# HCGM Next Grow Decision > Status: MSP next-grow gate v0.1 > Date: 2026-06-26 > Boundary: this is a decision gate, not a public essay, tutorial chapter, > theorem promotion, guide chapter, book draft, or benchmark plan. > Scope update, 2026-07-11: this file governs the earlier four-entrance > propagation / teaching / theory / authority review phase. It is provenance, > not the current global gate for the regression, binary-treatment, and > continuous-treatment paper program. Start from `README.md`, `RESTART_HERE.md`, > and `paper/README.md` for current routing. > Relocation, 2026-07-16: references below to `paper/arxiv-hcgm-regression/` > preserve the historical phase decision. The active flagship is now > `paper/learning-discoscm/`; do not execute the old paths. ## Why This Exists The HCGM project now has four strategic entrances: ```text propagation -> entrances/WHY_HCGM_FIRST.md teaching -> toy/TEACHING_PATH.md and toy/TEACHING_CONTRASTS.md theory -> entrances/THEORY_ENTRANCE.md and paper/arxiv-hcgm-regression/ authority -> entrances/AUTHORITY_ENTRANCE.md and entrances/FIRST_CHOICE_FRAMEWORK_LEDGER.md ``` The current risk is not lack of material. The risk is growing all four at once. This file turns the owner-review packet into a smaller action gate so the next agent can move quickly without promoting a manifesto, tutorial, theorem, guide, book, or benchmark too early. Priority rule: prefer the smallest move that reduces restart ambiguity or selects one P2 branch. If a proposed action requires coordinating multiple public surfaces, theorem claims, tutorial chapters, guide/book prose, or benchmark infrastructure, fold it back into Strategic Deferral unless it is explicitly accepted as the one P2 branch. Anti-repeat rule: if a proposed next action only restates the same F1 gate, adds another status/audit/restart surface, adds another copy affordance, or rechecks already verified projection plumbing, stop unless it fixes a concrete reviewer-visible blocker. Record any useful remainder as a seed, then return to collecting the real F1 owner-or-reader response. Read first: ```text README.md FOUR_ENTRANCES_READER_FEEDBACK.md F1_READER_REQUEST_2026-06-26.md paper/arxiv-hcgm-regression/README.md toy/TEACHING_CONTRASTS.md ``` For durable entrance context, use `entrances/`. For owner-review, dry-run, verification, and P1 promotion evidence, use `_archive/phase-2026-06-26/`. P0 has already passed and P1 has already been promoted. For the fastest next owner pass, use the P2 Fast Decision Table below. ## Current Default Current state after owner approval: ```text P0 defaults accepted: yes P1 paper motivation paragraph promoted into main.tex: yes P2 branch selected: teaching card deck F1 reader-feedback gate created: yes F1 shared-spine check exposed: yes F1 agent dry-run prepared: yes F1 reader request packet prepared: yes F1 fast owner reply path exposed: yes F1 review snapshot exposed: HCGM-F1-2026-06-26-review-v1 F1 owner-or-reader pass recorded: no P2 fast-path signal: mainline first; seed-bypass nonessential ripple Causal AI book assimilation ledgers: added under entrances; folded unless F1 or owner review selects a receiving branch ``` Until the owner changes it, keep this as the default next-grow order: ```text P0. Done: owner-review first-choice wording, public-strength boundary, teaching contrast order, and next grow surface. P1. Done: exactly one paper motivation paragraph seed from paper/arxiv-hcgm-regression/README.md was promoted into main.tex. P2. Selected fast path: keep the minimal entry-site teaching-card deck live and easy to read, starting with "Factual evidence vs target condition" and "Aggregate trend vs target unit distribution." Keep calibration theorem, public essay, benchmark, and guide/book growth folded. F1. Current next move: collect one compact reader-feedback pass across propagation, teaching, theory, and authority using the copy-paste `F1 reader feedback` template, or the shorter `F1 fast owner reply` if the owner is reviewing, before choosing any further branch. The response must include a shared-spine signal: do the four entrances feel like one HCGM-first argument? The response should cite `HCGM-F1-2026-06-26-review-v1`. The agent dry-run and