Pi Sych

Visible project state · bounded delegation · human-owned decisions

Attribution and intellectual influences

Pi Sych is distributed under the repository’s MIT docs/LICENSE.md. This file records the known direct influences identified in project history, the owner’s source inventory, and the current package design. It also separates intellectual influence from host platforms, integrations, and first-party material. Pi Sych does not include or relicense source text or code from the works and projects cited below unless a file says otherwise.

Writing and reasoning methods

Prose

Sarnecka (2021) informs the topic-sentence and reverse-outline practices, hamburger paragraph, familiar-to-new information flow, end-weight, concrete wording, global-to-local revision, and defeasible rather than absolute writing rules.

Pullum’s critiques of mechanical grammar advice inform Pi Sych’s diagnosis of grammatical passives, rejection of blanket active-voice rules, and treatment of passive voice as a choice about topic, information structure, and agency. The package does not infer a passive from a form of be alone and does not treat every agentless passive as a defect.

These sources support strong but defeasible defaults. They do not make every other prose recommendation a claim attributable to Sarnecka or Pullum.

Hypothesis generation and perspectivism

McGuire (1989, 1997) informs the deliberate generation of multiple accounts, contrary cases, scope and moderator questions, rival mechanisms, and observations that discriminate them. His heuristic catalogue is the source for the compact transformations in the shared hypothesis-generation method. Pi Sych keeps generation distinct from support: a heuristic can produce a candidate but cannot make it evidentially supported.

Pi Sych uses a compact practical subset: contrary cases, reversed causation, moderators, multiple accounts, counterforces, deviant cases, conflict reconciliation, extreme conditions, re-operationalization, decomposition, restatement, analogy, and discriminating study sequences.

Zahavy (2026) was the inspiration for a narrower grounding safeguard: a language model may diversify, formalize, compare, analyze scope, and propose discriminating tests for candidate hypotheses, but it must not invent the sensory, experiential, empirical, or literature material that supposedly motivated them. Pi Sych adopts this operational risk, not the paper’s stronger conclusion that present LLMs are structurally incapable of an abductive scientific jump.

Argument and claim analysis

The argument-analysis and claim-evidence methods consolidate procedures from Pi Sych’s earlier theoretical, empirical, review, research, and analysis guidance. They do not adopt or reproduce one external formal system. McGuire’s perspectivism also informs their attention to serious rivals and discriminating implications, but claim-to-artifact provenance and the premise/inference/scope distinctions are first-party syntheses rather than an attribution to McGuire alone.

Package and harness design

Three owner-supplied references inform the package’s view that model behavior depends on the deployment harness and on the quality of task-specific context, not only on base-model capability.

Weng’s (2026) inspires the treatment of context, tools, action, artifacts, and evaluation as behaviorally material parts of a model’s deployment system. Pi Sych applies that lesson through explicit project files, bounded context packets, short-lived workers, and visible verification boundaries. Regression-aware retrospective proposals name the targeted component and predicted effect, then separate motivating cases from held-out checks. Pi Sych does not implement recursive self-improvement or autonomous prompt mutation.

Goedecke’s (2026) supports the decision to spend skill context on domain distinctions and model-specific failure modes rather than generic advice a capable model already follows.

The reported effects of retained reasoning and compaction inform Pi Sych’s attention to working-memory continuity, compaction, token use, and harness-sensitive evaluation. This citation does not claim that Pi Sych reproduces OpenAI’s settings, benchmark, or reported results.

Retrospective workflow inspirations

Three public projects inspired the narrow, cautious retrospective proposal format used in Pi Sych’s project guidance:

Pi Sych adopts only the idea that retrospective lessons can be proposed for human review. It does not copy their code, install their hooks, mine unattended transcripts, or adopt automatic edits or commits.

Platforms, integrations, and first-party material

Pi is the host platform and extension system for Pi Sych. The package currently targets the earendil-works/pi distribution of Mario Zechner’s Pi coding agent. The project also grew out of the author’s move from OpenCode to Pi; that is experiential lineage, not a code or API dependency.

pi-mcporter and its MCPorter runtime provide the optional remote-research bridge.

Plannotator provides browser annotation and code-review interfaces. These are dependencies or integrations, not sources for the writing and reasoning methods. Their own licenses govern their packages.

templates/revealjs-baseline.css is adapted from the Pi Sych author’s own talk styles, not third-party CSS. The argument-analysis and claim-evidence methods likewise preserve first-party lineage from earlier Pi Sych guidance.

Limits

Attribution records influence, not correctness or universal authority. It does not establish that a local edit, hypothesis, argument, citation, or harness choice is sound.