R&D
The research behind the product
RPCS-1's recommendations come from receiver laws derived in the IMM research program — an open, numbered paper series with a public claim ledger. This page is the technical layer: the framework, the derivations, the pre-registered checks (including the ones that failed), and the mapping between the product names on the site and the scientific names in the papers.
Product names ↔ scientific names
The site uses production names for the five receiver primitives. The papers, the SDK field names, and the API use the scientific identifiers. Same objects, two registers:
| Product name | Scientific name | SDK field | What it sets |
|---|---|---|---|
| Memory | Temporal Integration (TI) | ti | How much history your agent holds onto |
| Agility | Update Elasticity (UE) | ue | How fast it changes its mind when the world changes |
| Gain | Signal Gain (SG) | sg | How loud incoming signals get before it reacts |
| Trigger | Firing Threshold (FT) | ft | How much pressure it takes to notice something |
| Commit | Action Rule (AR) | ar | How much evidence it needs before it acts |
The framework underneath
Three blocks every bounded observer is forced to implement
IMM Paper 18 rebuilds the RPCS-1 foundations from the minimal requirements for being an observer at all: estimate a changing quantity from noisy observations, act on it under time constraints. Under those named assumptions, what boundedness forces is three coupled functional blocks — and the five RPCS-1 primitives are the measurement coordinates laid over them.
Estimate
How much history to keep, how fast to update. Derived law: integrate less and update faster when the world changes faster.
Detect
One alarm channel, two knobs. Derived law: turn gain up and liberalize the criterion as the change rate rises.
Commit
Accumulate evidence to a bound, then act. Derived law: commit faster when the world changes faster — thresholds scale with stakes-to-urgency.
When the environment changes faster, every derived law moves the same way:
Honest labels, kept honest: these are derived under named assumptions, with pre-registered numerical checks — three of which failed and are reported in full, with the repairs. The exact exponents are conditional on the environment class. That level of disclosure is the point. Read how the framework works →
Tested before it was shipped
Pre-registered checks, pass and fail alike
Every law behind the recommendations was checked numerically against criteria fixed beforethe data was generated. Here's the scorecard — including what failed:
| Check | Predicted | Measured | Verdict |
|---|---|---|---|
| Update-elasticity exponents (R-2) | +0.667 / +0.5 / +0.5 | +0.651 / +0.496 / +0.454 | PASS |
| Sampling-rate exponents (R-3) | +0.667 / +0.5 / +0.5 | +0.620 / +0.493 / +0.488 | PASS |
| Detection value (R-4): deaf receiver | interior optimum | deaf loses 3.2× | PASS |
| Commitment stakes law (R-5) | slope ≈ 1 | 0.977 | PASS |
| Criterion log-law slope (R-4) | −0.51 | −0.33 | FAIL — cut from claims |
Three registered checks failed during development; one was corrected and re-run, one is cut from the claims entirely, one was traced to a discretization artifact. All three are reported in the paper — because a scorecard you can trust has to include the misses. Full claim ledger →
Go deeper
The IMM research program
The open paper series the receiver laws are derived in, with its public claim ledger and status vocabulary.
Primitive reference
The five primitives in full: definitions, measurement, and how each maps to runtime settings.
SDK & API docs
Getting started, MCP server, platform mappings, and worked examples.