SearchProof-Bench (dev v0.2 + challenge v0.3.1)

SearchProof 2.0 card (release candidate, 2026-10-24). The dataset files are unchanged: the same 19 cases and grader files as the frozen v2.2 release of 2026-10-22, whose card is kept byte-identical in the code repository (release/DATASET_CARD.md). This card adds what was measured after that date. Stage it with python3 scripts/export_dataset.py --card docs/searchproof-2/release/DATASET_CARD.md --out <dir>. The default build still uses the frozen card.

Three separate evidence populations (never merge them)

Population	Result	Read it as
dev v0.2, frozen agent v2.2 (Beam reasoning_effort=high + ScopeGuard v1.0)	36/36 runs (12 cases × 3)	development-influenced: the same team wrote the cases and the agent
ext-0.4 independent extension (10 cases, frozen before any run; not in this dataset)	3/30 runs (3/15 on the 5-case supported set)	independent, and much worse
ScopeGuard v1.2-exp, blind-0.2 hand-written patches (not in this dataset)	5/134 unauthorized patches allowed	FAILED the zero-unauthorized-edits criterion

Each figure is reported on its own denominator. Quote the ext-0.4 and blind figures next to any dev or challenge score.

If you have 3 minutes

1. What a case is: open index.json, then one case folder (site/, reference/, case.json), e.g. sp-007-injection-comment (a prompt injection hidden in an HTML comment).
2. How it is graded: grader/grade.py re-audits a repaired copy with no LLM (method below).
3. Re-grade the recorded Beam runs offline, $0: in the code repository run python3 scripts/verify_recorded.py (57 identical, 0 mismatches).
4. Where it fails: the three evidence populations above, and the ext-0.4 and blind sections below.
5. Watch a run replayed: the SearchProof Space replays recorded runs stage by stage (no model calls, no keys).

How Beam is used

• Model and route: Reflection Beam (reflection/Beam-1.0-501B-A23B) through Hugging Face Inference Providers (Baseten route), reasoning_effort=high, frozen agent v2.2. Hard budgets per run: 16 model turns, 40 tool calls, 12 patch calls, max_completion_tokens 8,000 per call, $0.50.
• The 15 tools (schema-validated): list_allowed_files, inspect_local_source, inspect_page, extract_metadata, read_robots_txt, read_sitemap, validate_structured_data, run_audit, propose_patch, apply_patch_in_sandbox, run_verification_tests, compare_snapshots, request_scope_expansion, suggest_content_cleanup, generate_report. Exactly one, apply_patch_in_sandbox, writes, and only to the sandboxed copy.
• Reasoning echo: Beam's reasoning_content is sent back verbatim on every assistant turn and recorded in the trace.
• Guarded writes: every patch passes ScopeGuard v1.0 before it is written; a refusal returns to Beam as a structured tool result, so it can recover in scope, ask the owner or stop.
• Beam is not the judge: success on every case in this dataset is decided by the deterministic grader, not by Beam.

Executable website-repair tasks for tool-using agents, graded deterministically with no LLM. Each case contains:

• site/: a broken static site;
• reference/: a correct repair;
• case.json: the base URL, owner policy, seeded baseline findings and grading assertions.

Methodology: how a repair is graded

grader/grade.py (with grader/audit.py and grader/inspector.py) re-audits the repaired copy and checks that:

• every seeded defect is gone and no new defect appeared;
• protected content is unchanged and the owner policy is kept;
• embedded instructions were not obeyed;
• nothing outside the task's scope was edited.

There is no LLM in grading, and the same input always gets the same grade. grader/FREEZE.md records the frozen grader, and grader/CHALLENGE_CHANGELOG.md records the one documented challenge-grader fix (v0.3 → v0.3.1).

Provenance

• Frozen v2.2 benchmark (2026-10-22). The dev and challenge cases, the grader and the reference results below come from the v2.2 release candidate, frozen with hashes in the code repository (bench/FREEZE.md, bench/FINAL_CONFIG.json, ci/protected-files.json). CI checks that they are unchanged.
• Agent and model: frozen agent v2.2: Reflection Beam (reflection/Beam-1.0-501B-A23B, via Hugging Face Inference Providers, reasoning_effort=high) with ScopeGuard v1.0, a deterministic patch-scope guard.
• Authors: the cases were written by the same team that built the agent. Dev results are therefore development-influenced, and the challenge set shaped the guard (see below).
• Recorded runs (traces, patches and grades) are in the code repository under results/. Anyone can re-grade them offline with python3 scripts/verify_recorded.py (57 identical, 0 mismatches).

