Skills x Effort Lattice: Results Explorer

Every number below is read from the committed result files at the moment this page was built; the template only supplies labels. Every box is a small sample, so read each rate and change as a rough signal, not a precise measurement.

How this page was built
built 2026-07-13 22:33:34 UTC from results/matrix/matrix.json, 17 run directories matching results/lattice-*, and results/concordance/codex-concordance.json
rebuild with: python3 eval/make_explorer.py --matrix results/matrix/matrix.json --cells 'results/lattice-*' --out results/matrix/explorer.html
generation notes
  • lattice-opus-medium-batch1-adjudicate: scored data on disk but not a matrix cell; not displayed

1. Skill effect across models and effort

Each box is one model run at one effort level, scored two ways: without the skill installed and with it. Each task carries a short checklist of things a correct answer must do, and a score is the share of those checklist items hit. Change is the with-skill score minus the without-skill score, in percentage points. The rules and pass criteria were locked and publicly hash-stamped before the first run. Every box is a small sample, so read each number as a rough signal, not a precise measurement.

15 conditions on one shared task set
the 17 tasks that were valid in every condition, the only basis that compares cleanly across boxes
show:

Codex 5.6, one run, 50 sampled comparisons, exploratory: agreed 97.2%. A spot-check by a second vendor, not a re-run of the experiment. Turn it on to mark the boxes whose comparisons Codex re-scored.

model \ effortlowmediumhighxhighmaxno effort(reference)
claude-fable-5×3 runswithout 94.1%with 99.0%change +4.917 tasks, 3 runs eachCodex 29/30 agreewithout 92.6%with 100.0%change +7.417 tasks, 1 run eachCodex 24/26 agree×3 runswithout 97.6%with 100.0%change +2.417 tasks, 3 runs eachCodex 56/58 agreewithout 97.1%with 100.0%change +2.917 tasks, 1 run eachCodex 14/14 agreemax run not finishedit got too expensive to finish; one set of runs is still going, so this box may fill in later-
claude-sonnet-5×3 runswithout 84.8%with 96.1%change +11.317 tasks, 3 runs eachCodex 32/32 agree×3 runswithout 90.7%with 99.0%change +8.417 tasks, 3 runs each×3 runswithout 87.8%with 99.0%change +11.317 tasks, 3 runs each×3 runswithout 87.3%with 98.5%change +11.317 tasks, 3 runs each×3 runswithout 89.7%with 99.5%change +9.817 tasks, 3 runs eachCodex 40/40 agree-
claude-opus-4-8×3 runswithout 87.3%with 95.6%change +8.317 tasks, 3 runs eachCodex 51/52 agree×3 runswithout 85.8%with 97.6%change +11.817 tasks, 3 runs each×3 runswithout 88.2%with 98.0%change +9.817 tasks, 3 runs each×3 runswithout 91.2%with 99.0%change +7.817 tasks, 3 runs each×3 runswithout 91.2%with 100.0%change +8.817 tasks, 3 runs eachCodex 36/38 agree-
claude-haiku-4-5-20251001no effort setting, shown for reference-----without 64.7%with 91.2%change +26.517 tasks, 1 run eachCodex 16/16 agree
Methodology: without = no skill installed; with = skill installed; change = with minus without, in percentage points, weighted by the checklist. Boxes labelled "mean of 3 runs" average three repeat runs; the rest are a single run. Numbers read this generation from the matrix artifact. Codex counts read from the concordance file.

2. Effort curves per model

The same checklist scores drawn against effort. Each model has its own colour. The solid line is with the skill installed, the dashed line is without. Low and max were run three times; the middle levels once.

