Translation Bridge · the receiver laws, pointed at you
AI that adapts to your cognitive style — not the other way around.
RPCS-1 turns an agent's operating conditions into derived runtime settings. The bridge is the same model pointed the other way: you are the receiver. Five questions place you on the five primitives; the result compiles into deterministic rendering directives any AI can obey — instead of a system prompt it will skim and forget.
The framework underneath
Three blocks every bounded observer is forced to implement — including you
IMM Paper 18 derives what any bounded system must do to estimate a changing world from noisy signals and act in time: estimate → detect → commit. The five primitives are the measurement coordinates over those blocks. An agent is one such observer. A human reading a message is another. The bridge measures your coordinates and renders for them.
01
TI · UE
Estimate
How much history to keep, how fast to update.
How much context you need before the point lands (TI), and how readily you revise when pushed back on (UE). The bridge renders structure and revision posture to match — bottom line first for a low-TI reader, full context first for a high-TI one.
02
SG · FT
Detect
One alarm channel, two knobs: gain and criterion.
How much tone you read into a message (SG), and how explicit it must be before subtext registers (FT). A high-FT reader gets intent stated outright — no idiom, no hints. That is not a deficit; it is a criterion setting.
03
AR
Commit
Accumulate evidence to a bound, then act.
When an answer should commit versus surface the options (AR). A high-AR profile gets the best reading picked and answered; a low-AR profile gets the alternatives laid out before anything is locked in.
Same blocks, same laws, different receiver. That is the whole trick — and why the profile is coordinates on a derived chart, not a personality quiz. Read the framework →
The problem with prompting a mind-reader
System prompts describe you in prose and hope. The bridge replaces the hope with parameters — the same move the tuner makes for agents.
Standard prompting
The bridge
Static instructions
Long system prompts get skimmed, diluted, or overridden as context grows.
Structured receiver profiles
Preferences compile into explicit, versioned rendering directives that survive context growth.
Implicit adaptation
You rely on the model to infer your communication parameters from conversation statistics — a design premise we state as a testable prediction, not a finding, is that this inference is unreliable.
External parameter-setting
Structure, warmth, explicitness, revision posture, and ambiguity handling are set as explicit knobs, outside the model’s discretion.
Cognitive friction
Exploratory probes get read as factual claims; executive-function load gets read as disinterest.
Precision translation
Claims are tagged separately from probes; evidence is separated from speculation; intent survives the channel.
What ships today
Each feature is labeled with its actual status. That's the house style.
Behavioral calibrationLive
Five forced-choice questions — one per primitive — place you continuously on the five axes. Never a category label: ASD, ADHD, and AuDHD are regions of the same continuous space, not three boxes. Answers stay in your tab.
/calibrate →ReceiverProfile JSONIn SDK
Your profile exports as clean, versioned JSON with deterministic rendering directives — portable to any AI agent or LLM workflow. Schema published at /v1/receiver-profile.json.
/v1/receiver-profile.json →Governed refinement + masking mirrorIn SDK
The quick intake is a noisy prior. It refines from real interaction at a rate governed by your own UE — and the gap between where you set yourself and where your messages actually land (your masking shape) is surfaced back to you, never silently acted on.
Privacy-tiered data controlRoadmap
Tiered partitioning — Core Card, Working Profile, Private Archive — so profiles stay safe across public and enterprise endpoints. Designed; not yet shipped.
The schema, not the sales pitch
Five primitives, continuous [0,100] — never category labels. Directives derive deterministically from the profile; the profile is the source of truth.
{
"$schema": "https://rpcs1.dev/v1/receiver-profile.json",
"version": "1.0",
"profile": { "TI": 20, "SG": 25, "FT": 80, "UE": 75, "AR": 75 },
"directives": {
"structure": "bluf", // Estimate: bottom line first
"warmth": "minimal", // Detect: flat and factual
"explicitness": "explicit_literal", // Detect: no idiom, no hints
"revision": "open_challenge", // Estimate: pushback welcome
"ambiguity": "commit" // Commit: pick a reading, answer
},
"meta": { "source": "intake", "generator": "@rpcs1/core" }
}// matches ReceiverProfile + deriveRenderingDirectives in @rpcs1/core · schema: /v1/receiver-profile.json
Grounding — at exactly its earned grade
Most AI products claim more evidence than they have. This one is built on a public claim ledger, and this page obeys it.
- DerivedThe three-block receiver model and its matching laws are derived under named assumptions from standard optimality theory (IMM Paper 18) — checked against pre-registered numerical criteria, with the failures reported and cut.
- RegisteredThat a five-item intake usefully predicts a human’s communication preferences is a registered, testable hypothesis. The human-side validation battery is pre-registered with pass/fail criteria fixed before data generation — and has not yet been run. Results will be published either way.
- OpenWhether in-context AI receivers can self-tune these parameters from conversation statistics alone is an open, testable question. The bridge externalizes the parameters so you don’t have to bet on the answer.
// no claim exceeds its derivation