#Pyroduct

#Features

• Paper-executable, not paraphrased
  Every state, transition and phase is literal data with a source citation: gloss() carries the text's own wording verbatim, and each construct is tagged 原文直述 (verbatim from the source) or 模型整理 (model reconstruction). Page numbers are not part of that discipline: P.xx appears only where the source survey actually recorded one — multi/group/society — and never in the subject machine's own tables.

• One language, four units
  Subject (主体) / group (群体) / society (社会) share the same state + trigger + transition vocabulary at three scales: src (34·53·11), group (20·44·9, member config → emergence with R7 write-back), society (13·28·8, lifeworld vs. system). multi is the normative face: rules R1–R14 for two subjects — regime precedence, validity-claim timing, the two-sincerity distinction (两诚区分), unconcealment–concealment (解蔽—遮蔽) — each rule citing the thesis.

• Two-character names only
  Position and phase designations name a structural relation — nothing else. The discipline is locked by tests and printed by render_naming_table().

• Drive slots and a step contract
  Trigger::slot sorts all 49 triggers into eight slots (Occasion 处境 / Orientation 取向 / Antecedent 先行 / Act 行动 / Feedback 反馈 / Hold 持存 / CoBeing 共在 / Unveil 揭蔽) with an exhaustive match — an unplaced trigger fails the build. step(position, slot) answers only Moved (迁) / Held (守) / Undecided (未定) / Same-destination (同归) / No-way (无路); of 272 position × slot combinations, 228 honestly answer No-way. The machine never invents a move — and since 0.1.26 it never invents which sentence moved it either: on the three position × slot pairs where several triggers lead to one target it answers 同归 (destination known, cause unnamed) instead of naming the first one in the table. step_by_trigger(position, sentence) takes the sentence as given.

• An agent whose updates must clear a gate
  evolution treats an agent's config as Genome, enumerates single-step candidates (prompt / strategy / memory / code), writes "better" as an Objective, fences off constitutional invariants, and records every accept / reject / revert with its reason in an append-only, serializable Ledger. Because every trail names the trail it came from, that ledger is also a tree — roots, children, depth, ancestry, leaves, height, lineage — and the lineage it reports is checked, not assumed: a non-root's parent must be a root or an accepted change, and differ from it by exactly one mutation. Gates come in two grades: hand-written weights, or a DoubleML causal effect (θ̂ > 0 and the 95% lower bound > 0). The subject cycle run_cycle drives real positions by drive slot and puts Cultivation (修习) and Opening (开放) through the same gate — no evidence, no advance (没有证据就不前进).

• A coordinator that refuses to close
  coordinator adds what a research agent needs beyond a state machine: tasks and artifacts, planning over a dependency graph, external evidence judged by confidence interval (not by votes), content memory that keeps superseded findings, credit with IP provenance, and a deliberation / decision split — deliverables carry revision marks, residue ownership and reopen conditions. The runtime is replayable and persists to disk. (The "think" step is still a deterministic stand-in, no LLM yet.)

• Dependency isolation
  Four named third-party dependencies, each confined to its consumer: moonbitlang/async@0.22.7 → cmd/coord (native disk I/O) and tools/ifacescan / cmd/ifacescan (the .mbti drift tripwire, native); riantr/moonbit_doubleML@0.164.56 → dmlref / causal; riantr/snn_mbt@0.166.2 → snnref (native); riantr/moonbit_static_analysis@0.4.39 → audit. Every other package uses official moonbitlang/* only.

#What you get

• CLI — moon run cmd/main -- <face> prints any layer as a report. 37 subcommands; an unknown one exits non-zero rather than printing the default report.
• JSON bridge — cmd/jsoncli takes one JSON request and returns one JSON reply {ok, kind, rendered, faces}. This is what the two agent plugins (plugin/ for MiniMax, .dsh/ for DeepSeek) spawn; it is the only surface an agent needs.

#Four layers that check the machine itself

#How this plugs into a running agent

• constitutional — three invariants (R10 residue ≥ 1; R2/R14 no declarative voice; R3 still open to the other) that no score can buy. A violation is rejected outright.
• selection — either hand-written weights, or a DoubleML causal effect: accept only if θ̂ > 0 and the 95% lower bound > 0.

#What the causal claim assumes

1. No unmeasured confounding given X (E[ε | X, T] = 0). This is the assumption that "find all confounders" cannot deliver — confounders are not discovered by the estimator, they are assumed away by the conditioning set. If a confounder is not in X, no amount of machine learning inside the estimator removes it.
2. Overlap / positivity — both treatment states must occur at every X. The repo's data satisfies it by construction (treatment is a continuous function of X plus noise), but the repo does not implement an overlap diagnostic — a real deployment must check it.
3. A correctly specified or learnable nuisance m(X). Cross-fitting protects against overfitting the nuisance; it does not protect against a nuisance that cannot be learned from the data at hand.
4. A well-defined intervention. T must be a manipulable treatment with a coherent counterfactual, not a proxy for something else.

