Generative Specification White Paper — Decision Record

This file is the ADR (Architecture Decision Record) for the white paper itself. It applies GS’s own discipline to the artifact that defines it: every structural decision, critique, and deliberate preservation is recorded here so that future revision sessions start with full context rather than reconstructing it from git blame.

Rule: If a structural change is made to the paper — framing, section placement, content added or removed — an entry goes here. Commit messages say what changed; this file says why, what was considered and rejected, and what must not be undone without understanding why it was done.

Format 

## [Date] — [Short label]
**Trigger:** what prompted the change
**Decision:** what was changed and why
**Preserved:** content explicitly kept despite pressure to cut or change
**Rejected options:** alternatives considered and not taken
**Open:** unresolved questions flagged for the next round

Late March 2026 — Initial submission: proto-v1.0 (e42ee53)

Trigger: First arXiv-targeted draft integrating AX adversarial experiment results.

Decision: The paper opened cold with the theoretical apparatus — Morris semiotic tripartition in the abstract, Chomsky structural analogy in §4.1.a, seven properties immediately. No narrative warm-up. The argument sequence was: abstraction ladder → theoretical gap → GS principle → artifact grammar → six case studies → experiments → implications → conclusion.

What was already settled and never changed afterward:

  • The seven property names: Self-describing, Bounded, Verifiable, Defended, Auditable, Composable, Executable. All seven names present and final in v1.0.
  • The Martin-sense “paradigm” disambiguation (not Kuhn’s sociological sense).
  • The Morris (1938) attribution for the semiotic tripartition.
  • The Chomsky analogy framed explicitly as structural, not formal equivalence.
  • ForgeCraft conflict-of-interest disclosure.

What was missing:

  • No prologue. The paper opened with theory.
  • Abstract stated thesis rather than promising empowerment.
  • Related Work was §5 (after the solution) rather than §3.5 (before it).
  • Seven properties lacked per-property failure-mode grounding.
  • The formula was I(S) ≈ 1/S (dimensionally weak — S=1 gives I=0 but S=0 does not give I=∞ cleanly).
  • “Community convergence theorem” — not yet renamed to “principle.”
  • SlopCodeBench (Orlanski et al.) not yet cited.
  • §8.13 The Engineer Elevated did not exist.
  • §8.14 The Adoption Ladder did not exist.

Late March 2026 — Adversarial audit fixes: v1.1 (b10903b)

Trigger: Expert philologist review. Multiple linguistic attack surfaces identified in the Chomsky analogy and related vocabulary.

Decisions:

  1. Chomsky correction: GS’s term “valid” is a deliberate extension of Chomsky’s “grammatical” — not a misappropriation. Disambiguated explicitly with a note that the extension is intentional.
  2. “Generative capacity” → “derivation precision” throughout. The prior term overloaded Chomsky’s formal vocabulary.
  3. “Community convergence theorem” → “community convergence principle.” “Theorem” implies formal proof; “principle” is the correct epistemic weight.
  4. Analogy-mapping table added to §4.1.a to make the Chomsky structural parallel explicit and bounded.
  5. Plain language summary added after the abstract — the Jacquard loom / foundry historical arc from Babbage to present.
  6. §8.12 “The Application Gate” added.
  7. §9.5 “Practitioner Path” added.
  8. Mutation testing framing corrected: v1.0 incorrectly called line coverage a “hallucinated” metric. Fixed.

Preserved: The entire seven-property structure, the Morris/Martin/Chomsky theoretical apparatus, the six case studies as existence proofs.

Persistent artifact introduced: The §9.x Discussion sections (§9.1–§9.5) appear without a parent ## 9. Discussion heading. This numbering artifact persists through all subsequent versions and causes ToC generation issues.

Late March / Early April 2026 — §4.2 subheadings + abstract compression (6c66f24)

Trigger: Internal structural review — §4.2 was a dense unbroken prose block.

Decisions:

  1. Four named subheadings added to §4.2 (The Three-Tier Taxonomy): “The Syntactic and Semantic Tiers,” “The Pragmatic Tier,” “Prior Occupants of the Pragmatic Surface,” “The Derivability Obligation.”
  2. Abstract compressed — phrasing tightened in the paradigm definition paragraph.
  3. “Derivability obligation” given its own named subsection — the concept is now retrievable by a stateless reader without reading the full §4.2.

