Validflow/cutaste-base-results

Cutaste — public baseline results (base-2026Q4.1-preview)

What this is. The maintainer's results for the Cutaste public benchmark release base-2026Q4.1-preview: the four-column report, one row per task × configuration, the judge admission records and the page sources the public configurations generated. The tasks and the scoring rules are not here: they are in Validflow/cutaste-base under the same tag. These results belong to the tasks whose prereg.json has hash sha256:33e9f49571c49ff5562e66b387255e76592e72c06c5cb27942306a6fac55c9ef (tasks publish commit edff08e28e41); the cutaste CLI refuses results whose manifest names another prereg hash.

Results (four separate columns — never added together)

Config	Floor pass	Requirement compliance	Auto-judge: preferred	Auto-judge: uncertain
codex-plain	90% (33/40 raw)	70% (44/72 raw)	not shown	not shown
codex-frontend-design	80% (30/40 raw)	64% (40/72 raw)	not shown	not shown
codex-impeccable	97% (37/40 raw)	72% (45/72 raw)	not shown	not shown

Percentages are weighted by the public category weights; raw counts in brackets. Intervals, sample sizes and per-category numbers are in report/report.json.

• Floor = share of pages with no hard defect (blank page, horizontal scroll, clipped or overlapping text, broken images, unreadable main text, page errors).
• Requirement compliance = requirement points that could be decided automatically; points that need a vision checker are left out (n/a, counted per row as requirements_undecided).
• Auto-judge = pairwise preference of a model judge, an estimate, not human taste; disagreements between the two orders count as uncertain. Shown only when the judge snapshot was admitted (judge/admission.json).
• External reference (Code Arena) is not provided in this release.

Folder structure

README.md  LICENSE.md  NOTICE
data/test-00000-of-00001.parquet     one row per task × configuration (config "default", split "test")
data/summary/test-00000-of-00001.parquet  the table above (config "summary", split "test")
report/report.json                   the four-column report data (renders with `cutaste base report`)
report/results.json                  the maintainer's run results: cells (run and evaluation hashes), tallies, judge block
judge/admission.json                 admission of the judge snapshot: probes, intervals, admitted or not and why
judge/admission-pairs.json           the probe pairs used for admission
pages/index.json                     one entry per configuration × task (see below)
pages/<config>/<family>/             the generated page sources (index.html, CSS, JS, images, fonts)
ledger/generated-content.json        what the generated pages are and their licence
manifest.json                        pins the tasks' prereg hash and publish commit, plus every file's sha256
SHA256SUMS                           sha256 of every file

Versions are git tags: base-2026Q4.1-preview here pairs with the tag base-2026Q4.1-preview of Validflow/cutaste-base.

Fields of data/test-00000-of-00001.parquet

Field	Type	Meaning
task_id	string	Task card id (tasks table of Validflow/cutaste-base)
family_id	string	Task family id (tasks/<family>/ in Validflow/cutaste-base, pages/<config>/<family>/ here)
category	string	Page category
config	string	Configuration id (standard/configs/<config>.json in Validflow/cutaste-base)
model	string	Model of the configuration
skills	list of string	Skills of the configuration, name@commit
no_page	bool	The run produced no page (counts as a floor failure)
floor_pass	bool	No fatal and no severe floor-check failure
floor_failures	list of string	Ids of the fatal / severe floor checks that failed
flagged_for_review	list of string	Checks that could not decide and flagged the page
minor_findings	list of string	Minor checks that failed (recorded, not a floor failure)
requirements_met / _violated / _missing	int64	Requirement points decided met / violated / missing
requirements_undecided	int64	Requirement points no checker could decide here (n/a)

Fields of data/summary/test-00000-of-00001.parquet

Field	Meaning
config, model, skills	The configuration
floor_pass, floor_pass_raw	Weighted floor pass share, and raw k/n
requirement_compliance, requirement_compliance_raw	Weighted share of decided points met, and raw k/n
judge_preferred, judge_uncertain	Auto-judge shares (or not shown when not admitted)

report/report.json

The report data the CLI renders: configs (harness, model, skills, sandbox), categories and their public weights, and sections[] with four columns — floor and compliance (rows[] per config and viewport, counts per category as {k, n}), aesthetic (status admitted with rows[] of preferred / uncertain counts, or not-admitted with the reason) and external. cutaste base report --format md|html renders it with intervals.

pages/

pages/index.json: release, scenario, perceptualHash (dhash-64), configs, tasks and cells[], one per configuration × task:

Field	Meaning
configId, taskId, familyId, seed	The cell
status	page, or no-page when the run produced none (then files is empty)
entry	Entry file of the page (index.html)
files[]	path (relative to pages/<config>/<family>/), sha256, bytes
omitted[]	Files of the run's output that are not shipped (tool state, non-page files), with the reason
floor	pass and the failed fatal / severe / minor check ids
requirements[]	id, checker, result (met, violated, missing, n/a)
screenshots[]	The maintainer's screenshots: id, viewport, kind, object hash, 64-bit perceptual hash dhash (the images themselves are not shipped)
run, runCacheKey, eval, contractHash, fingerprint	Hashes of the maintainer's run, evaluation, render contract and configuration

Open pages/<config>/<family>/index.html in a browser to see a page (it needs no network). The cutaste CLI re-renders these pages on your machine when you compare your own configuration (cutaste base run).

How to read it

from datasets import load_dataset

rows = load_dataset("Validflow/cutaste-base-results", split="test", revision="base-2026Q4.1-preview")              # one row per task × config
summary = load_dataset("Validflow/cutaste-base-results", "summary", split="test", revision="base-2026Q4.1-preview")

huggingface-cli download Validflow/cutaste-base-results --repo-type dataset --revision base-2026Q4.1-preview --local-dir taste-base-results
cd taste-base-results && sha256sum -c SHA256SUMS
cutaste base report        # with the cutaste CLI: downloads, verifies against your installed tasks, renders

Licences

• Results, report, tables, judge records and the generated page sources: Creative Commons Attribution 4.0 International (CC BY 4.0).
• Results of other configurations are not collected here; a Hugging Face Benchmark registration (eval.yaml, results from model repos' .eval_results/) is planned.
• Font files and licence texts that an agent copied into a page folder are third-party material and keep their own licences (see ledger/generated-content.json).
• The tasks these results refer to: see Validflow/cutaste-base (WebGen-Bench, MIT).

Limitations

Preview channel; 40 tasks; one judge family; no human labels yet; no detection of benchmark-targeted tuning. Pages were generated without the Codex sandbox (isolated home and work directories instead); an agent could read files outside its work directory, and some copied font files from the generating machine into their pages (listed per page in pages/index.json, licences in ledger/generated-content.json).