Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/xxyoudeadpunkxx/pa-pvp-mini/llms.txt

Use this file to discover all available pages before exploring further.

pa_pvp mini is calibrated for material with internal structure — steps, claims, constraints, rules, expected behavior, or dependencies. The more structure an artifact has, the more the reviewer can find to challenge. An artifact with no internal logic gives the reviewer nothing to anchor a finding to, and findings with no anchor are excluded by the protocol.

​
Strong candidates

These material types consistently produce high-quality findings because they carry the properties that make adversarial review meaningful: stated intent, traceable logic, and discrete elements that can break independently of each other.

Plans and proposals

Decision memos, project plans, and proposals with steps, criteria, or explicit success conditions. The reviewer targets gaps between stated goals and the mechanics used to reach them.

Prompts and AI workflows

System prompts, chain-of-thought templates, and AI instruction sets with rules, constraints, or behavioral expectations. Conflicting rules and underspecified edge cases surface quickly under adversarial pressure.

Procedures and checklists

Step-by-step processes, runbooks, and onboarding flows. The reviewer examines whether steps are sequenced correctly, whether preconditions are stated, and whether any step can silently fail without detection.

Technical specifications

Specs, API designs, architecture decisions, and data models. These carry implicit constraints and behavioral contracts that the reviewer can test for completeness and internal consistency.

Bug fixes and code changes

Proposed fixes, refactors, and patches with a stated problem and stated intent. The reviewer checks whether the fix actually resolves the stated problem and whether it introduces new failure modes.

AI-generated answers

Outputs that sound plausible but have not been pressure-tested. AI answers often carry confident language over shaky assumptions — adversarial review surfaces where the logic breaks before you act on it.

Policy and rule sets

Access policies, decision rules, and governance documents. The reviewer looks for rule gaps, conflicting conditions, and cases where the policy produces undefined or unintended behavior.

Research outlines

Structured research plans and literature reviews with explicit claims, scopes, or methodologies. The reviewer challenges unsupported claims and structural gaps in coverage or reasoning.

Instructions and structured arguments

Step-by-step instructions written for someone else, and structured arguments that lead from premises to conclusions. The reviewer checks whether the instructions cover preconditions and failure cases, and whether the argument’s logic holds under pressure.

​
What makes a good artifact

The core question is: does the artifact contain things that can break independently? The more of these properties are present, the more the reviewer can produce findings against:
  • Steps — a sequence where order matters and a wrong step produces a wrong outcome
  • Claims — stated assertions that can be tested for accuracy or support
  • Constraints — explicit limits on what the artifact allows, requires, or excludes
  • Rules — conditional logic describing when something applies or does not apply
  • Dependencies — relationships where one element relies on another being correct or present
  • Expected behavior — descriptions of what should happen under specific conditions
  • Internal logic — a line of reasoning that leads from a premise to a conclusion
An artifact does not need all of these. Even one or two is enough to produce valid findings. The threshold for a useful review is low — a three-step procedure or a five-rule policy is sufficient. The ceiling is higher: a complex spec with many interacting constraints will generate more findings with more structural weight.

​
Poor fits

pa_pvp mini is less useful when the material has no internal structure for findings to anchor to. Purely creative writing — fiction, poetry, free-form narrative — has no claims to challenge and no logic to break. Raw unstructured notes that represent thinking-in-progress rather than a stated artifact give the reviewer nothing to evaluate against. Very short content with no internal logic (a single sentence, a label, a heading without body) will return few or zero findings, which is correct behavior — the protocol does not pad output to reach a number. If you find yourself getting no findings, the artifact is likely too short or too unstructured. Add more content or use a different review approach.

​
Labeling your artifact

pa_pvp mini uses an explicit block label to identify the material under review. The label is: 
ORIGINAL_ARTIFACT:
[your material here]
The label matters because the reviewer is instructed to ignore chat memory, prior session assumptions, and anything not contained in one of the explicit input blocks. If your material is not labeled, the reviewer may still process it — but labeling removes ambiguity about what is artifact and what is context. It also enables precise evidence anchoring, since every finding is required to reference ORIGINAL_ARTIFACT directly. The label is also the boundary. Everything outside it is not the artifact. This matters when you include context, instructions, or notes alongside your material: put those outside the block, and the reviewer will treat only the labeled content as the thing under review.
Submit the full artifact rather than an excerpt or a representative sample. The section concentration rule limits the reviewer to no more than two findings per section (with a hard cap of four in cases of concentrated structural failure). This means the reviewer must distribute attention across the whole document — it cannot spend all five findings on one section. If you submit only part of the artifact, the review covers only that part, and you lose the coverage enforcement the protocol provides. A complete submission is always stronger than a partial one.

Build docs developers (and LLMs) love

Get started for freeTalk to us