Preserved: All §4.2 content. The subheadings are organizational overlays, not restructuring.

Open: The §9 header absence noted but not fixed in this commit.

Early April 2026 — v1.3 Narrative restructure for approachability (b9cf29c)

Trigger: Approachability failure — the paper opened cold with theory. Reviewers unfamiliar with the domain were unable to locate the human problem the paper solved. Winston’s “5 S’s” rhetorical structure applied.

Major decisions:

  1. Prologue added: “The Week That Should Not Have Taken a Week.” A personal narrative about OAuth/DNS configuration consuming four days. The paper’s first non-technical entry point.

  2. Abstract rewritten as empowerment promise. The opening sentence changed from “The ladder has been moving…” to “You will finish this paper able to build any software system you can describe with precision, without writing a line of code yourself.” This framing was locked here and never subsequently changed.

  3. Related Work moved from §5 → §3.5 (before the paradigm claim rather than after). Independent corroboration of the problem now precedes the solution.

  4. Seven properties derivation added: Each property now grounded by naming the specific failure mode it prevents. The canonical framing: “not a taxonomy constructed in advance, but a record of what breaks and why.”

  5. SlopCodeBench added (Orlanski et al., arXiv:2603.24755) as a third independent validation thread in §3.5. This is the first version where the paper cites an independent empirical finding that structural deterioration rises in 80% of AI agent trajectories — directly corroborating GS’s anomaly claim.

  6. Context compaction added as the mechanistic link between SlopCodeBench’s finding and GS’s response.

  7. CodeSeeker named explicitly in §6 tooling.

  8. Accessibility pass (Nadjet review): Martin’s three paradigms explained with significance; ADR, CLAUDE.md, sentinel, Chomsky hierarchy, all experiment acronyms defined at first use.

  9. §8.6.1 “The API-First Future” added.

  10. §8.13 “The Engineer Elevated” added.

  11. §8.14 “The Adoption Ladder” added.

  12. §11 “Onwards” added as closing section.

Preserved: OAuth/DNS story moved to §8.6.1 rather than cut.

Rejected options: Opening directly with the Orbiter story (considered; rejected because it was more abstract than the OAuth/DNS personal narrative). The OAuth/DNS story was the first prologue choice.

Structural debt introduced: Related Work now appears in two locations — §3.5 (problem-side corroboration) and §5 (relationship to prior principles). This split is intentional but creates a navigation burden that must be managed.

April 5, 2026 — Prologue replacement: v1.2/v1.3 (309ec58)

Trigger: Academic pre-publication scrutiny. The OAuth/DNS prologue was personal but not structurally load-bearing. A reviewer noted that the paper needed an opening that was simultaneously accessible and directly demonstrative of the paradigm claim.

Decision: OAuth/DNS → Mars Climate Orbiter.

The Orbiter (1999, $327.6M, pound-force vs. newton-seconds) replaced the personal week-of-pain narrative as the prologue. Key structural argument: the Orbiter failure was a specification boundary failure, not an implementation failure. The same class of failure is produced at AI generation speed with no new mechanisms required.

The OAuth/DNS story was moved to §8.6.1, not cut.

Why the Orbiter is better:

  • It is historically documented, attributed, and verifiable.
  • It names the failure class (implicit unit convention = implicit context) precisely.
  • It scales: a $327M government contract is legible to any audience.
  • It is structurally integrated with the paradigm claim rather than narratively adjacent to it.

Preserved: The empowerment promise abstract (unchanged). The OAuth/DNS story (moved, not cut).

Rejected options: Keeping the OAuth/DNS prologue. The trade was narrative warmth against structural argument. The Orbiter won on structural grounds.

Note for future revisions: Some readers found the OAuth/DNS prologue more personally relatable. If the target audience shifts from academic (arXiv) to practitioner (workshop, Substack), a personal prologue may be more appropriate. The OAuth/DNS story in §8.6.1 is available for that recontextualization.

April 5, 2026 — Restriction as activation mechanism: v1.3 reframe (d549bf6)

Trigger: Recognition that the paper’s treatment of restriction was purely negative — restriction as constraint. A deeper structural insight was not yet named.

