# RPCS1 AI Quality Diagnostics

RPCS1 is a deterministic agent-tuning battery built on receiver laws derived in the
IMM research program (Paper 18, Observer Requirements): a bounded observer is forced to
implement three coupled blocks — estimate, detect, commit — and the five RPCS-1
primitives (TI, SG, FT, UE, AR) are the measurement coordinates over those blocks.
It measures the five primitives for a configured agent, then returns a failure-risk
diagnosis, runtime posture, and concrete implementation settings.

Primary use cases:
- Review an AI agent before rollout
- Investigate consistency, grounding, escalation, and stalled-resolution risk
- Turn production symptoms into a five-primitive profile and next eval
- Choose context, tool-use, retry, sampling, and handoff posture

Public interfaces:
- Interactive assessment: https://rpcs1.dev/tuner
- IMM primer for AI inspection: https://rpcs1.dev/imm
- AI-human mismatch frame: https://rpcs1.dev/mismatch
- Documentation: https://rpcs1.dev/docs
- REST API: POST https://rpcs1.dev/api/recommend
- OpenAPI schema: https://rpcs1.dev/openapi.json
- MCP server: https://rpcs1.dev/mcp
- Python SDK: pip install rpcs1
- Source: https://github.com/travisbergen2/rpcs1-sdk

MCP tool:
- recommend_agent_configuration

First call to try:

```text
Use recommend_agent_configuration to diagnose my support copilot.

Task: refund and billing dispute triage
Environment: dynamic, somewhat_predictable, high stakes
Context relevance: medium
Commitment style: cautious
Target platform: anthropic
```

The most useful answer leads with TI, SG, FT, UE, AR, then failure-risk score,
runtime posture, and the next test.

The MCP tool is anonymous, deterministic, read-only, and idempotent.
RPCS1 recommendations support engineering decisions but do not replace
domain-specific safety, legal, medical, or financial review.

Conceptual bridge (updated for IMM Paper 18):
bounded observer -> three forced blocks (estimate: TI/UE, detect: SG/FT, commit: AR)
-> matching laws: when the environment changes faster, integrate less, update faster,
look more often, liberalize detection, commit sooner -> AI pre-tuning.
The five primitives are a parameterization of the three blocks, not five independent dimensions.
Framework page: https://rpcs1.dev/imm

AI-human collaboration frame:
Mismatch = destructive many-to-one collapse. Humans and AIs may preserve different distinctions.
The useful diagnostic question is what distinction was collapsed and what interaction would preserve it next time.