Three-Paper Roadmap — the cascade, the owed experiments, the anti-monoculture plan

2026-09-22. Decided with JC: build three papers off the Compendium master, not one over-bundled submission. This sheet is the map to sign BEFORE any paper is written: for each paper — the one-sentence thesis, the single contribution, the spine, what is evidenced today, what is owed, the anti-monoculture strategy, and the target venue. Plus the shared replication / preregistration / outsourcing plan and the ship sequence.

Principle (from the focus decision): split by EVIDENCE, not by academic-vs-industry. What has an experiment behind it ships in the base; the ambitious-but-unproven becomes an extension that must earn its claim; the practice/tooling is the product. Nothing load-bearing is deleted — it is relocated to where its evidence supports it.

0. The cascade (what anchors what) 

COMPENDIUM (canonical master, superset, hypotheses marked)
   │  published as a citable tech report (arXiv / Zenodo DOI) — NOT submitted as a paper
   ├── PAPER 1  (BASE, peer-reviewed)      ── derivability + capacity-relative + rubric
   ├── PAPER 2  (EXTENSION #1)             ── cheap rigor + the revival model   (cites P1)
   └── PAPER 3  (EXTENSION #2)             ── the externalized guarantee        (cites P1)

INDUSTRIAL BODY (not journal papers): field guide + product (Companion/Ledger) +
   an experience/industry-track report (e.g. ICSE SEIP) + Substack
  • Compendium = the “academic-industrial body that links everything” JC asked for. It stays the superset; the papers are derived, focused submissions that cite it. Give it a DOI so it is citable without being over-bundled.
  • Base ships first and does not wait on any owed experiment (it is already evidenced).
  • Extensions ship on their own timeline, gated on experiments they must run/earn.

PAPER 1 — BASE (peer-reviewed submission)

Working title: Derivable Correctness: the Capacity-Relative Value of Externalized Specification.

Thesis (one sentence): A program’s correctness can be ratified by a reader from the specification, contracts, and audit trail — without reading the implementation — and the value of that externalized specification is capacity-relative: greatest where the model is weak, receding as it strengthens.

Single contribution: the derivability lens + the seven-property rubric that operationalizes it

  • the capacity-relative descriptive finding + honest bounds. (Descriptive/measured — NOT the predictive revival model, which is Paper 2.)

Spine (~12pp):

  1. Problem: the assistant does everything from one entry point; what does “correct” mean when a human never reads the implementation.
  2. Derivability (premise): correctness ratified from what is externalized; the stateless reader as the vivid case (central but underlying, not the headline).
  3. GS as the discipline: the artifacts that make a program derivable and ratifiable.
  4. The seven properties = the expanded definition of correctness (the rubric as instrument).
  5. Evaluation (methods + threats-to-validity): RX (derivability), KX (sentinel/retrieval), EX (the governance gate), AX (early wins and the pre-registered saturation null), ALX (machine-checkable oracle), SX/TX (the mechanism).
  6. The capacity-relative finding (the honest one): CR + AX2 — the value recedes as the model strengthens because the failure mode it targets stops occurring. This explains the saturation rather than hiding it — it is a strength, not a concession.
  7. Limitations, replication package, related work.

Evidenced today: RX, KX, EX, AX (incl. the null), ALX, SX/TX, CR, AX2. Fully evidenced.

Owed: none for the claim. Only the replication package (publish benchmarks, harnesses, oracles, preregistrations) and the related-work fill (specification-driven / pragmatic-tier / declarative lineage).

Anti-monoculture in the base: cross-vendor already present (AX2 = GPT/Gemini/Claude); non-memorized benchmark already present (Pastura, CR); report the null in the same voice as the positives. Ship the artifact for badge review; invite replication. (Human raters = deferred to DX2, stated as a limitation — do NOT overclaim.)

Sentinel + governance ARE here: the sentinel is evidenced (KX) as the operationalization of Bounded/derivability, stripped of mythology; the governance properties (Auditable, Defended, Executable) + EX (the gate catching 15 defects) are here as evidenced rubric properties. What is NOT here is the claim “governance is the durable value” — that is Paper 3.

Venue: IEEE (Software / Access) or an SE venue; a focused empirical/position hybrid.

Status: ~35% drafted (per ieee-strategy-calibration); spine above is the missing structure. Ready to write on sign-off.

PAPER 2 — EXTENSION #1: cheap rigor + the revival model

Working title: Cheap Rigor: a Capacity-Gated Model of Discipline Revival.

Thesis: the AI executor makes validated-but-dead-on-cost practices cheap again — but a practice revives only where defect exposure λ > 0; its revival value = f(capability gap), and its shape depends on the practice class (governance = flat/durable, defect-catching = capability-hump, structural = receding). A predictive portfolio model selects which practices to revive for a given project and what each buys.

Single contribution: the class characterization (validated-but-dead-on-cost) + the predictive model + its validation + the surprising negatives (which stay dead and why). Not “AI is cheap.”