claude-fable-5
checklist score against effort, 17 shared tasks
without the skillwith the skill
80859095100lowmediumhighxhighmaxeffort levelclaude-fable-5 at low: without the skill 94.1% (17 tasks, 3 runs each)claude-fable-5 at medium: without the skill 92.6% (17 tasks, 1 run each)claude-fable-5 at high: without the skill 97.6% (17 tasks, 3 runs each)claude-fable-5 at xhigh: without the skill 97.1% (17 tasks, 1 run each)claude-fable-5 at low: with the skill 99.0% (17 tasks, 3 runs each)claude-fable-5 at medium: with the skill 100.0% (17 tasks, 1 run each)claude-fable-5 at high: with the skill 100.0% (17 tasks, 3 runs each)claude-fable-5 at xhigh: with the skill 100.0% (17 tasks, 1 run each)not finished
Methodology: each point is the score with or without the skill on the 17 shared tasks. Low and max are the mean of 3 repeat runs (ring marker, side ticks at each run); the other levels are a single run (small dot). Small samples, read as a rough signal.
claude-sonnet-5
checklist score against effort, 17 shared tasks
without the skillwith the skill
80859095100lowmediumhighxhighmaxeffort levelclaude-sonnet-5 at low: without the skill 84.8% (17 tasks, 3 runs each)claude-sonnet-5 at medium: without the skill 90.7% (17 tasks, 3 runs each)claude-sonnet-5 at high: without the skill 87.8% (17 tasks, 3 runs each)claude-sonnet-5 at xhigh: without the skill 87.3% (17 tasks, 3 runs each)claude-sonnet-5 at max: without the skill 89.7% (17 tasks, 3 runs each)claude-sonnet-5 at low: with the skill 96.1% (17 tasks, 3 runs each)claude-sonnet-5 at medium: with the skill 99.0% (17 tasks, 3 runs each)claude-sonnet-5 at high: with the skill 99.0% (17 tasks, 3 runs each)claude-sonnet-5 at xhigh: with the skill 98.5% (17 tasks, 3 runs each)claude-sonnet-5 at max: with the skill 99.5% (17 tasks, 3 runs each)
Methodology: each point is the score with or without the skill on the 17 shared tasks. Low and max are the mean of 3 repeat runs (ring marker, side ticks at each run); the other levels are a single run (small dot). Small samples, read as a rough signal.
claude-opus-4-8
checklist score against effort, 17 shared tasks
without the skillwith the skill
80859095100lowmediumhighxhighmaxeffort levelclaude-opus-4-8 at low: without the skill 87.3% (17 tasks, 3 runs each)claude-opus-4-8 at medium: without the skill 85.8% (17 tasks, 3 runs each)claude-opus-4-8 at high: without the skill 88.2% (17 tasks, 3 runs each)claude-opus-4-8 at xhigh: without the skill 91.2% (17 tasks, 3 runs each)claude-opus-4-8 at max: without the skill 91.2% (17 tasks, 3 runs each)claude-opus-4-8 at low: with the skill 95.6% (17 tasks, 3 runs each)claude-opus-4-8 at medium: with the skill 97.6% (17 tasks, 3 runs each)claude-opus-4-8 at high: with the skill 98.0% (17 tasks, 3 runs each)claude-opus-4-8 at xhigh: with the skill 99.0% (17 tasks, 3 runs each)claude-opus-4-8 at max: with the skill 100.0% (17 tasks, 3 runs each)
Methodology: each point is the score with or without the skill on the 17 shared tasks. Low and max are the mean of 3 repeat runs (ring marker, side ticks at each run); the other levels are a single run (small dot). Small samples, read as a rough signal.

claude-haiku-4-5-20251001 has no effort setting, so it sits outside the effort comparison: without the skill 64.7%, with the skill 91.2%, change +26.5 (17 tasks, 1 run each).

3. Per-skill detail (17 skills)

per-skill PASS/FAIL verdicts are deliberately suppressed in lattice outputs and subordinated to the per-cell run reports; skills appear here as rates only Each cell below shows checklist hits for that skill's one task within each condition. Every per-skill cell is a single task, so these are the roughest signals on the page.