Decision: Added to §4.1.a: “The act of ruling things out is also the act of activating what was always present.”

The AI’s training corpus contains the full formal tradition — Hoare logic, type theory, design-by-contract, REST, deontic logic, domain-specific vocabularies. Naming these constructs in a specification does not teach the AI; it unlocks what the model already holds. The restriction is the activation mechanism, not merely a discipline.

Figure 1 added: The Generative Specification Domain Stack — two-axis table showing Chomsky’s upward generative reach against Martin’s downward restriction, with the full tier stack from syntax to business/ethics. First visual element in the paper.

Significance: This reframe changes the paper’s second-order claim. Beyond “GS is a methodology,” it now claims “GS is the mechanism by which formal specification traditions become accessible without specialist training.” This is a stronger and more provocative claim.

Preserved: The restriction framing in all prior sections (restriction as discipline, restriction as liberation). The new framing adds a dimension rather than replacing.

April 2026 — Name all novel concepts, enforce pillar hierarchy (3f09b85)

Trigger: Reviewer critique: the paper introduced novel concepts (coined terms) without flagging them explicitly. A reader could not distinguish GS’s own vocabulary from borrowed vocabulary.

Decisions:

  1. All coined terms flagged with “coined here” + nearest established concept:
    • “Architectural constitution” → coined here; maps to ISO/IEC 42010
    • “Phase-collapse” → coined here; no prior exact term
    • “Hardening surface” → coined here; maps to attack surface in security literature
    • “Derivability obligation” → coined here; nearest: specification completeness (Parnas 1972)
  2. Each property definition expanded with:
    • Automatable checks note
    • Explicit prior literature citations (Parnas, Conway, Beck, Jia & Harman, ISO/IEC)
  3. Bounded property: mechanical justification for 300-line limit linked to AI tool read budgets.
  4. Verifiable property: write-completion ≠ compile-success argument added.
  5. Formula updated: I(S) ≈ 1/S → I ∝ (1-S)/S. The new form is dimensionally correct (S=1 gives I=0; S=0 gives I=∞). No explicit decision note in the commit; the change was made in passing. This note is its record.
  6. §9.4 renamed: “From Theoretical Claim to Running Instrument” → “The Convergence Spiral: Expected Iterations as a Function of Specification Completeness.”

Preserved: All seven property names (no renames). The formula shape (only precision-corrected).

Note for future revisions: “From Theoretical Claim to Running Instrument” was a more descriptive title for §9.4. “The Convergence Spiral” is more evocative but less precise about what the formula actually does. If the section is ever extracted as a standalone technical appendix, restore the original name.

April 13, 2026 — Contract Sufficiency + adversarial critique pass (f9f13f1)

Trigger: Final adversarial critique pass before arXiv submission.

Decisions:

  1. §4.4 “Contract Sufficiency: The What-How Distinction” added. Closes a gap identified in critique: the paper needed a principled account of where the specification obligation ends and implementation discretion begins. The spec states what must be true; it does not prescribe how the AI achieves it. Promiscuous specification (over-specifying how) is a distinct failure mode.

  2. §8.2.2 “The Industrial Threshold” added.

  3. AX study: seven → eight conditions (one more post-hoc condition from iterative gap-closing).

Preserved: The pre-registration framing for the AX experiment series.

April 14–15, 2026 — Six structural fixes + three content restorations (session: 74db953f)

Trigger: Multi-round adversarial critique session. Six vulnerabilities identified in the then-current v1.4. Three content losses identified by auditing v1.0 and v1.1 against current state.

Six structural fixes:

Fix 1 — Abstract: Kuhn/Martin framing separated. Prior: The abstract conflated Martin’s structural criterion (what defines a paradigm) with Kuhn’s sociological criterion (when a community adopts one). These are different claims with different falsifiability profiles. Decision: Rewritten to keep them distinct. “The Martin-sense structural case is complete is the claim this paper advances.” Whether the Kuhnian transition occurs is left explicitly as a community determination question.

Fix 2 — §4 opening: GS removal differs in kind from predecessors. Prior: GS’s removal was listed alongside structured programming, OOP, FP as a member of the same category. Decision: Added paragraph making the categorical difference explicit: prior paradigms made implicit context inconvenient (human readers compensate); GS makes it structurally absent because the reader that would have compensated does not exist.

