Figure 4 · E004

Effect size against bare Claude Code

Cohen's d expresses how far an arm's accuracy distribution sits from bare Claude Code's in units of their pooled standard deviation, so it is comparable across batteries and models.

protocols
battery
model
columns

Pooled models shows one column per arm, its models averaged with equal weight — the readable overview. Per model shows one column per model for every arm, for like-for-like reads at a fixed model. This choice applies to every figure on the site and is remembered.

Effect size against bare Claude Codehigher is betterZeroShot + MultiTurn · both batteries
0.00.51.01.51.359 modelsTrueArchitectTrueArchitect1st of 80.006 modelsCC Barebare harness7th of 81.288 modelsCursor Barebare harness2nd of 80.176 modelsGraphifyindexing tool5th of 80.146 modelsGitNexusindexing tool6th of 8-0.076 modelsCodeGraphindexing tool8th of 80.246 modelsCodebaseMemoryindexing tool3rd of 80.206 modelsSerenaindexing tool4th of 8
Data behind this figure ZeroShot + MultiTurn · both batteries
armrolecolumnCohen's d vs bare Claude Code, accuracyrunscellsminmaxsdrankruns in the repository
TrueArchitectTrueArchitect9 models1.352002165.00100.008.191st of 8Haiku 4.5 · Sonnet 5 · Opus 4.6 · Opus 4.8 · Opus 5 · Fable 5 · GPT 5.6 Luna · GPT 5.6 Terra · GPT 5.6 Sol
CC Barebare harness6 models0.002012240.00100.0011.537th of 8Haiku 4.5 · Sonnet 5 · Opus 4.6 · Opus 4.8 · Opus 5 · Fable 5
Cursor Barebare harness8 models1.281201255.00100.0011.422nd of 8Haiku 4.5 · Sonnet 5 · Opus 5 · GPT 5.6 Luna · GPT 5.6 Terra · GPT 5.6 Sol · Grok 4.6 · Composer 2.5
Graphifyindexing tool on Claude Code6 models0.172012240.00100.0011.635th of 8Haiku 4.5 · Sonnet 5 · Opus 4.6 · Opus 4.8 · Opus 5 · Fable 5
GitNexusindexing tool on Claude Code6 models0.142012255.00100.009.216th of 8Haiku 4.5 · Sonnet 5 · Opus 4.6 · Opus 4.8 · Opus 5 · Fable 5
CodeGraphindexing tool on Claude Code6 models-0.072032240.00100.0012.948th of 8Haiku 4.5 · Sonnet 5 · Opus 4.6 · Opus 4.8 · Opus 5 · Fable 5
CodebaseMemoryindexing tool on Claude Code6 models0.242012260.00100.009.673rd of 8Haiku 4.5 · Sonnet 5 · Opus 4.6 · Opus 4.8 · Opus 5 · Fable 5
Serenaindexing tool on Claude Code6 models0.202032250.00100.0010.924th of 8Haiku 4.5 · Sonnet 5 · Opus 4.6 · Opus 4.8 · Opus 5 · Fable 5

Figure 4. Cohen's d expresses how far an arm's accuracy distribution sits from bare Claude Code's in units of their pooled standard deviation, so it is comparable across batteries and models. Each column is the mean d over cells that contain both the arm and bare Claude Code at the same model, battery and protocol; bare Claude Code is the zero line by definition, and arms with no shared Anthropic cell have no value. By convention d of 0.2 is small, 0.5 medium and 0.8 large; TrueArchitect is the only arm above the large threshold.

How this figure is computed

Definition. Within a cell shared by the arm (n₁ runs, accuracies with mean m₁ and standard deviation s₁) and bare Claude Code (n₂, m₂, s₂), d = (m₁ − m₂) / s_p with s_p = √(((n₁−1)s₁² + (n₂−1)s₂²) / (n₁+n₂−2)). Standard deviations use the n−1 denominator.

Aggregation. The column is the equal-weight mean of d over the cells the arm shares with bare Claude Code. A cell with a single run on either side contributes a degenerate deviation and is reported as such in the table. Codex and Cursor at non-Anthropic models have no bare Claude Code counterpart and are therefore absent, not zero.

Reading it. Higher is better; negative values would mean the arm is worse than bare Claude Code. The figure answers "is the difference large relative to run-to-run noise", the question a mean alone cannot answer.

Provenance

Computed at build time from epochs/E004/summary/runs.json of the proof package (export tool tabench export public 0.1.25, scorer era tabench-1.1). Each row of the table above links to the run directories in the repository, where run.json, the answers, the verdicts and the transcript of every run can be read.