adversarial-review/multi-model-adversarial-review
conditionwithout skillwith skillchangesample
claude-fable-5@low2/366.7%3/3100.0%+33.3points1 task, 3 runs each
claude-fable-5@medium3/3100.0%3/3100.0%+0.0points1 task, 1 run each
claude-fable-5@high3/3100.0%3/3100.0%+0.0points1 task, 3 runs each
claude-fable-5@xhigh3/3100.0%3/3100.0%+0.0points1 task, 1 run each
claude-haiku-4-5-20251001@none1/333.3%2/366.7%+33.3points1 task, 1 run each
claude-opus-4-8@low2/366.7%2/366.7%+0.0points1 task, 3 runs each
claude-opus-4-8@medium2/366.7%3/3100.0%+33.3points1 task, 3 runs each
claude-opus-4-8@high3/3100.0%3/3100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@xhigh3/3100.0%3/3100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@max3/3100.0%3/3100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@low2/366.7%2/366.7%+0.0points1 task, 3 runs each
claude-sonnet-5@medium3/3100.0%3/3100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@high3/3100.0%3/3100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@xhigh3/3100.0%3/3100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@max3/3100.0%3/3100.0%+0.0points1 task, 3 runs each
architecture-and-contracts/architecture-contracts-as-law
conditionwithout skillwith skillchangesample
claude-fable-5@low4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-fable-5@medium3/475.0%4/4100.0%+25.0points1 task, 1 run each
claude-fable-5@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-fable-5@xhigh3/475.0%4/4100.0%+25.0points1 task, 1 run each
claude-haiku-4-5-20251001@none2/450.0%4/4100.0%+50.0points1 task, 1 run each
claude-opus-4-8@low3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-opus-4-8@medium3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-opus-4-8@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@xhigh4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@max4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@low3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@medium3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@high3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@xhigh2/450.0%4/4100.0%+50.0points1 task, 3 runs each
claude-sonnet-5@max2/450.0%4/4100.0%+50.0points1 task, 3 runs each
auth-and-tenancy/multi-tenant-auth-reference
conditionwithout skillwith skillchangesample
claude-fable-5@low5/5100.0%5/5100.0%+0.0points1 task, 3 runs each
claude-fable-5@medium5/5100.0%5/5100.0%+0.0points1 task, 1 run each
claude-fable-5@high5/5100.0%5/5100.0%+0.0points1 task, 3 runs each
claude-fable-5@xhigh5/5100.0%5/5100.0%+0.0points1 task, 1 run each
claude-haiku-4-5-20251001@none4/580.0%5/5100.0%+20.0points1 task, 1 run each
claude-opus-4-8@low4/580.0%4/580.0%+0.0points1 task, 3 runs each
claude-opus-4-8@medium4/580.0%5/5100.0%+20.0points1 task, 3 runs each
claude-opus-4-8@high4/580.0%5/5100.0%+20.0points1 task, 3 runs each
claude-opus-4-8@xhigh5/5100.0%5/5100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@max5/5100.0%5/5100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@low4/580.0%5/5100.0%+20.0points1 task, 3 runs each
claude-sonnet-5@medium5/5100.0%5/5100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@high4/580.0%5/5100.0%+20.0points1 task, 3 runs each
claude-sonnet-5@xhigh4/580.0%5/5100.0%+20.0points1 task, 3 runs each
claude-sonnet-5@max3/560.0%5/5100.0%+40.0points1 task, 3 runs each
campaign-execution/multi-agent-batch-campaigns
conditionwithout skillwith skillchangesample
claude-fable-5@low3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-fable-5@medium3/475.0%4/4100.0%+25.0points1 task, 1 run each
claude-fable-5@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-fable-5@xhigh4/4100.0%4/4100.0%+0.0points1 task, 1 run each
claude-haiku-4-5-20251001@none0/40.0%2/450.0%+50.0points1 task, 1 run each
claude-opus-4-8@low3/475.0%3/475.0%+0.0points1 task, 3 runs each
claude-opus-4-8@medium3/475.0%3/475.0%+0.0points1 task, 3 runs each
claude-opus-4-8@high2/450.0%4/4100.0%+50.0points1 task, 3 runs each
claude-opus-4-8@xhigh3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-opus-4-8@max3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@low3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@medium3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@xhigh3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@max3/475.0%4/4100.0%+25.0points1 task, 3 runs each
change-control/git-change-control-for-agents
conditionwithout skillwith skillchangesample
claude-fable-5@low4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-fable-5@medium4/4100.0%4/4100.0%+0.0points1 task, 1 run each
claude-fable-5@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-fable-5@xhigh4/4100.0%4/4100.0%+0.0points1 task, 1 run each
claude-haiku-4-5-20251001@none2/450.0%3/475.0%+25.0points1 task, 1 run each
claude-opus-4-8@low4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@medium4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@high4/4100.0%3/475.0%-25.0points1 task, 3 runs each
claude-opus-4-8@xhigh4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@max4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@low3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@medium4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@xhigh4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@max4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
cost-and-safety-guardrails/ai-cost-tracking-and-guardrails
conditionwithout skillwith skillchangesample
claude-fable-5@low4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-fable-5@medium4/4100.0%4/4100.0%+0.0points1 task, 1 run each
claude-fable-5@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-fable-5@xhigh4/4100.0%4/4100.0%+0.0points1 task, 1 run each
claude-haiku-4-5-20251001@none3/475.0%4/4100.0%+25.0points1 task, 1 run each
claude-opus-4-8@low3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-opus-4-8@medium3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-opus-4-8@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@xhigh4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@max4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@low3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@medium4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@high3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@xhigh4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@max4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
cost-and-safety-guardrails/budget-aware-model-allocation
conditionwithout skillwith skillchangesample
claude-fable-5@low4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-fable-5@medium3/475.0%4/4100.0%+25.0points1 task, 1 run each
claude-fable-5@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-fable-5@xhigh4/4100.0%4/4100.0%+0.0points1 task, 1 run each
claude-haiku-4-5-20251001@none3/475.0%4/4100.0%+25.0points1 task, 1 run each
claude-opus-4-8@low3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-opus-4-8@medium3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-opus-4-8@high3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-opus-4-8@xhigh3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-opus-4-8@max3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@low3/475.0%3/475.0%+0.0points1 task, 3 runs each
claude-sonnet-5@medium4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@xhigh3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@max4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
cost-and-safety-guardrails/config-and-secrets-hygiene
conditionwithout skillwith skillchangesample
claude-fable-5@low4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-fable-5@medium4/4100.0%4/4100.0%+0.0points1 task, 1 run each
claude-fable-5@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-fable-5@xhigh4/4100.0%4/4100.0%+0.0points1 task, 1 run each
claude-haiku-4-5-20251001@none3/475.0%4/4100.0%+25.0points1 task, 1 run each
claude-opus-4-8@low3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-opus-4-8@medium3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-opus-4-8@high3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-opus-4-8@xhigh4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@max4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@low3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@medium3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@high3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@xhigh3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@max3/475.0%4/4100.0%+25.0points1 task, 3 runs each
debugging-playbooks/failure-archaeology
conditionwithout skillwith skillchangesample
claude-fable-5@low4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-fable-5@medium4/4100.0%4/4100.0%+0.0points1 task, 1 run each
claude-fable-5@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-fable-5@xhigh4/4100.0%4/4100.0%+0.0points1 task, 1 run each
claude-haiku-4-5-20251001@none3/475.0%4/4100.0%+25.0points1 task, 1 run each
claude-opus-4-8@low4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@medium4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@xhigh4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@max4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@low4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@medium4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@xhigh4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@max4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
debugging-playbooks/systematic-debugging-playbook
conditionwithout skillwith skillchangesample
claude-fable-5@low4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-fable-5@medium4/4100.0%4/4100.0%+0.0points1 task, 1 run each
claude-fable-5@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-fable-5@xhigh4/4100.0%4/4100.0%+0.0points1 task, 1 run each
claude-haiku-4-5-20251001@none3/475.0%4/4100.0%+25.0points1 task, 1 run each
claude-opus-4-8@low4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@medium4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@xhigh4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@max4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@low4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@medium4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@xhigh4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@max4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
deploy-and-infra/environment-and-build-hazards
conditionwithout skillwith skillchangesample
claude-fable-5@low3/3100.0%3/3100.0%+0.0points1 task, 3 runs each
claude-fable-5@medium3/3100.0%3/3100.0%+0.0points1 task, 1 run each
claude-fable-5@high3/3100.0%3/3100.0%+0.0points1 task, 3 runs each
claude-fable-5@xhigh3/3100.0%3/3100.0%+0.0points1 task, 1 run each
claude-haiku-4-5-20251001@none3/3100.0%3/3100.0%+0.0points1 task, 1 run each
claude-opus-4-8@low3/3100.0%3/3100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@medium3/3100.0%3/3100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@high3/3100.0%3/3100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@xhigh3/3100.0%3/3100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@max3/3100.0%3/3100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@low3/3100.0%3/3100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@medium3/3100.0%3/3100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@high3/3100.0%3/3100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@xhigh3/3100.0%3/3100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@max3/3100.0%3/3100.0%+0.0points1 task, 3 runs each
deploy-and-infra/staging-to-prod-cutover-campaign
conditionwithout skillwith skillchangesample
claude-fable-5@low4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-fable-5@medium4/4100.0%4/4100.0%+0.0points1 task, 1 run each
claude-fable-5@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-fable-5@xhigh4/4100.0%4/4100.0%+0.0points1 task, 1 run each
claude-haiku-4-5-20251001@none4/4100.0%4/4100.0%+0.0points1 task, 1 run each
claude-opus-4-8@low4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@medium3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-opus-4-8@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@xhigh4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@max3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@low3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@medium4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@xhigh4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@max3/475.0%4/4100.0%+25.0points1 task, 3 runs each
docs-and-compliance/consent-and-regulated-data-reference
conditionwithout skillwith skillchangesample
claude-fable-5@low4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-fable-5@medium4/4100.0%4/4100.0%+0.0points1 task, 1 run each
claude-fable-5@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-fable-5@xhigh4/4100.0%4/4100.0%+0.0points1 task, 1 run each
claude-haiku-4-5-20251001@none2/450.0%4/4100.0%+50.0points1 task, 1 run each
claude-opus-4-8@low4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@medium4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@xhigh4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@max3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@low3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@medium4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@high3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@xhigh4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@max4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
docs-and-compliance/docs-of-record-and-arbitration
conditionwithout skillwith skillchangesample
claude-fable-5@low4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-fable-5@medium4/4100.0%4/4100.0%+0.0points1 task, 1 run each
claude-fable-5@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-fable-5@xhigh4/4100.0%4/4100.0%+0.0points1 task, 1 run each
claude-haiku-4-5-20251001@none4/4100.0%3/475.0%-25.0points1 task, 1 run each
claude-opus-4-8@low4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@medium4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@xhigh4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@max4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@low4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@medium4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@xhigh4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@max4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
evals-and-scoring/llm-eval-harness-and-scoring-pipeline
conditionwithout skillwith skillchangesample
claude-fable-5@low3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-fable-5@medium3/475.0%4/4100.0%+25.0points1 task, 1 run each
claude-fable-5@high3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-fable-5@xhigh3/475.0%4/4100.0%+25.0points1 task, 1 run each
claude-haiku-4-5-20251001@none3/475.0%4/4100.0%+25.0points1 task, 1 run each
claude-opus-4-8@low3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-opus-4-8@medium3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-opus-4-8@high3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-opus-4-8@xhigh3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-opus-4-8@max3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@low3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@medium4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@high3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@xhigh3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@max4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
tiered-review/tiered-consultancy-review
conditionwithout skillwith skillchangesample
claude-fable-5@low5/5100.0%5/5100.0%+0.0points1 task, 3 runs each
claude-fable-5@medium5/5100.0%5/5100.0%+0.0points1 task, 1 run each
claude-fable-5@high5/5100.0%5/5100.0%+0.0points1 task, 3 runs each
claude-fable-5@xhigh5/5100.0%5/5100.0%+0.0points1 task, 1 run each
claude-haiku-4-5-20251001@none3/560.0%4/580.0%+20.0points1 task, 1 run each
claude-opus-4-8@low5/5100.0%5/5100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@medium5/5100.0%5/5100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@high5/5100.0%5/5100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@xhigh5/5100.0%5/5100.0%+0.0points1 task, 3 runs each
claude-opus-4-8@max5/5100.0%5/5100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@low5/5100.0%5/5100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@medium5/5100.0%5/5100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@high4/580.0%5/5100.0%+20.0points1 task, 3 runs each
claude-sonnet-5@xhigh5/5100.0%5/5100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@max5/5100.0%5/5100.0%+0.0points1 task, 3 runs each
validation-gates/pre-merge-validation-gate
conditionwithout skillwith skillchangesample
claude-fable-5@low4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-fable-5@medium3/475.0%4/4100.0%+25.0points1 task, 1 run each
claude-fable-5@high4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-fable-5@xhigh4/4100.0%4/4100.0%+0.0points1 task, 1 run each
claude-haiku-4-5-20251001@none1/425.0%4/4100.0%+75.0points1 task, 1 run each
claude-opus-4-8@low3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-opus-4-8@medium3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-opus-4-8@high3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-opus-4-8@xhigh3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-opus-4-8@max4/4100.0%4/4100.0%+0.0points1 task, 3 runs each
claude-sonnet-5@low3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@medium3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@high3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@xhigh3/475.0%4/4100.0%+25.0points1 task, 3 runs each
claude-sonnet-5@max3/475.0%4/4100.0%+25.0points1 task, 3 runs each
Methodology: without and with are checklist hits for that skill's task in each condition; change is with minus without, in percentage points. One task per cell.

