Conflicting ISC minimums and effort tiers between PAI/SKILL.md and PAI/Algorithm/v3.7.0.md

[catchingknives]

Summary

PAI/SKILL.md and PAI/Algorithm/v3.7.0.md both define effort tiers and ISC (Ideal State Criteria) minimums, but they disagree on the numbers, tier count, and time budgets. Since both files ship with PAI and are loaded during different execution paths, this creates ambiguity for every PAI user about which spec is authoritative.

Specific Conflicts

ISC Minimums

Effort Tier	Algorithm v3.7.0.md	SKILL.md	Discrepancy
Standard	8	8	✅ Same
Extended	16	33	2x difference
Advanced	24	64	2.7x difference
Deep	40	128	3.2x difference
Comprehensive	64	256	4x difference

Effort Tier Count

• Algorithm v3.7.0.md defines 5 tiers: Standard, Extended, Advanced, Deep, Comprehensive
• SKILL.md defines 7 tiers: Instant, Fast, Standard, Extended, Advanced, Deep, Comprehensive

Time Budgets

Tier	Algorithm v3.7.0.md	SKILL.md
Extended	<8min	<4min
Advanced	<16min	<8min

Capability Selection Methodology

• Algorithm v3.7.0.md: Free selection from skills + platform capabilities during OBSERVE phase
• SKILL.md: Full 25-capability audit with explicit USE/DECLINE/N/A assignment for each

Response Depth Names

• CLAUDE.md (which points to Algorithm v3.7.0): NATIVE / ALGORITHM / MINIMAL
• SKILL.md: FULL / ITERATION / MINIMAL (ITERATION is not defined in CLAUDE.md)

ITERATION Mode — Defined in SKILL.md, Referenced But Undefined in Algorithm v3.7.0.md

Algorithm v3.7.0.md's Critical Rules section says:

every response uses exactly one (ALGORITHM, NATIVE, ITERATION, or MINIMAL)

But neither CLAUDE.md nor Algorithm v3.7.0.md defines what the ITERATION template looks like. Only SKILL.md defines it (as a condensed format for follow-ups within an Algorithm run: CHANGE + VERIFY + voice line).

In practice, the AI produces ITERATION mode regularly based on what it picks up from SKILL.md:

🤖 PAI ALGORITHM ═════════════
🔄 ITERATION on: [context of what changed]
🔧 CHANGE: [what changed]
✅ VERIFY: [evidence]
🗣️ Nairyo: [summary]

If SKILL.md is non-canonical (as implied by CLAUDE.md pointing exclusively to Algorithm v3.7.0.md), the ITERATION template has no canonical home. It should be explicitly defined in Algorithm v3.7.0.md alongside the other modes.

Why This Matters

CLAUDE.md explicitly points to PAI/Algorithm/v3.7.0.md as the canonical Algorithm spec:

MANDATORY FIRST ACTION: Use the Read tool to load PAI/Algorithm/v3.7.0.md, then follow that file's instructions exactly.

But PAI/SKILL.md also exists in the system tier and is referenced by agent definitions. When the AI loads both (or loads one in one context and the other in another), it gets contradictory instructions — particularly around how many ISC to generate, which directly affects output quality. And ITERATION mode — which is used frequently — only has a template definition in the non-canonical file.

Suggestion

Consolidate to a single canonical Algorithm/ISC spec:

1. Have SKILL.md reference Algorithm v3.7.0.md rather than redefining the same concepts with different numbers
2. Move the ITERATION mode template into Algorithm v3.7.0.md so it has a canonical definition
3. This eliminates ambiguity and prevents the AI from getting conflicting instructions depending on which file it reads first

---

Found during a full PAI setup audit. Both files are identical to origin/main — this is not a local divergence.