Bridge Test (TX) — Architecture as AI-Derivability

Runnable protocol. Grounding + literature: ../../docs/white-paper/ieee-access/bridge-test-grounding.md. Theory: Compendium §4.1 (the bridge). This supersedes the prompting-altitude design in ../bridge-strong/ + ../bridge-weak/ FOR THE BRIDGE CLAIM; those two prompts are retained as the instrument for the sibling specificity-dial study (§ Specificity Dial below).

Claim under test

Human-maintainability structure == AI-derivability. The architectural disciplines (SOLID interfaces + SRP, hexagonal/layered known-locations, tests-as-contracts, intention-revealing names) were built so a HUMAN derives intent without reverse-engineering; the SAME structure lets a stateless AI reader derive intent without reverse-engineering. Measured on the READ side.

The gap we fill (from grounding doc)

Naming (Le/Li 2025), comments (Vitale 2026), and smells (Xue 2025) → AI comprehension are shown. No study manipulates the architectural disciplines on the same system measuring AI comprehension + modification + token cost. “Architecture as AI-derivability” is OPEN.

Twins (the only free variable is structure)

  • M (mud) = a real default-prompt Conduit generation, NO bridge signal. Source: an existing runner/runs/naive/{i}/project rep. Selection rule: pick the rep whose measured smell profile best matches the documented default profile (Long Method / God Class / poor names / no interfaces; see grounding §”natural default”). Cite that M is representative, not fabricated.
  • D (disciplined) = a behavior-preserving disciplined refactor of that same M: extract SOLID interfaces + SRP units, hexagonal ports/adapters (known locations), intention-revealing names, tests-as-contracts. NOTHING about observable behavior changes.
  • Oracle (behavioral equivalence proof) = the shared Hurl suite ../evidence/hurl/*.hurl (13 files: auth, articles, comments, profiles, favorites, feed, pagination, tags, + 5 error suites). D and M are admissible as twins ONLY when the identical suite passes GREEN on both. If default M fails an oracle case behaviorally, fix behavior MINIMALLY (no restructuring) and log it; M’s structure stays untouched so it remains a genuine default-smell artifact.

Confound control: architecture ABOVE naming

Naming is a known-strong channel (Le/Li). To attribute the effect to ARCHITECTURE and not just names, run two D contrasts:

  • D-full = architecture + naming disciplined.
  • D-arch = architecture disciplined, names held at M-quality (obfuscated/generic), a ClassEval-Obf-style naming control. Effect of D-arch over M isolates the architecture channel.

Dependent variables (fresh, stateless assistant per twin, per task)

Each assistant gets ONLY the twin’s source tree, no session history, no bridge narration.

  1. Comprehension — a fixed bank of ~18 questions across three types: intent (“what does X guarantee?”), location (“where is invariant Y enforced?”), invariant (“what breaks if Z changes?”). Score token-level F1 against fixed ground truth. Metric: mean F1.
  2. Modification — a SWE-bench-style task set (add/change a feature, e.g. add article tag-rename cascade, add favorite-count denormalization). Metric: does the shared Hurl oracle stay GREEN after the patch? (binary pass) + tokens spent + read-breadth.
  3. Cost / reverse-engineering — tokens consumed and read-breadth (files opened, targeted vs whole-tree sweep) per Ma-2026 accounting. Reverse-engineering-cost proxy.

Prediction

D → accurate, cheap, targeted reads; M → error-prone, expensive, whole-codebase sweeps. Closes the lifecycle-economics arc: AX (GS costs MORE to generate) + KX + TX (disciplined structure is CHEAPER to read/comprehend/extend) = the real answer to the token-cost objection (pay at generation, recover across maintenance). NO token-reduction % asserted.

Protocol (binary first)

  1. Select M — measure smell profile across naive/{0..4}, pick best-matching rep. [measure.cjs]
  2. Oracle-baseline M — run Hurl suite against M; record green/red; minimal behavior-only fixes.
  3. Build D-full and D-arch — behavior-preserving refactor; Hurl suite must stay green on both.
  4. Author DV instruments — the 18-Q comprehension bank + ground truth; the SWE-bench-style modification tasks + their acceptance (oracle-green); the token/read-breadth accounting harness.
  5. Run — k stateless assistants per {M, D-arch, D-full} × {comprehension, modification}.
  6. Stats — paired per twin; Mann-Whitney U + Cliff’s delta + Holm-Bonferroni (reuse stats.py). Binary D-vs-M first; discipline-ladder (strip one discipline at a time) only if binary separates.

Pre-registered caveats

Varies by language/prompt/project (Zhu, Paul) — run on >=1 system, ideally add a 2nd language/domain for external validity. Architecture effect must be shown ABOVE the naming effect (the D-arch contrast). M is one representative default, not a claim about all defaults.

Specificity Dial (sibling study, reuses bridge-strong/weak)

Distinct claim (Compendium §4.5.1): specificity (not specification) is a DIAL; low specificity → high AI freedom → light verification; high specificity → heavy human control. ../bridge-strong/ (conceptual = low specificity) and ../bridge-weak/ (mechanical = high specificity) are the two dial ends. Design (later): vary specificity × stakes, measure time-to-acceptable-quality + verification cost. Not part of the binary bridge test; documented here so the prompts are not lost.