Skip to content

moncface/scr

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 
 
 

Repository files navigation

Shared Context Reduction (SCR)

License: CC BY 4.0

The systematic elimination of information already mutually known between communicating agents.


What Is SCR?

When two parties share prior knowledge, most transmitted information is redundant. SCR is the principle of removing that redundancy — not by compressing data, but by not sending what both sides already know.

  • An LLM that knows Python syntax doesn't need you to explain it.
  • A robot that has mapped a room doesn't need you to describe the furniture.
  • A voice assistant that knows your routine doesn't need you to spell it out.
  • A car's AI that has driven a route 200 times doesn't need turn-by-turn directions.

SCR is not compression. Compression encodes the same information in fewer bits. SCR transmits less information, because the missing portion already exists at the receiver.


Five Principles for Safe Reduction

Removing shared context only works if both sides agree on what may be inferred from the gaps. Without constraints, the receiver fills gaps with guesses — and guesses, at scale, become failures.

These five principles form a decision hierarchy, from broadest to strictest.

1. Boundary of Omission

The receiver may only infer universally established facts: programming language syntax, framework conventions, mathematical constants, physical laws.

It must not infer: culture-dependent knowledge, time-sensitive information, personal experience, or emotional nuance.

Example: A Django project that omits language:python — Python can be safely inferred. A project that omits its status — "everything is fine" cannot be inferred. Project health is not a universal fact.

Theoretical basis: Grice's Cooperative Principle (1975).

2. Prohibition of Implicature

The receiver must not generate implied meaning beyond what is explicitly stated.

test:pass means "tests passed." It does not mean "code quality is high," "ready to deploy," or "no bugs remain."

Theoretical basis: Grice's theory of conversational implicature. Implicature is culture- and context-dependent. Automated implicature has no guarantee of matching sender intent.

3. Verification of Presupposition

When the receiver encounters uncertain presuppositions, it must pause inference and request clarification rather than guess.

bug:auth-again — "again" presupposes a prior occurrence. Without a record of a previous auth bug, the receiver interprets this only as "auth-related bug" and does not infer recurrence.

Theoretical basis: Presupposition theory in formal semantics.

4. Safe Interpretation Selection

When input permits multiple interpretations, the receiver must select the lowest-risk interpretation.

status:debugging could mean a status report (low risk), a help request (medium risk), or a warning (high risk). The receiver selects the status report. If the sender meant otherwise, they can clarify.

Theoretical basis: Austin and Searle's Speech Act Theory (1962/1969).

5. Precision Guarantee

Numbers, negations, conditions, and exceptions are never inferred. If a value requires precision, it must be explicitly stated.

  • timeout:long → must not be resolved to a number. Must ask.
  • auth:no-password → scope of negation must be confirmed.
  • deploy:staging → conditions and constraints must be verified.

Theoretical basis: Scalar implicature and negation scope theory.


Principle Hierarchy

These principles are not equal. They form a decision chain where each level tightens the constraint:

Principle 1 (highest priority):
  → What CAN be inferred? Draw the boundary.

Principle 2:
  → Within that boundary, generate NO implied meaning.

Principle 3:
  → If a presupposition is uncertain, STOP and ask.

Principle 4:
  → If multiple interpretations remain, choose the SAFEST.

Principle 5 (strictest):
  → If precision is needed, NEVER guess. Always ask.

Relationship to LNDF

SCR is the general principle. LNDF (LLM-Native Data Format) is the first formal implementation of SCR, applied to structured data transmission between humans and LLMs.

SCR    = What to remove    (the principle)
LNDF   = What to write     (the format)
5 Principles = What to infer (the safety constraints)
  • SCR defines the process: eliminate mutually known information.
  • LNDF defines the sender's discipline: six principles for writing minimal, intent-focused data.
  • The Five Principles define the receiver's constraints: what may and may not be reconstructed from omitted information.

Applicability

SCR is medium-independent and agent-independent. It applies wherever information is transmitted between parties who share prior context:

  • LLM communication (text, API, agent frameworks)
  • Robotics and physical AI (sensor data, status reporting, command interpretation)
  • Voice interfaces (conversational assistants, in-car systems)
  • IoT and edge computing (device-to-cloud telemetry, firmware updates)
  • Multi-agent systems (agent-to-agent coordination)

Theoretical Foundations

The Five Principles draw from established work in linguistics and philosophy of language:

  • Grice, H.P. (1975). Logic and Conversation. — Cooperative Principle, conversational implicature
  • Austin, J.L. (1962). How to Do Things with Words. — Speech act theory
  • Searle, J.R. (1969). Speech Acts. — Illocutionary force, indirect speech acts
  • Brown, P. & Levinson, S.C. (1987). Politeness: Some Universals in Language Usage. — Face-threatening acts in communication
  • Sperber, D. & Wilson, D. (1986). Relevance: Communication and Cognition. — Relevance theory, optimal inference

Specification Governance

This specification is maintained by Hiroaki Tachibana (moncface).

  • Change proposals: Submit via GitHub Issues
  • Final decision authority: The maintainer (BDFL model)
  • Versioning: Semantic versioning (current: v0.1)
  • Canonical URL: This README is the authoritative specification

License

Specification and documentation: CC BY 4.0

The concepts described are freely available for use, adaptation, and extension with attribution.


Hiroaki Tachibana / moncface April 2026

About

Shared Context Reduction (SCR) — the systematic elimination of mutually known information between communicating agents.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors