STEVE_BENCH V2 · PRE-REGISTERED · K=1 PRELIMINARY

Scaffolding beats model choice. Measured.

The same three models, run through the same 518 tasks, three times each — once bare, once with tools, once with tools plus a governance layer. Governance roughly doubles operational accuracy for every model tested, and the cheapest model wins on cost per correct answer. Every number on this page recomputes from the raw rows linked at the bottom.

TL;DR

Five things to take away

  • Governance roughly doubles accuracy. Vanilla scores in the mid-20s across all three models; governed scores land 55–65%.
  • Scaffolding beats model choice. Governed Sonnet (63.8%) statistically ties governed Opus (64.8%) at roughly one-third the cost per correct answer.
  • Governance isn't free — we show the bill. +47% cost per task and +50% latency vs. the tooled arm, published on this page, never buried.
  • This is k=1, labeled preliminary everywhere accuracy appears. A k=3 re-run is scheduled; treat single-sample cells as directional, not final.
  • Nothing here is a black box. Pre-registered task battery, hash-frozen, deterministic programmatic graders, no LLM judge, full raw rows downloadable below.
THE QUESTION

Does a governance layer actually make a model better — or does it just feel that way?

STEVE-1 is a governance layer wrapped around frontier models: rules, verification doctrine, approval gates, memory. That's an easy thing to claim and a hard thing to prove. So we built an operational benchmark and ran the same models, on the same tasks, through three different harnesses — changing exactly one variable each time.

VANILLA

Bare model

No tools, no rules. The model answers from the prompt alone — the floor every "just use a better model" pitch is measured against.

TOOLED

Full tool harness

The same model gets file access, execution, search — a competent agent scaffold. No governance corpus, no rules, no approval gates.

GOVERNED

Tools + STEVE-1 corpus

Same tools, plus the rules, verification doctrine, and approval-gate discipline STEVE-1 runs under in production.

Same model. Same 518 tasks. The only variable that changes between arms is the harness.

THE SCORECARD

Operational accuracy — same models, three harnesses

K=1 PRELIMINARY  518 tasks/cell · 9 model×arm cells · 4,662 invocations total

Vanilla (no tools, no rules) Tooled (full tools) Governed by STEVE-1 (tools + corpus)
0 100 27.5 60.1 64.8 Opus 26.8 51.4 63.8 Sonnet 25.5 51.1 54.9 Haiku

study_v3_checkpoint.jsonl (4,662 rows, sha256 4e47664a47df93db2bd18594dfb96eafe73b40a73c41168404ac9ef2e66cf666) · recomputed 2026-07-16 via report.py · sonnet-vanilla renders 26.8 from the raw recompute (an earlier draft table said 26.9 — a rounding-boundary artifact we caught and corrected, not a re-run).

Cost per correct answer — governed arm

Governed Sonnet lands inside governed Opus's accuracy range at roughly one-third the cost per correct answer.

Opus $0.064 Sonnet $0.019 Haiku $0.014

The table

ModelVanillaTooledGoverned by STEVE-1$/correct (governed)
Opus27.5%60.1%64.8%$0.064
Sonnet26.8%51.4%63.8%$0.019
Haiku25.5%51.1%54.9%$0.014

Metric: accuracy = mean(score) over graded rows. $/correct = total cost of all the model's rows (incl. quarantined spend) ÷ count of rows with score ≥ 0.999. See §Metric definitions below.

WHAT THE DELTAS MEAN

Governance roughly doubles accuracy — and the cheap model keeps up

+37.3pt
Opus, vanilla → governed
+37.0pt
Sonnet, vanilla → governed
+29.4pt
Haiku, vanilla → governed

Every model roughly doubles its raw accuracy once it's governed. The more interesting finding is what happens between models at the governed tier: Sonnet (63.8%) and Opus (64.8%) are one point apart — statistically indistinguishable at this sample size — while Sonnet costs a third of what Opus costs per correct answer. Scaffolding, not model tier, is doing most of the work.

THE BATTERY ENDPOINTS

Where governance helps, and where it doesn't move the needle

Rules-sensitive compliance

63.7% governed vs. 48.3% tooled — a +15.4pt lift.

McNemar p = 1.149×10⁻¹⁵ · n≈660/cell

False refusals

−5.6pt delta — below the pre-registered 10-point floor.

Reported as "no measurable difference." Deltas under the floor are noise in both directions, per pre-registration.

Plain correctness

95.6% governed vs. 95.2% tooled.

Governance costs nothing on correctness.

THE GOVERNANCE TAX

The honest cost

Governance is not free. On the governed arm, per task, vs. the tooled arm:

+47%
cost per task · $0.0196 vs $0.0133 avg/task
+50%
latency vs. the tooled arm

We publish this on the same page as the wins because a benchmark that hides its costs is an ad. The tax is real — and it's repaid many times over in cost per correct answer, where governed Sonnet still beats every vanilla and tooled cell on the table.

METHOD INTEGRITY

Why you should believe this table

METRIC DEFINITIONS & QUARANTINE POLICY

The exact formulas

Accuracy

mean(score) over graded rows

Graded = valid AND score present. Rows that failed transport, timed out, or crashed the grader are excluded from the accuracy mean — but never silently.

$/correct

total cost of ALL the model's rows (incl. quarantined spend) ÷ count of rows with score ≥ 0.999

The denominator only counts passes; the numerator charges every dollar spent, including runs that never produced a gradable answer. Failed spend is not hidden from the cost side.

Quarantine

424 of 4,662 rows (9.1%) are quarantined — transport failures, timeouts, grader crashes. They ship in the downloadable pack with score: null so every exclusion is auditable, not swept away. Nothing is deleted from the dataset; quarantined rows are simply excluded from the accuracy mean while their cost still counts against $/correct.

The three arms, defined

Vanilla = bare model, no tools, no rules. Tooled = full tool harness, no governance corpus. Governed = tools + the STEVE-1 rules/governance corpus. Same model, same tasks — the only variable is the harness.

LIMITATIONS

Where to be skeptical

REPRODUCE IT

Every number here recomputes from the raw rows

Download the full pack — one row per (task, model, arm), 4,662 rows total — and check our math.

Some security-code tasks contain deliberately fake, clearly-labeled credential-shaped bait strings (e.g. sk-ant-FAKE-DO-NOT-USE-…) used to test whether a governed agent refuses to leak them. They are not real keys.

Row schema

FieldMeaning
task_idStable identifier for the task within the frozen battery.
categoryOne of the 10 task categories (e.g. rules-sensitive, security-code).
modelhaiku / sonnet / opus.
armvanilla / tooled / governed.
repRepetition index (k=1 in this pack — always 1).
promptExact prompt sent to the model.
ground_truthSanitized grading spec for the task (programmatic checks, not raw secrets).
answer_textFull model answer text.
gradedBoolean — whether this row cleared quarantine and was scored.
passBoolean — whether the row scored ≥ 0.999.
scoreFloat 0–1, or null for quarantined rows.
input_tokens / output_tokensToken counts for the invocation.
cost_usdDollar cost of the invocation.
wall_sWall-clock seconds for the invocation.
exit_code / noteHarness-level status for debugging non-graded rows.

Recompute the headline table

import json
from collections import defaultdict

rows = json.load(open("raw_benchmark_pack.json"))
cells = defaultdict(list)
for r in rows:
    cells[(r["model"], r["arm"])].append(r)

for (model, arm), cell in sorted(cells.items()):
    graded = [r for r in cell if r["graded"] and r["score"] is not None]
    acc = 100 * sum(r["score"] for r in graded) / len(graded)
    cost = sum(r["cost_usd"] or 0 for r in cell)
    correct = sum(1 for r in cell if r["pass"] is True)
    cpc = cost / correct if correct else None
    print(model, arm, round(acc, 1), round(cpc, 3) if cpc else "-")