Fix 3 — §4 opening: Enforcement objection answered at the top of §4. Prior: The answer to “the compiler knows nothing about the spec” was buried in §4.3. Decision: Moved to §4 opening as “On enforcement.” paragraph. The gate stack (commit hooks, CI, MCP boundaries) IS the compiler for GS’s restriction class. Removed from §4.3 to avoid duplication.

Fix 4 — Tier table: Evidence column corrected. Prior: T3 cited “Chronicle/Railway, §7.7” but §7.7 is the COMPASS ETL case, not Chronicle/Railway. Decision: T1 → “Demonstrated: production cases + AX + ALX”; T2 → “Demonstrated: AX”; T3 → “Demonstrated: COMPASS ETL (§7.7)”; T4 → “Demonstrated: COMPASS/The Eye (§7.7)”. The word “Proven” was replaced throughout with “Demonstrated” — the epistemic claim is calibrated to what the evidence actually supports.

Fix 5 — §3: Auto Dream promoted from footnote to body. Prior: Auto Dream was a footnote. Decision: Promoted to §3 body as a complementary approach. Auto Dream (bottom-up, reactive consolidation) vs. GS (top-down, preventive specification) are not competing. Both address the stateless reader problem at different layers.

Fix 6 — Orphaned footnotes removed. Three footnote definitions ([^1], [^2], [^6]) remained in the file after their inline references were removed during prior condensation. Pandoc generated warnings. [^6] content promoted to §3 body (Auto Dream paragraph). [^1] and [^2] removed. Second pandoc run: zero warnings.

Three content restorations:

Restoration 1 — §8.13: Synthetic/synaptic/synoptic triad. Present in v1.0 and v1.1. Lost during narrative restructure (b9cf29c). The triad names three properties of cross-domain specification skill: Synthetic (combining domains produces insight neither holds alone), Synaptic (the productive surface is the boundary between domains), Synoptic (holding multiple domains simultaneously enables pattern recognition rotation cannot achieve). Restored in full.

Restoration 2 — Provenance section. Present in v1.0. Lost during structural hardening. Records the paper’s own authorship process: the author named the doors; the AI supplied the contents of the rooms. The Blavatsky inversion. The blind-session rejection of the Austin speech act framing. Restored between Acknowledgements and References.

Restoration 3 — About the Author. Present in v1.0. Lost during structural hardening. Brief biographical note including: “The first professional line of code was written in 2006. The last one typed by hand was written sometime before June 2025.” Restored at end of document.

April 16, 2026 — §5 disciplines table + §4.1.f harness/cascade + §7 reframe + §8.13 evolution (this session)

Trigger: User review identified four structural gaps:

  1. Structural disciplines described in prose; no table; acronyms unexplained.
  2. “Spec is the program” claim uncaveated — harness as guarantor not stated.
  3. Six projects framed as experiments; their role as method-development substrate not distinguished from the controlled experiments.
  4. §8.13 stated that specification skill compounds for cross-domain practitioners but did not show the trajectory — the evidence for expertise accumulation.

Decisions:

§5 — Disciplines prose → table. The three-tier taxonomy (syntactic/semantic/pragmatic) existed in prose across §4.1 and §5. Converted to a ten-row table in §5 with columns: Discipline, Tier, What it is (with all acronyms spelled out), Freedom removed, GS property satisfied, What GS adds. Every acronym (SOLID, TDD, DDD, BDD, CI/CD, OOP, FP) is spelled out with a plain-language description of what the discipline does and why it exists. The prior bullet list (four GS-only contributions) was consolidated into a closing paragraph rather than kept as separate bullets — it now serves as a summary after the table rather than a repetition of it.

§4.1.f — “On the role of the harness” paragraph added. The claim “spec is the program” requires the harness to be a guarantee rather than an assertion. New paragraph states this explicitly: the spec establishes intent; the harness certifies the derivation was faithful. T2 is not optional scaffolding; it is constitutive of the GS guarantee.

