Compile: A Fourth Primitive for Canon-Keeping AI Agents

In Engine, Rules, and Canon (Salvo 2026b), the author proposed that a canon-keeping learner requires exactly three primitive operators: observe, decide, and remember. The three-primitive minimality claim is correct in the theoretical limit where the canon can be directly consulted during decision, but a runnable reference implementation exposes a missing operator: the canon must be materialized into a bounded, well-typed view before the decision step can consult it. This document names that missing operator compile(C, x) → P and extends Salvo (2026b) §3.2 by one primitive. The compile primitive has two structural requirements, both inherited from §7 of Salvo (2026b): the substrate of the compiler must be disjoint from the substrate of the generator performing the decide step, and the final packet must be assembled by a deterministic non-LLM process even if LLM-aided selection is used inside the compiler. We introduce a fifth witness primitive, g_fidelity, that checks whether the generator's output is grounded in the compiled packet's substrate, and a sixth primitive, rule_compliance, that enforces an explicit rule set (forbidden libraries, commands, and regex patterns) at every tool-call commitment. A pre-dispatch secret scrubber blocks turns whose compiled packet contains a recognized secret pattern. Polycode v1.1.2, the runnable reference implementation at github.com/polylogicai/polycode, ships these primitives under MIT license. A nine-test rigorous validation harness (four functional, five security) is included in the release and was executed as part of this paper's preparation with all nine tests passing. The test harness results are reported verbatim in §7 and §8 of the PDF. Disclosure. Polycode is inspired by Claude Code's public architecture at docs.claude.com. Polycode is not affiliated with Anthropic. No Claude Code code or system prompts are copied into polycode. Anthropic (compile tier) and Groq (generator tier) are consumed as arms-length paid API providers. Neither provider has reviewed or endorsed this paper or the software. Chain: This paper is derived from and cites Salvo 2026b, continues and cites Salvo 2026c, and is supplemented by the polycode and polybrain-kernel repositories. The paper PDF is Bitcoin-anchored via OpenTimestamps (paper.pdf.ots) at deposit time. License: CC BY 4.0 (this document). MIT (polycode source). Axiom: substrate cannot witness itself.

Authors

Institutions

Publication Details

Journal
Open MIND
Published
2026-04-15
DOI
https://doi.org/10.5281/zenodo.19600252
Primary Topic
Security and Verification in Computing
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
article

Compile: A Fourth Primitive for Canon-Keeping AI Agents

Andrew Salvo
Open MIND
Security and Verification in Computing
article

Compile: A Fourth Primitive for Canon-Keeping AI Agents

Andrew Salvo
article en

Abstract

In Engine, Rules, and Canon (Salvo 2026b), the author proposed that a canon-keeping learner requires exactly three primitive operators: observe, decide, and remember. The three-primitive minimality claim is correct in the theoretical limit where the canon can be directly consulted during decision, but a runnable reference implementation exposes a missing operator: the canon must be materialized into a bounded, well-typed view before the decision step can consult it. This document names that missing operator compile(C, x) → P and extends Salvo (2026b) §3.2 by one primitive. The compile primitive has two structural requirements, both inherited from §7 of Salvo (2026b): the substrate of the compiler must be disjoint from the substrate of the generator performing the decide step, and the final packet must be assembled by a deterministic non-LLM process even if LLM-aided selection is used inside the compiler. We introduce a fifth witness primitive, g_fidelity, that checks whether the generator's output is grounded in the compiled packet's substrate, and a sixth primitive, rule_compliance, that enforces an explicit rule set (forbidden libraries, commands, and regex patterns) at every tool-call commitment. A pre-dispatch secret scrubber blocks turns whose compiled packet contains a recognized secret pattern. Polycode v1.1.2, the runnable reference implementation at github.com/polylogicai/polycode, ships these primitives under MIT license. A nine-test rigorous validation harness (four functional, five security) is included in the release and was executed as part of this paper's preparation with all nine tests passing. The test harness results are reported verbatim in §7 and §8 of the PDF. Disclosure. Polycode is inspired by Claude Code's public architecture at docs.claude.com. Polycode is not affiliated with Anthropic. No Claude Code code or system prompts are copied into polycode. Anthropic (compile tier) and Groq (generator tier) are consumed as arms-length paid API providers. Neither provider has reviewed or endorsed this paper or the software. Chain: This paper is derived from and cites Salvo 2026b, continues and cites Salvo 2026c, and is supplemented by the polycode and polybrain-kernel repositories. The paper PDF is Bitcoin-anchored via OpenTimestamps (paper.pdf.ots) at deposit time. License: CC BY 4.0 (this document). MIT (polycode source). Axiom: substrate cannot witness itself.

Open MIND
Pennsylvania State University (US)
Peace, Justice and strong institutions
Openalex Percentile: Top 6%
Security and Verification in Computing
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.