GitHub repository | 日本語

Conveyor Courier Gallery — 202607_delegate_implement_bench

Browser-playable Web exports of the Conveyor Courier implementations generated by the benchmark. Click a model name to play the representative run (median composite score across adopted reps).

Reference implementation (the benchmark example): Play

Summary

Values are copied from benchmarks/202607_delegate_implement_bench/impressions.en.md. See the footnotes in that file for metric definitions.

Bold model names scored perfect functionality (hidden tests, 70/70) on the representative run.

ModelAuto-graded score (sum)Code quality (sonnet, sol)Parent + child cost (median)Wall clock (median)Note
claude-sonnet-5 (Claude)290.003.8 (4.3, 3.3) — only rep1 reaches 5, internal type abandonment in 2/3; large judge divergence, representative = operator adjudication$0.97 + $1.30 (total $2.27, measured)8.6 minQuality leader, incident-free. Two zero-warning runs
swe-1.7 (Devin)280.064.1 (4.0, 4.3) — only rep1 regresses to untyped Dictionary$1.41 + $0 (total $1.41, bundled)12.5 min+57.00 over swe-1.6. Its former 100-point run fell to 98.06 on the view click check
gpt-5.5 (Codex)280.004.0 (4.3, 3.8) — only rep2 reaches 5, untyped _items in 2/3$1.00 + $2.34 (total $3.48, converted)5.8 minPerfect functional score in every rep, incident-free. Zero warnings in 1/3
cursor-grok-4.5 (Cursor)276.003.3 (3.2, 3.5) — item swings between typed class and raw Dictionary$1.25 + $0.12 (total $1.38, converted)4.8 minFastest tier, fewest tokens. No view deductions — rose to 4th in the re-grade
devin-deepseek-v4-pro (Devin)270.063.4 (3.3, 3.5) — raw Dictionary internals, only rep2 reaches 4$1.53 + $2.05 (total $3.58, estimated)11.1 minPerfect functional score in every rep, but slowest tier and pricey
claude-opus-4-8 (Claude)270.003.5 (3.3, 3.7) — int usage leaks into the View layer in 2/3$0.93 + $1.42 (total $2.35, measured)8.4 minChild aces functional tests every time. All failures were parent/harness side
gpt-5.6-sol (Codex)268.064.2 (4.2, 4.3) — untyped _items only in rep0$1.50 + $1.77 (total $3.10, converted)7.1 minA stable 88–90 in every rep. Roughly ties terra with the same quality at twice the unit price
cursor-gemini-3.1-pro (Cursor)268.062.5 (2.8, 2.3) — never introduces constants, largest variance (representative = operator adjudication)$1.55 + $0.58 (total $2.13, converted)8.4 minWarning-count swings and a stall habit. Highest token use among Cursor models
cursor-kimi-k2.7-code (Cursor)268.064.5 (4.3, 4.7) — rep2 carries typed Dictionary[K,V] generics throughout$1.44 + $0.16 (total $1.60, converted)10.6 minSplitter quirk did not reproduce. Type discipline stays top-tier
gpt-5.4-mini (Codex)266.12
※ rep view defect: black board
3.3 (2.8, 3.5) — typing regresses across reps (representative = operator adjudication)$1.01 + $0.84 (total $2.04, converted)11.1 minFewest type warnings but slow, with a determinism drop and stalls
gpt-5.6-terra (Codex)266.12
※ rep view defect: tofu arrows
4.2 (4.2, 4.3) — constant usage and board representation vary across reps$1.21 + $0.91 (total $2.38, converted)6.3 minMatches sol at roughly half the cost and faster — the efficiency pick
devin-glm-5.2 (Devin)266.12
※ rep view defect: no belt arrows, mouse dead
3.8 (4.0, 3.7) — stable across runs, int internals throughout$0.90 + $1.37 (total $2.28, estimated)5.9 minFast and functionally accurate. Pre-completion stalls are the one flaw
composer-2.5 (Cursor)264.28
※ view defect: mouse dead in all reps
3.8 (3.8, 3.8) — untyped _items Dictionary in all 3 runs$1.23 + $0.04 (total $1.29, converted)6.5 minStill incident-free. Measured child cost is the lowest of all models
composer-2.5-fast (Cursor)264.28
※ view defect: mouse dead in all reps
3.7 (3.7, 3.7) — untyped nested arrays in 2/3$1.49 + $0.26 (total $1.76, converted)7.8 minRose to a tie with the base model, but slower than it after re-measurement
gpt-5.3-codex-spark (Codex)264.173.6 (3.8, 3.5) — largest per-rep swing (5 / 3 / 3.5)$1.47 + unknown (total ≥ $1.47)4.5 minFast but unstable on details. 3–5× the input tokens of gpt-5.5
gpt-5.6-luna (Codex)250.553.3 (3.3, —) — raw Dictionaries throughout, only rep2 uses a class$1.46 + $0.49 (total $1.95, converted)7.3 minFailed the splitter exit-toggle test in every rep (follow-up 1 traced it to reasoning effort; fixed at xhigh)
swe-1.6 (Devin)223.06
※ rep view defect: tofu arrows
3.8 (3.7, 4.0) — only rep1 drops typing entirely$0.95 + $0 (total $0.95, bundled)3.9 minCheapest (bundled pricing) but lowest quality on both axes
claude-haiku-4-5 (Claude)86.11 (n=1)
※ rep view defect: mouse dead
2.7 (3.0, 2.5) (n=1) — untyped even at the API contract$1.87 + $0.66 (total $2.53, measured)9.3 minCompleted 1/7. Frequent stalls, least reliable