§4.1.f — “On cascade refinement” paragraph added. The tier hierarchy is conceptual; the implementation loop is recursive. T3/T4 failures are detectors; T1 is almost always the site of correction. Complete NFR register at T1 is a prerequisite for T3/T4 enforceability, not a documentation exercise.

§7 intro — Two-category reframe. The §7 italic preamble now explicitly separates the six production projects (method development substrate, existence proofs — “the rubric is their residue, not their premise”) from the controlled experiments (testing phase — AX, RX, BX, EX — designed to falsify rather than illustrate). The body paragraph adds the practitioner- arc sentence with forward reference to §8.13.

§8.13 — Practitioner evolution trajectory added. The arc from heavy AI dialogue → nudge-and-correct is evidence of expertise accumulation and an argument against the structural-dependency concern. Added as paragraph between the synoptic triad and the four structural changes list.

Preserved: The prior §5 closing sentences (“necessary but not sufficient” and the four GS contributions). The synoptic/synaptic/synthetic triad (restored in prior session, not touched). The §7 structural observation about SafetyCorePro/BRAD as natural control conditions.

May 9, 2026 — Tier framework restructure: 7 tiers → 6 tiers, stage-based (this session)

Trigger: Talk in 1 week. Pre-talk audit revealed a structural ambiguity in the prior 7-tier formulation: the harness was treated as its own tier (old T2) peer to the spec, when in fact it is the cross-cutting capability that recurs at every tier with stage-appropriate tests. This obscured the symmetry that each tier removes both an authoring obligation and a verification obligation simultaneously, where the verification removal is what makes the authoring removal safe.

Decision: Restructure §4.1.f from a 7-tier activity-typed cascade to a 6-tier stage-based cascade. Mapping:

Old New Stage
T1 (write code) + T2 (read generated code / harness) T1 Development
T3 (infrastructure) T2 Staging / Pre-prod
T4 (self-monitoring) T3 Production
T5 (BIOISO) T4 Evolution
T6 (problem-stated synthesis) T5 Synthesis
T7 (meta-telos) T6 Meta-telos

Old T1 and T2 collapse because verification at the development stage IS what makes the “you do not read generated code” removal safe — the same move stated twice. Renumbering propagates: proven tiers become T1–T4 (was T1–T5); research-frontier tiers become T5–T6 (was T6–T7).

Conceptual frame added (§4 callout, §4.1.f intro): Verification is not a tier; it is a cross-cutting capability that recurs at every stage with stage-appropriate tests — dev-time at T1, staging at T2, production runtime at T3, evolution-time (mutation gauntlet) at T4, colony-level at T5. This resolves the “harness as peer rung” ambiguity that the 7-tier formulation preserved.

Sections updated:

  • §4 opening: new “On verification: cross-cutting, not separate” callout added after “On enforcement” and before §4.1.
  • §4.1.f: full restructure. Title “Seven-Tier” → “Six-Tier”; both proven and research-frontier tier tables rewritten with a new “Stage” column and the “authoring + verification” pairing made explicit; six tier descriptions rewritten; “On the role of the harness,” “tiers are not independent,” and “On cascade refinement” notes updated for the new numbering and reframed around the cross-cutting harness.
  • §4.2 opening: “seven tiers” → “six tiers” (one-word swap; the seven properties are unchanged).
  • §3.5 industry prior art: tier renumbering — Tier 1 verification harness is no longer a separate (T2) target; the staging/production/evolution references shifted T3→T2, T4→T3, T5+→T4+.
  • §7 intro: “full T1–T4 proof” → “full T1–T3 proof” (the Conduit EX project covers what was old T1–T4 = new T1–T3).
  • §7.7 case heading and three subsection headings: “(T3 + T4)” → “(T2 + T3)”; “T3: Specification-Derived Infrastructure” → “T2: …”; “T4: Specification-Governed Self-Healing” → “T3: …”; “Formal Connection Between T3 and T4” → “Between T2 and T3”; body paragraphs in §7.7.3 renumbered consistently.

Reference counts: ~40 individual tier references updated across the paper. The Loom mutation tag string [GS T5] → [GS T4] was updated in §4.1.f Tier 4 description. The reachable cross-references for the philosophical/civilizational frames (Nous/Logos T1–T5 → T1–T4; Golden Century T5–T7 → T4–T6; Attention is All You Have T1–T7 → T1–T6; Ambient Engineering T6–T7 → T5–T6) were updated.

