The GS Paper Tree — one base, focused derivatives, one superset

2026-09-27. Decision (JC): build a tree of papers — one BASE, others DERIVED from it, each with a clear focus and a clear boundary of what it says and what it does not. This is fine — and correct — as long as the Compendium holds everything as the coherent superset. Refines THREE-PAPER-ROADMAP.md with concrete say/don’t-say boundaries taken from the actual GenerativeSpecification_WhitePaper.md content inventory.

The principle

Compendium = the master superset (integrity anchor). Everything lives there, cross-consistent, hypotheses marked. Not submitted. It is what makes multiple papers safe: each paper is a cut from one coherent body, so they cannot drift apart.
Each paper = ONE clear focus, cites the base, and explicitly does NOT carry claims that belong to a sibling. The over-bundling the blind reviews punished was exactly one paper carrying five focuses at once.
Derivatives earn their claim (their own experiments/argument); they don’t get the base’s evidence for free.

The tree

BASE — “Derivable Correctness: the stateless-reader discipline and its rubric”

The method, why it works, and the evidence it produces software a reader can ratify from outside the implementation — with honest capacity-relative bounds.

SAYS: the derivability obligation (stateless reader) · the seven-property rubric (the operationalization) · the three mechanisms — bridge + asymmetry, sentinel (bounded context), phase collapse · capacity-relative scaffolding (value highest where the model is weak, recedes as it strengthens) · the spec-as-contract root (#3, the abstraction-chain cap: RFC 2119/SLA form) · honest bounds.
Evidence: AX (structure), KX (retrieval/sentinel), EX (production), RX (reproducible), TX/SX (mechanism), ALX (formal tier), BX (rubric author-independence); MX/RND-1 as supporting.
Does NOT say: “governance is the durable value” (→ D1) · cheap-rigor/revival (→ D2) · the pragmatic-tier / paradigm-of-removal philosophy (→ D3) · the 21-term coined lexicon (cut; keep 2-3 load-bearing terms in prose) · the bolded promise “you will build correct systems” (reconcile to assurance/governable process, matching its own §4.1 concession).
Source: cut down from the current WhitePaper.md (remove the three exiled focuses + the lexicon + the overclaim). This is the real submission (~12-15pp).

D1 — “The Externalized Guarantee: what endures under strong models” (governance / durable core)

SAYS: in the single-entry-point world the durable value is the guarantee kept OUTSIDE the model (an LLM can’t be its own trustworthy verifier) + the regenerable spec imprint; the trust side does not recede. This is the academic home of the governance-product thesis.
Evidence OWED: the gating experiment (does an external non-LLM gate catch what strong-model self-verification misses, and under what conditions). Currently nulled by RND-1 under strong models — so this is the highest-risk paper; if it nulls broadly it becomes an honest experience/position report, not a defeated claim.
Does NOT say: the rubric/derivability framework (that’s the base it cites).

D2 — “Cheap Rigor: a capacity-gated model of discipline revival”

SAYS: the cheap executor revives validated-but-dead-on-cost practices; revival is exposure-gated (λ), not cost-gated; the shape taxonomy per practice class; the predictive portfolio model.
Evidence OWED: the pre-registered revival program (experiments/revival/PREREGISTRATION.md) — NX at higher k on the weak-model ladder, a 2nd/3rd practice, model-validation on the held-out fold.
Does NOT say: anything the base already carries; it uses the base’s capacity-relative finding as its starting point and turns it predictive.

D3 (optional) — “The Pragmatic Tier: a semiotic placement for AI-era programming discipline”

SAYS: the pragmatic-tier placement (Morris), the paradigm-of-removal framing (Martin’s sense), the theoretical home. A position/conceptual paper — no empirical burden.
Why separate: pulling the grand framing OUT of the base removes exactly the “grandiose framing” the novelty reviewers punished from the empirical paper, and lets it stand or fall as a position piece. Writable now (conceptual); lowest priority.

