Research Question: RQ-old-vs-new-exact-line-opus55

Findings — RQ-old-vs-new-exact-line-opus55

All figures: claim-office-example-mapping × opus-5-5-no-thinking (native subscription route) × Claude Code 2.1.280, n=5 per cell. Token and price columns include subagent consumption; total_tokens has meant the whole run since 2026-09-30, so the two isolated arms are no longer read short.

Two lines, each in a shared-context and an isolated-Refactor variant:

Overview

Decomposition is the axis on which the two lines separated on Opus 5 (RQ-current-ptdd-vs-exact-opus-native F-4.9.1). On TypeScript longest_function and unit_size_max are the same measure, as are avg_loc_per_function and unit_size_avg; only one of each pair is tabled.

Every cell is at verification_pct ≥ 0.96 and within budget in 5/5 runs, so correctness-gating excludes nobody. Multiple trophies in a row mark values whose σ bands overlap, not a fabricated tie.

Production LoC and Test LoC carry no trophy — less code is not better per se, and the row that matters for parsimony is Code Mass (APP).

cost_usd is a list-price comparison value, not an invoice: these runs went over the native subscription route. Tariffs in research/model-pricing.md.

F-4.12.1 — Finer decomposition is a property of the retired line, and it does not come from its subagent

The retired line decomposes more finely than the maintained one in both of its variants, isolated and shared alike.

Both retired cells sit below both maintained cells on both measures, so the property belongs to the line and not to the Refactor architecture. H1 confirmed on Opus 5.5: the answer from Opus 5 carries over.

It does not translate into a lower cognitive_max or mccabe_max on its own. exact-single-context-v3-no-subagent-cc holds the second-finest decomposition and simultaneously the worst cognitive_max of the field at 4.00 ± 2.83 and the worst mccabe_max at 4.20 ± 1.79 — the two widest σ values measured here. Fine average decomposition and a controlled worst case are different properties, and in the retired line only the isolated variant delivers both (2.92 and 2.40).

The suite follows the retired line too, which the decomposition rows do not predict: exact-hybrid-v2-testlist-fix-cc has the strongest Mutation Score at 0.98 ± 0.02 and lets only 2.2 ± 2.2 mutants survive, against 7.8 and 6.4 in the maintained cells. Read the pair, not the ratio — the mutant populations are within 15 of each other (121.0 to 135.2), so the survivor counts compare directly.

F-4.12.2 — Isolating the Refactor step has opposite signs on cost and the same sign on quality

Holding the line constant and varying only the Refactor architecture:

Every one of those six movements except the retired line's wallclock increase (954 ± 173 → 1117 ± 138) is larger than the standard deviations involved. H3 confirmed: the subagent decision cannot be stated as a workflow-independent rule.

The mechanism is visible in the token counts rather than in the quality columns. A shared-context Refactor step re-reads the whole conversation on every cycle; an isolated one starts from a fresh, small context and returns a diff. In the retired line, whose shared-context variant burns 102.6 M tokens, that saves far more than the isolation costs. In the maintained line, whose shared context is already compact at 24.4 M, it does not.

On quality the two lines move the same way. Isolating the step lowers cognitive_max in both: 2.40 → 2.00 in the maintained line and 4.00 → 2.40 in the retired one, and mccabe_max likewise (3.40 → 3.00 and 4.20 → 3.20). The effect is far larger in the retired line, and it is the only place where it clears the σ bands. So the architecture is not quality-neutral — it is cost-ambiguous.

The practical consequence: in the retired line the isolated arm is the one to take — it halves the tokens, cuts a third of the price and tightens the complexity profile for a 17 % wallclock penalty. In the maintained line it buys cognitive_max 2.00 ± 0.00 — deterministic across all five runs — for 2.9× the price, and costs one of five perfect runs on top.

F-4.12.3 — The retired line costs 1.9× the maintained line

Comparing each line's cheapest viable variant:

The retired line's finer decomposition (F-4.12.1) costs roughly twice the price and twice the tokens for a 21 % wallclock premium. Both cells are above the correctness gate and neither has a budget problem.

The comparison is only fair in this pairing. Against the maintained line's isolated arm the retired line is cheaper ($18.35 against $28.34), and against the retired line's shared arm it is cheaper still ($18.35 against $27.86) — the architecture choice moves price more than the line does. The headline factor therefore belongs to the two cheapest-viable variants and should not be quoted as "the retired line costs 2× more" without that qualifier.

F-4.12.4 — The retired line without its subagent burns the most tokens and is the least stable cell measured

exact-single-context-v3-no-subagent-cc is the only cell in this RQ that is worse than its sibling on every axis that separates at all.

It writes the least production code of the field and still spends the most tokens on it — 102.6 M against 24.4 M for the maintained shared-context arm, which produces 222 lines. The shared Refactor step re-reading the full conversation on every cycle is the mechanism F-4.12.2 identifies, and this cell is where it costs most.

The instability is the part that matters for a recommendation. cognitive_max σ 2.83 on a mean of 4.00 means individual runs land anywhere from a clean 2 to a 7; exact-hybrid-v2-testlist-fix-cc, the same line with the step isolated, holds 2.40 ± 0.89. A workflow whose worst-case complexity is a coin flip is not usable as a default even where its average decomposition looks good.

Correctness is unaffected — 5/5 perfect, the joint best in the field with exact-ptdd-v1-cc — so nothing here is a capability statement. It is a cost and variance statement.