#Comparison with the libraries this mirrors

#Terminology

• 处境 Occasion → Situation — an occasion is a particular ceremony; 处境 is the situation one finds oneself in.
• 修习 Discipline → Cultivation — 修养 is cultivation, and "discipline" has to stay free for the group's 规训 Disciplined, which means something else: binding the group and its members and putting freedom away.
• 自缚 Binding → Self-binding — the 自 is the whole point. You bind yourself; the table does not.
• 交织 Weaving → Interweaving — weaving can be one-sided, 交织 is mutual, and it pairs with the Interweave trigger.
• 无路 Block → No-way — a missing road, not a missing edge. 界域 frontier is a core concept in this model and "edge" would collide with it.
• 先行 Antecedent → Antecedent — kept. It was Anticipation, which names the attitude the slot produces rather than the structural relation it is.

#Phases — subject (11)

#Positions — subject (34)

#Drive slots (8) and the 49 triggers

#Step contract (5 outcomes)

#Group — phases (9)

#Group — states (20)

#Society — phases (8)

#Society — states (13)

#Truth regimes (5) · discourse types (4) · validity claims (4)

#Normative rules R1–R14

#Packages

#Discipline terms

#English rendering

#Quick Start

moon add riantr/pyroduct
moon test             # Total tests: 247, passed: 247, failed: 0.

moon run cmd/main -- report        # subject state machine
moon run cmd/main -- slots         # drive slots: 49 triggers → 8 slots
moon run cmd/main -- multi         # normative rules R1–R14 + scenarios
moon run cmd/main -- group         # group state machine
moon run cmd/main -- society       # two-level social model
moon run cmd/main -- evolution     # mutation / fitness / gate / persistence / heredity
moon run cmd/main -- cycle         # subject cycle: position × slot × causal feedback
moon run cmd/main -- coordinator   # research-agent coordinator
moon run --target native cmd/coord # real disk I/O: checkpoint round-trips identically

#Example

// Subject: positions, main line, drive slots, step contract
let states : Array[@src.SubjectState] = @src.all_states()      // 34 of them
let ok     : Bool = @src.reachable(@src.initial(), @src.terminal())
let moved  : @src.Outcome = @src.step(@src.Standing, @src.Occasion) // Moved

// One encounter of two subjects: regime, two sincerities, and claim timing all adjudicate
let me    = @multi.subject(1, @multi.rendezvous(), @multi.InterSubjective, 70, 70, 9)
let other = @multi.subject(2, @multi.rendezvous(), @multi.ChengRan, 80, 80, 9)
let claim : @multi.ValidityClaim = { kind: @multi.Truthfulness, phase: @multi.During }
let out   = @multi.encounter(me, other, @multi.Theoretical, claim)

// Society: two invariants locked by tests
let never_closes : Bool = @society.no_final_closure()

// Agent state machine: enumerate candidates → gate → append-only ledger
let agent  = @evolution.default_agent()
let ledger = @evolution.evolve(agent, @evolution.default_world(),
                               @evolution.Weighted(@evolution.truth_objective()), 3)

// Coordinator: run a demo study, write / read back the checkpoint
let rt   = @coordinator.new_runtime()
ignore(rt.step(@coordinator.demo_tasks(), @coordinator.demo_questions(),
               @coordinator.demo_workers(), @coordinator.demo_datasets(),
               @coordinator.demo_world(), @coordinator.demo_parent(),
               @coordinator.demo_candidate()))
let restored = @coordinator.restore(rt.checkpoint())
let same : Bool = restored.round_trips()

#Gates and drift detectors

#Used By

• cmd/main — the report entry for every layer above: report / slots / loop / multi / group / society / evolution / cycle / coordinator, dmlref / causal / audit / fleet / mutants / mbti, plus mermaid / dot / genesis / course / naming / principle / intuition, the five analysis faces pathsum / algebra / petri / aho / buchi, and the presentational all.
• cmd/coord — a native binary that runs the coordinator and round-trips its checkpoint through real disk I/O (.openseek/, gitignored).
• cmd/jsoncli — the JSON bridge the two agent plugins spawn (31 faces).

#Attribution

• Source text: 任勇祥《伽达默尔与哈贝马斯真理观比较》(Shanghai Academy of Social Sciences, 2008) and its appendix 《通往宁静之路：个体道德发生论纲要》. Every construct in the model cites the chapter / section / page of that text and explicitly distinguishes 原文直述 (verbatim from the source) from 模型整理 (model reconstruction).
• License: MIT (see LICENSE).