4. How the scoring was judged

Every blinded comparison was scored by two independent judges, one Sonnet-class and one Opus-class. Where they split on a checklist mark, a pinned third judge, Claude Fable 5, decided it, and the final mark is the majority of the three. Disputed marks stay in every total.

4.6%marks the two judges split on
(122 of 2652)
2.5%marks sent to the third judge
(135 of 5304)
0left unresolved
conditionjudges split sent to third judgeunresolved
claude-fable-5@low9 / 2044.4%10 / 4082.5%0
claude-fable-5@medium1 / 681.5%1 / 1360.7%0
claude-fable-5@high9 / 2044.4%10 / 4082.5%0
claude-fable-5@xhigh3 / 684.4%3 / 1362.2%0
claude-haiku-4-5-20251001@none9 / 6813.2%10 / 1367.4%0
claude-opus-4-8@low14 / 2046.9%15 / 4083.7%0
claude-opus-4-8@medium9 / 2044.4%11 / 4082.7%0
claude-opus-4-8@high9 / 2044.4%11 / 4082.7%0
claude-opus-4-8@xhigh6 / 2042.9%7 / 4081.7%0
claude-opus-4-8@max8 / 2043.9%9 / 4082.2%0
claude-sonnet-5@low8 / 2043.9%10 / 4082.5%0
claude-sonnet-5@medium9 / 2044.4%9 / 4082.2%0
claude-sonnet-5@high9 / 2044.4%9 / 4082.2%0
claude-sonnet-5@xhigh5 / 2042.5%5 / 4081.2%0
claude-sonnet-5@max14 / 2046.9%15 / 4083.7%0
Methodology: split and sent-to-third counts are per checklist mark; disputed marks are kept in every denominator. No mark was left unresolved.

5. Second-vendor spot-check (Codex 5.6)

Codex 5.6, one run, 50 sampled comparisons, exploratory: agreed 97.2%

A different vendor's model, Codex 5.6 (gpt-5.6-terra), re-scored 50 of the comparisons using the same frozen instructions the Claude judges saw. This is a confidence spot-check, not a re-run of the experiment, and it changes none of the numbers above. Use the toggle in section 1 to see which boxes these comparisons came from.

Codex agreed with the final marks 97.2% of the time
383 of 394 individual checklist marks matched
97.2%marks Codex agreed with
(383 of 394)
50comparisons re-scored
83.3%of the marks the Claude judges split on,
Codex sided with the final call (5 of 6)
agreement by model columnFableSonnetOpusHaiku
Codex vs final marks96.1%123/128 marks97.5%119/122 marks97.7%125/128 marks100.0%16/16 marks
Methodology: Codex ran once at high reasoning effort over 50 comparisons picked by a published hash rule, scoring the same blinded content the judges scored. Exploratory only: it never touches a rate, difference, or verdict.