TAS-184A: KV-cache and activation-state discipline audit

[AtlantisPleb]

Summary

• prove that KV cache and activation state are not silently carrying the
  interpreter in a way that breaks the article claim

Roadmap Position

• Phase: Phase 7. Close single-run posture, state discipline, and weight ownership
• Sequence: article-gap closure after TAS-156A
• Claim posture: necessary for article-equivalent closure, but not sufficient by itself to widen public capability claims

Description

• audit:
  • KV-cache size growth versus problem size
  • whether algorithmic state is encoded primarily in cache versus weights
  • whether truncation or reset breaks correctness
  • whether equivalent behavior survives under constrained cache
• define acceptable versus non-acceptable uses of:
  • KV cache
  • residual stream
  • attention history
• publish one explicit verdict on whether the canonical route is:
  • weight-dominant
  • activation-dominant
  • or mixed, with declared boundaries

Why This Matters

• without this, the route can look like "compute in Transformer weights"
  while really behaving like a state machine in cache or activations

Primary Surfaces

• psionic-runtime
• psionic-eval
• psionic-research
• psionic-provider

Validation

• KV-cache growth report
• cache truncation/reset sensitivity report
• constrained-cache equivalence checks
• dominance verdict checker

Claim Discipline

• Keep machine-legible boundaries explicit; partial progress must not be presented as full article equivalence.
• Preserve typed refusal, fallback, blocked, and suppressed posture where closure is still incomplete.
• Keep public language bounded to the declared article envelope rather than implying generic arbitrary-program closure from a partial green result.

Done When

• the repo has one explicit KV-cache and activation-state discipline verdict