request packet are only preflight / collection infrastructure and do not count as this pass. ``` Recorded P0 one-line approval phrase: ```text P0 defaults accepted. ``` Treat the same phrase without the final period as equivalent: ```text P0 defaults accepted ``` That punctuation-insensitive phrase opened only the P1 paper motivation paragraph path. It did not approve manifesto, full tutorial, guide/book, benchmark, shared package, broad learnability theorem, first calibration theorem promotion, or any P2 branch. ## P2 Fast-Path Selection Selected branch: ```text P2 teaching-card deck ``` Why this is the mainline grow: ```text It turns the already-approved reader order and existing toy evidence into a minimal reader-facing teaching deck, without promoting a theorem, public manifesto, guide/book chapter, benchmark, or shared package. It also gives the propagation and authority entrances concrete reader-facing anchors before broader public prose. ``` Owner fast-path signal: ```text 先完成主线，有些事情可以适当的 seed 绕过。 ``` Operational rule: ```text Keep the first two teaching cards visible, run the reader-feedback gate, verify the live projection, and seed-bypass nonessential audit/hash churn. Use the copy-paste `F1 reader feedback` packet, or the shorter `F1 fast owner reply` for owner review, instead of opening a broader survey. Both routes should answer the shared-spine check before selecting a branch. Read `_archive/phase-2026-06-26/F1_AGENT_DRY_RUN_2026-06-26.md` as preflight only when useful, then collect a real owner-or-reader response using `F1_READER_REQUEST_2026-06-26.md`. ``` This approves only a minimal teaching deck derived from `toy/TEACHING_CONTRASTS.md`. First cards: ```text Factual evidence vs target condition Aggregate trend vs target unit distribution ``` It does not approve a full tutorial chapter, large teaching section, public manifesto, guide/book prose, benchmark package, theorem promotion, or general website rebuild. ## P0 Defaults Recommended wording: ```text HCGM is the first modeling grammar to write down for causal modeling problems that require heterogeneous units, counterfactual prediction, mechanism-level reasoning, and explicit uncertainty over latent units. ``` Recommended boundary: ```text first modeling grammar to write down, not universal replacement ``` Recommended reader order: ```text ordinary prediction -> latent/world model -> graph / SCM reasoning -> selection and observed samples -> potential outcomes / estimands ``` Recommended next grow: ```text paper motivation paragraph before entry-site teaching cards ``` ## Decision Table | Entrance | Source to inspect | Allowed now | Grow only after | |---|---|---|---| | Propagation | `entrances/WHY_HCGM_FIRST.md` | tighten one sentence or boundary | P0 wording accepted | | Teaching | `toy/TEACHING_PATH.md`, `toy/TEACHING_CONTRASTS.md` | refine reader order or toy mapping | P0 reader order accepted | | Theory | `entrances/THEORY_ENTRANCE.md`, `paper/arxiv-hcgm-regression/theorem-targets.md` | choose theorem shape as a seed | exact target chosen and owner/paper checked | | Authority | `entrances/AUTHORITY_ENTRANCE.md` | refine claim ladder or promotion gate | stable wording plus theory/toy anchors | ## Strategic Deferral Keep these as seeds for now: ```text public manifesto full tutorial chapter guide / book chapter real-world benchmark general learnability theorem shared benchmark package project-wide website package non-Cauchy or nonlinear learnability generalization full causal-mechanism recovery claim ``` These directions matter, but they are not the current high-leverage move. ## P2 Branch Seeds These are seed-folded branches, not active tasks. Keep them here so the four- entrance ambition remains visible without forcing simultaneous growth. P1 is now promoted, and the first selected P2 branch is the teaching-card deck; keep any further branch folded unless the owner explicitly changes the gate. | Branch | Smallest future grow | Why deferred now | Trigger | |---|---|---|---| | Propagation | compress `entrances/WHY_HCGM_FIRST.md` + the crosswalk into a short public essay | public prose should not become the default before one teaching anchor is tested | owner chooses propagation branch | | Teaching | keep the first two-card entry-site deck small and collect feedback before expansion | teaching should stay a small deck until it proves useful | selected fast path; live projection exists; feedback gate created | | Theory | choose one calibration theorem target from `theorem-targets.md` as the next paper grow | broad learnability and mechanism recovery are still non-claims | reader feedback asks for sharper proof assumptions and owner chooses theory | | Authority | review the visible authority preface panel before growing longer authority prose | authority should synthesize stable teaching/theory evidence, not outrun it | reader feedback on the compact preface is stable and owner chooses authority | | Benchmark | seed a minimal benchmark question, not a benchmark package | prediction lift and robustness evidence are not established | reader feedback asks for predictive or robustness evidence and a concrete evaluation question is chosen | | Guide / book | outline one guide/book preface section from promoted claims only | old guide/book material is provenance, not independent-project authority | feedback says propagation, teaching, theory, and authority are clear enough for a chapter path | If a future task tries to start more than one branch, record the second branch as a seed here and keep the active grow narrow. ## P2 Branch Exit Checks Whichever P2 branch is chosen, it is complete only when the small grow has all of these: ```text source artifact evidence anchor non-claim boundary restart note projection update only if the public entry needs to change ``` Branch-specific exit checks: | Branch | Exit check | |---|---| | Propagation | the short essay points to the claim-to-evidence crosswalk and states the not-universal-replacement boundary | | Teaching | the teaching card names one comparator, one toy, one HCGM question, and one misconception it prevents | | Theory | the theorem target names assumptions, proof status, exact source file, and claims still outside the theorem | | Authority | the preface seed cites stable propagation, teaching, and theory anchors before making authority prose | | Benchmark | the benchmark question names the target claim, baseline, metric, dataset/synthetic generator, and non-claim | | Guide / book | the preface outline uses only promoted or explicitly seeded claims, and marks provenance from old `causal-mind/hcgm/` material | If a branch cannot satisfy its exit check, leave it as a seed and do not promote it as a completed P2 grow. ## Promotion Evidence If P1 happens, the evidence should include: ```text _archive/phase-2026-06-26/FOUR_ENTRANCES_OWNER_REVIEW.md outcome slots filled or explicitly referenced _archive/phase-2026-06-26/P1_PAPER_MOTIVATION_PROMOTION.md followed without strengthening the claim paper/arxiv-hcgm-regression/main.tex changed only for the motivation paragraph paper/arxiv-hcgm-regression/main.pdf rebuilt _archive/phase-2026-06-26/PHASE_VERIFICATION_2026-06-26.md updated with the new PDF/source hash entry-site projection updated only if the public page needs to reflect P1 ``` If the next grow is not P1, record why here first. The acceptable reason should be sharper than "it seemed useful"; it should explain which entrance would be damaged or stalled by doing P1 first. ## Current Non-Decision This file does not decide: ```text the first calibration theorem shape which guide/book material to promote whether to create a benchmark package whether to create a full HCGM website whether HCGM has already achieved public authority ``` Those remain open seeds. ## Post-Book Ripple The Causal AI book assimilation added durable entrance ledgers, not a new active branch: ```text entrances/CAUSAL_AI_BOOK_ASSIMILATION_SEED.md entrances/CAUSAL_AI_BOOK_EXAMPLE_QUESTION_BANK.md entrances/CAUSAL_AI_BOOK_NARRATIVE_TECHNIQUES.md entrances/CAUSAL_AI_BOOK_VISUAL_ASSIMILATION_LEDGER.md entrances/OWNER_REVIEW_ENTRANCES_2026-06-29.md ``` Use these only as review and receiving-surface guidance. They do not make guide/book, benchmark, theorem, teaching expansion, public essay, or slide-deck growth active. If the owner review says one of these should change the next move, record that as a real F1/owner signal before changing this gate.