PLUTONIUM-like Adversarial Peer Validation Protocol

Close decisions. Do not debate forever.

PA_PVP is an AI-native decision kernel for movement under uncertainty: forced verdicts, one executable next action, adversarial probes, rule-bound debt, and no narrative drift.

Successful first run: a forced verdict in [QUEUE], one [NEXT], and a derived [HUMAN_TABLE].

  • Forced verdict: DO NOW / DO LATER / DISCARD
  • One next action, not a pile of options
  • Reality probes when simulation stops being enough
PA_PVP banner

Why it exists

Doubt should generate motion, not delay.

Most decision systems fail in one of three ways: they keep debating, they simulate certainty, or they produce a list instead of a trajectory. PA_PVP is built to break all three failure modes.

Too much debate

No endless discussion loops. The kernel must force a verdict.

Too much simulation

When internal evidence is weak, the protocol escalates to probes or parks cleanly.

Too little movement

If the next step does not reduce uncertainty or change the world, the run is invalid.

What the kernel does

Adversarial critique in service of execution.

PA_PVP does not try to sound wise. It takes a plan, artifact, or target and converts it into critique, forced verdict, machine state, and a short executable path.

Forced verdict

Every valid run ends in a hard decision: DO NOW, DO LATER, or DISCARD.

Executable next

The protocol must return one next action, not an essay or a motivational summary.

Probe-first honesty

If confidence is fragile, PA_PVP forces a real-world test or parks instead of pretending.

Rule-bound debt

Near misses, open falsification debt, and circular dependencies become constraints, not notes.

PA_PVP verdict fork visualization

Surface

The output is meant to move something.

Input

<<<I id=E1 title="Short title">>>
<<<PLAN>>>
- ship X
- remove Y
- deploy by Friday
<<<END>>>

Output

STATE: DECISION.TEST.FINAL
[QUEUE] ... gate: ...
verdict: DO NOW
[NEXT] run the cheapest falsifier
[STEPS] max 5 atomic moves
PA_PVP execution chamber visualization

Reality inside the loop

Autonomy first. Human only when mechanically justified.

The default mode is ask_user=NONE. When internal evidence is not enough, PA_PVP either parks cleanly or, if allowed, asks for one minimal real-world probe and resumes with stronger evidence.

  • Snapshot-as-input continuation
  • One minimal probe instead of speculative certainty
  • Spent cycles and debt tracked across iterations

Start here

Use the kernel, then the guides.