Preserved: The seven specification properties (Self-describing, Bounded, Verifiable, Defended, Auditable, Composable, Executable). The closed-loop cascade structure of §4.2 (commit-type trigger surface, three-layer recording architecture, public-surface diff rule, judgment layer terminus, anti-drift formula). All §7 case study evidence and metrics.

Ambiguities encountered and how they were resolved:

  1. Loom [GS T5] mutation commit tag. The public Loom repository has been committing auto-mutations under the [GS T5] tag (old T5 = BIOISO). In the new schema, BIOISO is T4, so the description was updated to [GS T4]. This is a documentation change in the paper; whether the Loom commit tag itself is migrated in the repository is a separate operational decision that may need an addendum or a note explaining the historical-tag → new-tag mapping. Flagged for follow-up.

  2. §7 intro “T1–T4 proof”. Old T1–T4 = (write code, harness, infra, monitor) → new T1–T3 (development includes harness, then staging, then production). Translated as “T1–T3” with a parenthetical clarification of what those stages cover, since the literal numeric range changed but the coverage did not.

  3. §3.5 “Tier 1 (spec drives implementation)”. The industry SDD tools (SpecKit, OpenSpec, Kiro, Tessl) reach only the authoring half of new T1 — they do not treat the dev-time harness as constitutive. Reframed as “Tier 1’s authoring half” to preserve the original critique under the new model where T1 also implies the harness.

  4. §4.1.f “On the role of the harness”. Old text said “without a verification harness at T2 minimum is an assertion, not a guarantee.” In the new model T2 is staging, not the harness tier. Reframed to “the verification harness for the relevant tier” — preserving the load-bearing claim (harness is constitutive, not optional) while removing the tier-as-harness identification.

  5. Philosophical frames T5–T7 reference. “The Golden Century names the civilizational consequence when T5–T7 complete” — in the new schema T5–T6 are the research-frontier tiers. Translated as “T4–T6 complete” to preserve the spirit (the civilizational consequence kicks in once evolution-tier and the two research-frontier tiers are operational).

Rejected options:

  • Keeping the old harness as its own T2, just renumbering downstream tiers to span 1–6. Rejected: it would have preserved the “harness as peer rung” ambiguity the user explicitly identified as the structural debt being paid down.

  • Splitting development into T1a (write spec) / T1b (harness) sub-tiers. Rejected: the cross-cutting verification frame already captures this; a formal sub-numbering would re-introduce the very ambiguity the restructure is meant to remove.

Open for next revision:

  • Companion documents (PractitionerProtocol, GS_Experiment_Supplement) likely carry their own tier references; this session updated only the white paper. A separate sweep is needed before any companion is regenerated as PDF.
  • The Loom repo’s [GS T5] historical commit tag policy (continue under the old tag for historical continuity? rewrite to [GS T4]? ship a mapping doc?) — not a paper question but flagged because the paper now references [GS T4].

June 9, 2026 — §5.3 Field corroboration (W1: McBrokers) + Compendium/White-Paper split + review reframe

Trigger: First paying-cohort delivery (McBrokers “La Fragua,” four-day workshop, June 2–5 2026), full transcripts mined; plus a reviewer pass on the derived white paper (overall A−/B+, publishable after a short fix list).

Decision:

  1. Document split. The long document was repositioned as the Compendium (canonical, ~115 pp); a focused white paper (~8–10 pp, A Discipline of Derivability for the Stateless Reader) was derived from it for Zenodo/arXiv. The Onwards essay and conference submissions also derive from the Compendium.
  2. Title softened; paradigm demoted. Dropped “A Pragmatic Programming Paradigm” from the title; the pragmatic-tier placement is now an explicitly-labeled proposed lens, not load-bearing for the evidence, with “paradigm” scoped to Martin’s narrow sense and an explicit no-Kuhnian-claim caveat. The abstract now leads with the defended thesis and each experiment’s proof-point.
  3. §5.3 Field corroboration (W1) added as observational evidence — fenced (no control, no blind scoring, no pre-registration, single self-selected cohort) and tied to §4.3. It shows the controlled findings recur on a real team and codebase and surfaces the failure modes the controlled program then tests.
  4. Public-sourcing. Every cited exemplar was re-anchored to public repositories (the experiments + pragmaworks). ALX is correctly located in github.com/jghiringhelli/loom, not generative-specification.

