re-tag with stricter event-not-want rule
Hidden pocket discovery	persona-clustering + trigger-analysis	Look at low-frequency clusters explicitly
Title precision	enrichment	Add exclusions; tighten seniority
Firmographic combinations	enrichment	Test 2-attribute → 4-attribute combinations explicitly
Tech / behavioral signal	enrichment	Pull tech stack + hiring data, not just firmographic
Value prop authenticity	pocket-packaging	Replace any synthesized hook with verbatim quote
Named example density	pocket-packaging	Pull more calls/ files; require ≥3 per pocket
Priority matrix	priority-matrix	Re-score with CRM data joined
Action implications	priority-matrix	Add archetype per pocket
Anti-ICP / disqualifiers	pocket-packaging + priority-matrix	Mine stalled/lost calls explicitly

Provenance check (every iteration)

Spot-check 5 random quotes in pocket files. For each, open the referenced calls/<id>.md and verify the quote is there verbatim. If any fail: penalize Buyer Voice by 5 points and re-run pocket-packaging with strict provenance.

State update

{
  "step": "grading",
  "step_status": "completed",
  "iteration": <n>,
  "last_grade": <score>,
  "letter": "A+|A|B|C|D",
  "weakest_dims": ["...", "..."]
}