Research Question: RQ-tcr-variants-opus5

Findings — RQ-tcr-variants-opus5

Overview

All cells are correctness-complete enough for quality and efficiency comparison (verification_pct ≥ 0.90). Direction: correctness higher = better; complexity, Code Mass (APP), duration, tokens, and cost lower = better. Ties within the observed spread share trophies. Process and Git-history metrics are descriptive rather than intrinsically directional, so they receive no trophies.

Valid replicate counts are 4/5/5/4. One infrastructure-invalid EXACT Coding run and one method-invalid git-gamble run are excluded as documented in the RQ README.

F-4.8.1 — All three TCR variants deliver perfect external correctness

Classic TCR, native-Git TCRDD, and git-gamble TCRDD each reached verification_pct = 1.00 in every valid run. EXACT Coding remained near-perfect at 0.967 and passed its internal suite in all four valid runs.

The result does not support a correctness penalty from commit/revert discipline on this kata. The small EXACT Coding gap is one missed external scenario across four otherwise complete runs, not an internal-test failure.

F-4.8.2 — EXACT Coding buys decomposition and lower peak complexity

EXACT Coding produced substantially smaller functions and lower cognitive and mccabe_maxs than every TCR arm. The TCR variants were mutually much closer than any was to EXACT Coding on decomposition.

The clearest separation is average function size: EXACT Coding's 3.67 is less than half the 8.60–10.96 range of the TCR methods. This supports the interpretation that the predictive subagent architecture primarily purchases decomposition rather than correctness.

F-4.8.3 — Classic TCR is the efficiency winner

Classic TCR completed in 280 seconds with 3.9 M tokens and an estimated list-price cost of $3.38 per run. Native-Git TCRDD and git-gamble TCRDD were intermediate, while EXACT Coding required 3653 seconds, 63.2 M tokens, and $58.07 per run.

Classic TCR is roughly thirteen times faster and uses roughly one-thirteenth of EXACT Coding's tokens while retaining perfect external correctness. The cost figures are Requesty list-price estimates, not invoice amounts. The 18 valid RQ runs total $278.24 by that estimate.

F-4.8.4 — Native-Git and git-gamble TCRDD differ more in audit representation than product outcome

Both TCRDD implementations reached perfect correctness and generated a similar number of verified RED blocks, but their recorded protocol artifacts differ. Native Git retained explicit RED/GREEN phase commits and more explicit refactor steps; git-gamble's amend-based semantics retained no separate RED commits.

The zero retained RED commits for git-gamble is expected from its history-rewriting semantics and is not evidence that RED was skipped: marker-free analysis observed 25.8 verified RED blocks. Native Git makes phase history more directly visible, whereas git-gamble compresses it into the resulting commit history. The lower explicit refactor-step count for git-gamble nevertheless indicates a behavioral difference beyond history shape and does not support the hypothesis that tool enforcement automatically improves protocol adherence.