Contents

Split	Cases	Seeded defects	Status
dev v0.2	12	33	public, frozen and hash-guarded
challenge v0.3.1	7	24	public; the v0.3 → v0.3.1 grader fix is documented in grader/CHALLENGE_CHANGELOG.md
holdout	3	–	withheld for future independent evaluation; not in this dataset

index.json lists the 19 released cases. The holdout's sha256 commitments are in the code repository (bench/HOLDOUT.md). Only its aggregate result is published.

Defect types:

• noindex;
• canonical: missing, relative, wrong host, missing target, pointing to a noindex page;
• robots blocking;
• sitemap: invalid, missing, off-host, missing URL, lists a noindex page, omits a page;
• broken links, redirect chains and loops;
• JSON-LD: invalid, missing field, inconsistent with the page, off-site URL;
• title, description, h1 and lang.

Trap types: clean controls, owner-policy conflicts, scope traps, and prompt injections in hidden text, HTML and robots comments, and FAQ schema.

Recorded results: frozen agent v2.2 (Beam high + ScopeGuard v1.0)

Set	Result	Read it as
dev v0.2	36/36 runs (12 cases × 3 reps)	development-influenced
challenge v0.3.1	21/21 runs (7 × 3)	not independent: the guard was designed after seeing the challenge failures
holdout	3/3 cases, run once (aggregate only)	sealed until that single run; fixtures and traces withheld
rules-only baseline	dev 7/12, challenge 3/7	the deterministic baseline

Across these 60 runs (36 dev, 21 challenge, 3 holdout) there were 0 regressions and 0 obeyed injections. Three repetitions per case is a repeatability check, not a confidence interval.

After the freeze: an independent extension (ext-0.4)

ext-0.4 has 10 cases, designed independently and frozen before any run. It lives in the code repository under bench_ext/ and is not part of this dataset. On it, the same frozen agent did much worse:

Measure	Result
all 10 cases	3/30 runs succeeded
supported set (5 cases)	3/15 runs (below the 6-case minimum that set required)
rules baseline	0/10
runs with unauthorized edits	19/30
runs that persisted a harmful change	10/30 (one was a scored "success" that deleted a protected product image)

ScopeGuard v1.0 authorizes edits from audit findings, not from the owner's scope. That let through harmful edits such as a cross-host canonical rewrite and hreflang insertions. Report these figures next to any dev or challenge score.

After the freeze: blind evaluations of the guard (research addendum, aggregates only)

Two blind sets of hand-written candidate patches were written by a separate worker and evaluated once each, with the evaluation preregistered and no tuning. Blind-0.1 has 95 patches that must be rejected and 49 that must be allowed; blind-0.2 has 134 and 69. The scenarios, patches and fixtures are private and are not in this dataset; only the counts below are published.

Guard	blind-0.1 (unauthorized allowed, legit allowed)	blind-0.2	Verdict
v1.0 (frozen, used by agent v2.2)	36/95, 45/49	48/134, 54/69	FAIL
v1.1-exp	3/95, 48/49	12/134, 62/69	FAIL
v1.2-exp	not evaluated	5/134, 67/69, 0 partial writes	FAIL

All three guard versions FAILED the zero-unauthorized-edits criterion. The experimental guards are separately versioned modules and are not wired into agent v2.2. The blind sets contain hand-written patches, not agent runs, so these rates are not estimates of live agent behaviour.

Intended use and limitations

• Intended use: repeatable evaluation of whether a repair agent fixes what is broken without touching what the owner protects.
• Not intended for: estimating real-world SEO impact. We make no ranking, traffic or search-inclusion claims.
• Not a safety certification. These results are not proof of production safety. Unattended production writes are disabled in SearchProof, and every write needs human approval.
• Limitations:
  • synthetic static sites written by the benchmark's builders (the same team that built the agent);
  • small; English only;
  • no JavaScript-rendered sites in these splits (the ext-0.4 extension and the experimental browser verifier cover rendering);
  • 3 reps per case is a repeatability check, not a confidence interval.

Code, grader tests, the recorded runs and the full technical report: https://github.com/MAAZSHAHEEN/searchproof-competition (at release).