Evidenced today (as instances, not proof): Loom/ALX (the extreme instance — formal spec → compiler), NX k=1 (clean negative → revival is capacity-gated, not cost-gated; N-version surfaces generator uncertainty), CR (a revival datum, receding class).

Owed (the scientific risk — this is the reason to do the experiments):

  1. NX k≥3 on a weak model (qwen ladder) + harder problems so λ > 0 even on the frontier — the test of whether N-version revives where the frontier does not.
  2. A 2nd/3rd practice (mutation testing, PBR, formal property spec) shown to revive and to stay dead, the model predicting both.
  3. Model-validation: does predicted revival correlate with measured benefit across (practice × project) cells? Bootstrap from the ledger; prospective via chronicle-ledger.
  4. The automatable revival tool (revival_v0.cjs → point at a GS-ready project, get the actionable portfolio out).

Anti-monoculture = the outsourcing engine: this is the paper to run as a Registered Report (protocol frozen, in-principle acceptance, null publishes) and to ship the benchmark + harness + oracle as a first-class artifact with a call for replication. Every independent replication cites us AND removes the proponent-authored critique. Recruit runners via St. Thomas capstones, Gabriel/BYU. We run the first honest pass; others generalize.

Venue: Registered Report track (ESEM / an RR-accepting SE venue), or empirical SE.

Status: model + ledger drafted; experiments designed, mostly not run. Gated on the owed runs.

PAPER 3 — EXTENSION #2: the externalized guarantee (the durable core)

Working title: The Externalized Guarantee: what endures when the model does everything.

Thesis: in the single-entry-point world, the value that outlives any model is the guardrail kept OUTSIDE the model (an LLM cannot be its own trustworthy verifier) + the regenerable spec imprint (code as residue). This is where GS evolved — the governance layer as the durable claim.

Single contribution: the durable-core argument, made falsifiable by a gating experiment.

Evidenced today: EX (the external gate caught 15 defects against a live system) — an existence proof, not a general claim.

The honest tension (why this is the riskiest paper): RND-1 currently nulls the independent-verification arm under strong models — a strong model catches much of what an external gate would. So the claim is not yet earned. The paper’s whole job is the experiment that resolves it:

  • Owed — the gating experiment: show under what conditions an external non-LLM gate catches what model self-verification misses (stakes, model strength, defect class). If it nulls broadly, this becomes an honest experience/position report (“here is where the external guarantee still pays, and where the model absorbed it”), NOT a defeated claim.

Anti-monoculture: same engine — preregister the gating protocol; publish the harness; invite replication. The null is publishable and honest either way.

Venue: governance/assurance or SE-in-practice track; possibly an experience report if it nulls.

Status: thesis clear, experiment un-run, claim currently un-evidenced. Lowest priority, highest risk — write last.

The shared plan — replication, preregistration, outsourcing (the anti-monoculture spine)

The three critiques are separate problems with separate fixes:

Sub-problem Fix Who does it
Proponent-authored Pre-registration (protocol frozen before data) + third-party replication Us (design) → neutrals (replicate)
Single-benchmark 2-3 problem families + non-memorized (Pastura) + cross-vendor (AX2) Us
No human raters DX2 redesigned (independent raters/control) Us — cannot be outsourced

The outsourcing engine (JC’s idea, made concrete): publish each experiment’s benchmark + harness + oracle + preregistration as a first-class research artifact, run it as a Registered Report where possible, and issue a call for replication. Benchmark/dataset artifacts are among the most-cited SE outputs; every independent run cites us and simultaneously dissolves the proponent-authored critique. Recruitment channels already in hand: St. Thomas capstones (adjunct route), Gabriel / BYU networks, the biologist co-author (cross-discipline eyes), and the stateless external judges as an automated complement (never a substitute for human raters).

The honest boundary: replication by others takes time — the base cannot wait on it, and the human-rater arm (DX2) we must run ourselves. So: base ships now with a full replication package; extensions run their first honest pass by us + open the terrain for others.

Ship sequence

  1. Now: write Paper 1 (base) spine → JC sign-off → full draft. Ship the replication package. Re-label the Compendium as internal master; register it (arXiv/Zenodo) for a citable DOI.
  2. In parallel: run the owed Paper 2 experiments (NX k≥3 weak + λ>0; a 2nd practice), preregister them, publish the benchmark-as-artifact with a call for replication.
  3. After the base lands and P2 experiments read out: draft Paper 2.
  4. Last / gated on the gating experiment: Paper 3 — as a full claim if the experiment supports it, as an honest experience report if it nulls.

Decision requested

Sign off on: (a) the three-paper cascade + Compendium-as-citable-master; (b) Paper 1 as the base to write now (fully evidenced, sentinel + governance-properties + EX included, durable-core CLAIM excluded); (c) Paper 2/3 as extensions that must earn their claims via the owed experiments; (d) the anti-monoculture engine (preregister + benchmark-as-artifact + call for replication + DX2 for human raters). On sign-off I write the Paper 1 spine for your signature.