Baseline condition (does not go through the delegation protocol; do not compare directly with the table above):

ConditionAuto-graded score (sum)Code quality (sonnet, sol)Parent + child cost (median)Wall clock (median)Note
fable-direct (no delegation)300.004.6 (4.7, 4.5) — stable across runs, top type discipline$2.62 + $0 (total $2.62, measured)5.7 minBaseline: the parent (Fable) implements directly

Follow-up 1 (2026-07-11, gpt-5.6-luna reasoning effort A/B: medium vs xhigh)

Conditional measurement (A/B); do not compare directly with the main summary table.

ConditionAuto-graded score (sum)Code quality (sonnet, sol)Parent + child cost (median)Wall clock (median)Note
gpt-5.6-luna@medium252.50
※ rep view defect: tofu arrows
3.7 (3.7, 3.7) — large per-rep variance persists (3.0/3.5/4.5)$1.59 + $0.71 (total $2.30, converted)8.0 minDefault effort. Same as round 2 under the old allocation.
gpt-5.6-luna@xhigh282.23
※ rep view defect: headless self-exit loses 3 view tests
3.8 (3.5, 4.2) — zero warnings in 2/3 but design effectively unchanged (representative = operator adjudication)$1.34 + $1.75 (total $3.09, converted)13.0 minModel-layer tests 3/3 perfect, one 100-point run; view deductions (tofu, self-exit) added

Follow-up 2 (2026-07-11, gdscript-quality skill × main bench: claude-sonnet-5 / claude-opus-4-8 with/without skill A/B)

Conditional measurement (A/B); do not compare directly with the main summary table.

ConditionAuto-graded score (sum)Code quality (sonnet, sol)Parent + child cost (median)Wall clock (median)Note
claude-sonnet-5@noskill282.004.2 (4.2, 4.2) — only rep2 is fully typed at 4.5$1.35 + $1.41 (total $2.64, measured)9.9 minBaseline. Down from main 290.00
claude-sonnet-5@gq298.06
※ rep view defect: mouse dead in rep2
5.0 (5.0, 5.0) — typed Dictionary all through, ◎$1.42 + $2.03 (total $3.46, measured)12.4 min100 / 100 / 98.06. Only rep2 lost view click test
claude-opus-4-8@noskill270.003.0 (2.7, 3.3) — rep1 is untyped-container mix at 2.0$1.36 + $1.35 (total $2.84, measured)9.0 minBaseline. Matches main 270.00
claude-opus-4-8@gq300.004.0 (4.0, 4.0) — identical across reps; only _grid: Array[int] is untyped$1.26 + $1.94 (total $3.25, measured)10.5 min100 points in every rep

Follow-up 3 (2026-07-14, gdscript-quality skill × Cursor harness: composer-2.5 with/without skill A/B)

Conditional measurement (A/B); do not compare directly with the main summary table.

ConditionAuto-graded score (sum)Code quality (sonnet, sol)Parent + child cost (median)Wall clock (median)Note
composer-2.5@noskill268.183.4 (3.3, 3.5) — internal untyped collections same as main$1.18 + $0.04 (total $1.23, converted)5.0 minBaseline. Roughly matches main 264.28
composer-2.5@gq292.23
※ view defect: mouse dead in 2/3 reps
4.2 (4.2, 4.3) — Dictionary[K, V] introduced in all reps; rep2 scored 5.0 by both judges$1.13 + $0.06 (total $1.18, converted)6.2 minZero warnings in all 3 reps (+24.05). View mouse dead remains, so 300 not reached

Other entries

GameModelRepresentative runScoreRep countStatusDetail
composer-2.5@gq-pilotcomposer-2.5@gq-pilot20260714T045450Z-composer-2.5_gq-pilot-rep084.171exportedshared engine