The BASE core (locked, JC 2026-09-28)

The base’s minimum sufficient core = the three load-bearing mechanisms (bridge, sentinel, phase collapse) + the whole substrate + the rubric. That is the core of GS. Then add the items that directly frame or bound them (below). Everything else → a derivative.

Core (must): derivability / stateless reader (the premise) · the substrate (spec + cascade
- sentinel + hooks/gates + verification, with the contract-spec root, #3) · the bridge + asymmetry · the sentinel (bounded context) · phase collapse · the seven-property rubric.
Add (frames/bounds the core): capacity-relative scaffolding (the honest bound on the bridge) · read-asymmetry (the read side of the bridge) · cost inversion / “the spec is the program” (the consequence of phase collapse + substrate) · generative execution (the verify step) · tokens-per-correct-output (light).
NOT in base: governance-is-the-durable-value CLAIM (→ D1) · cheap-rigor/revival (→ D2) · pragmatic-tier philosophy (→ D3) · the coined lexicon.

Contribution distribution

| Contribution | Paper | |—|—| | Derivability / stateless reader · substrate · bridge · sentinel · phase collapse · rubric · capacity-relative · read-asymmetry · cost inversion · generative execution · contract-spec root (#3) | BASE | | The externalized guarantee (external non-LLM verifier + regenerable imprint is the durable value) | D1 | | Cheap rigor + the capacity-gated revival model + the portfolio selector | D2 | | Pragmatic-tier placement · paradigm-of-removal framing | D3 |

Experiment distribution

Rule: an experiment may be cited by more than one paper for different claims (EX, ALX), but each claim is owned by exactly one paper.

New experiments to create (JC willing to build published ones)

Prioritized by academic leverage — each addresses a blind-review weakness (monoculture: proponent-authored / single-benchmark / single-model / no-human-raters). Publish each as a pre-registered artifact with a call for replication (the anti-monoculture engine).

For the BASE (strengthen the core against the monoculture critique):

Second benchmark / domain replicating AX’s structure result off Conduit → breaks single-benchmark. HIGH leverage, most needed.
Capability ladder for the capacity-relative claim (multi-rung, not n=2) → strengthens TX/capacity-relative. Can share infrastructure with the D2 revival ladder. HIGH.
Human-rater study (DX2 redesigned) for the rubric → breaks no-human-raters. HIGH; the one that cannot be automated (needs disinterested raters).

For D1: the gating experiment — required; the paper does not exist without it. For D2: the revival grid — already pre-registered (experiments/revival/).

Boundary rules (how to decide where content goes)

Evidenced now → base. Ambitious-but-unproven → the derivative that must earn it. Philosophy → D3.
A claim lives in exactly one paper. If two papers want it, it belongs to the more specific one and the other cites it.
Coinages: keep only the 2-3 load-bearing terms (stateless reader, and the mechanism names) in prose; no lexicon table. New coinages → the Compendium glossary, not the papers.
The Compendium carries the union, with each hypothesis marked and each paper’s scope noted, so the papers never contradict.

Sequencing

Now: cut the BASE from WhitePaper.md (remove D1/D2/D3 material + lexicon, reconcile the overclaim, keep the factual fix). Ships first — it is fully evidenced.
Parallel: run the D2 revival experiments (already pre-registered); draft D2 when they read out.
After the base lands: D1 gated on the gating experiment (or as an honest experience report); D3 whenever.
Compendium stays the superset; register it (arXiv/Zenodo DOI) as the citable master.

What happens to the current WhitePaper.md

It is the raw material, not the base. The base is a focused cut of it. Either (a) transform WhitePaper.md into the base in place (removing the exiled material), or (b) keep it as the long-form “everything” white paper (Compendium-adjacent) and create a new short base file. Recommendation: (a) — one focused base is what submits; the long-form role is already the Compendium’s.