Preserved (the Auditable trail — do not quietly undo):

  • The ~70% token-reduction figure is conceded UNPROVEN. It was stated to the W1 cohort, on camera, as “un número mágico… no tengo grandes pruebas.” The paper therefore does not claim a measured end-to-end token reduction: §4.3 flags the per-correct-output economy as a reasoned argument supported only by the measured KX retrieval proxy. This concession is the rationale for that hedge and must not be re-inflated into a measured claim without a powered experiment.
  • The strongest internal exemplars — a fintech decision engine (conclave) and a HIPAA data platform (lumen) — are client-confidential and deliberately not cited as public evidence.
  • The honest negatives that are the paper’s strongest credibility signal: the AX v4 regression (non-monotonic trajectory) and the conceded-unproven token-reduction figure.

Rejected options: (a) presenting W1 as controlled evidence — it is not; (b) omitting W1 — a methodology’s claim to practice is incomplete without one account of it meeting a real team; (c) keeping conclave/lumen as cited exemplars — unverifiable to readers and a client-confidentiality risk.

Open: A controlled, multi-cohort human-participant study is the path from corroboration to confirmation — and the powered test of the token economy that would let the per-session figure be claimed rather than reasoned.

July 29, 2026 — Controlled practitioner study removed from public materials

Trigger: Confidentiality. The controlled practitioner study was conducted at the author’s employer and cannot be published.

Decision: The controlled practitioner study was removed from all public materials (Compendium, white paper, field guide, two-pager, book drafts, supplement, and this record). Its data is preserved in a private archive. Practitioner-transfer evidence is now carried by the observational McBrokers field corroboration (a paying-client workshop cohort, June 2026), which remains in §7.8.A of the Compendium and is explicitly fenced as observational — no control, no blind scoring, no pre-registration. Arguments that had leaned on the removed study (the guidance-circularity closure, the layered-validation triad, the transferability claim, the tier-demonstration citations) were rewired to the remaining evidence — the six production projects, AX, ALX, BX, RX — plus the observational corroboration, with a controlled human-participant study now noted throughout as future work.

Preserved: All claims and numbers of the remaining experiments (AX, EX, KX, ALX, RX, BX, MX, RND-1). The observational McBrokers corroboration. No replacement numbers were fabricated; unsupported claims were softened rather than re-sourced.

Open questions for next revision (v3.0)

  1. §7.8 controlled human-participant study — When a controlled, pre-registered, multi-cohort study is available, integrate its results, update pre-registered prediction outcomes, and update the follow-up-studies section.

  2. §9 header — Verify ## 9. Extended Applications and Technical Notes (or equivalent) exists before §9.1. The missing §9 header is a structural artifact present since v1.1 and has never been explicitly fixed.

  3. Formal/informal discipline naming pass — The table in §5 uses “Syntactic” and “Semantic” as tier labels. Consider whether “structural disciplines” (informal tier) and “correctness disciplines” (Loom formal tier) should be introduced as named categories for the broader GS vocabulary. A full pass through the paper would be needed.

  4. DDD/ORM roadmap — DDD and ORM as informal restriction layers for Loom. These were identified as relevant to the ForgeCraft roadmap but not yet in the paper.

  5. Provenance section meta-note — Consider a footnote or sidebar noting that the paper itself is now governed by the same artifact discipline it describes: PAPER-DECISIONS.md is its decision record; critique rounds are its harness. This is a meta-demonstration worth one sentence in the Provenance section.

  6. NFR → T3/T4 cascade insight — The insight that T3/T4 failures trace to T1 NFR incompleteness is now in §4.1.f as a paragraph. If a controlled study produces data on this (teams that had complete vs. incomplete NFRs at T1 and their T3 failure rates), promote the paragraph to a subsection with evidence.

  7. The formula title (§9.4) — “The Convergence Spiral” is evocative but less precise than the v1.0 title “From Theoretical Claim to Running Instrument.” If §9.4 is ever extracted as a standalone technical appendix, restore the original.