2026-08-18

Language as Architecture — thesis v2

Status: DRAFT for operator review. v1 is docs/research/language-as-archetecture.md (filename spelling preserved as history); this file is the revised version, kept separate by operator instruction. Nothing in v1 is edited.

Abstract

v1 proposed that language, ontology, and constraints participate in software architecture, and asked for falsifiable experiments. v2 revises the central claim from "consistent vocabulary reduces the context a model needs" to a stronger, now-measured statement: names are model inputs treated as facts; a misleading name is a defect that injects false facts into every inference that touches it, and the lexicon is therefore a system component with its own tests, ratchets, and refactoring machinery. Context reduction is a corollary, not the claim. The revised version is grounded in the E-series measurements and the tier-true rename rehearsal, all run on this repository during sprints 4.8–4.12.

What changed from v1

1. v1 treated naming as one instrument among many. v2 elevates it: E2 showed the brand prefix was not neutral — models read it as an architectural claim ("foundation") and answered wrong with confidence. Misnaming is not missing information; it is misinformation. 2. v1 asked for experiments; v2 has results. The hypothesis ledger below records which of H1–H12 now have evidence from real use (no synthetic traffic, per operator rule). 3. v1 described enforcement abstractly. v2 names the mechanism that worked: ratified registries (single-source TSV seeds) feeding mechanical detectors (D9–D14) with shrink-only grandfather baselines, so the standard ratchets forward and never silently regresses. 4. v2 adds refactorization-as-data: a whole-workspace rename planned as database rows (term-level before/after per site), applied by a verify-then-replace exporter into a disposable worktree, proven by probes, replayable. The failure mode is "row rejected," never "silent wrong edit."

Diagrams

The vocabulary loop as operated:

flowchart LR
    M[measure in corpus DB] --> R[operator ratifies]
    R --> S[registry seed TSV]
    S --> D[detector D9-D14]
    D --> B[shrink-only baseline]
    B --> W[backfill work queue]
    W --> M

The rename-as-data pipeline (sprint 4.12 rehearsal):

flowchart LR
    CD[crate_desired_state] --> P[rename_edit plan rows]
    P --> X[exporter: verify then replace]
    X --> WT[disposable worktree]
    WT --> PR[probes: build, test, tools-doc-truth, check-deps]
    PR -->|red: new failure class| RL[tested rule added]
    RL --> P
    PR -->|green| V[verdict to operator; merge is stage 2]

Hypothesis ledger (v1 H-numbers, evidence to date)

| H | v1 claim (short) | Status | Evidence | |---|---|---|---| | H4 | route by judgment density, not size | supported | E1: haiku 26s vs sonnet 109s on Q1, quality-signaled routing seed (docs/research/doc-audit-study-2026-08-16.md) | | H7 | mechanical boundaries resist erosion | supported | layers.toml + check-deps; D10–D12 ratchets held through backfill | | H8 | controlled lexicon reduces inconsistent abstraction | supported | word stock 178→121 mechanically; D13 one-noun-one-referent found live corpus already clean | | H10 | harness-controlled beats unconstrained over time | partial | ratchet baselines only shrink across sprints 4.10–4.12; no controlled comparison arm yet | | H12 | semantic repair of deltas beats regeneration | supported | rename rehearsal: 35,879 planned site edits, stale rows fail loud and re-plan, nothing regenerated | | others | H1–H3, H5, H6, H9, H11 | untested | no measurement yet; candidates for the E-series |

New hypotheses (v2 additions)

degrades model answers below the no-name baseline. Measured once: E2, current names 41–61% name-only accuracy with systematic "foundation" misreads; tier-true names 100% on both model tiers (docs/research/e2-name-comprehension-2026-08-16.md).

probe gate produce a large refactor with zero silent wrong edits, at the cost of iterated rule discovery. Being measured now: rehearsal telemetry, series tier-true-rename-2026-08-16, nine failure classes each closed by a tested rule.

quality-signaled real runs converge to cheaper models without raising the refusal rate. Harness built (propose_routing); needs accumulation.

Telemetry

All quantitative claims trace to run_measurement in corpus_sourcecode (series: doc-audit-2026-08-16, e2-2026-08-16, tier-true-rename-2026-08-16) and to the study documents cited above. The held-out control for the central claim — same tasks, old names vs tier-true names, measuring task success and context consumed — is designed but not yet run; it is the single most important missing experiment and blocks any peer-review-grade version of this thesis.

All writing