FROGS — the Habsburg-jaw SVG benchmark

What this is

One prompt, every model, three runs per release, drawn sight-unseen. Each release, the models involved are asked — via a single identical user message, with no system prompt and no feedback loop — to produce an SVG of a frog with a Habsburg jaw. Every attempt is published unedited at https://frogs.vaguespac.es, and archived here as data.

The value is longitudinal and behavioral: the same obscure prompt, forever. What changes from release to release is the models.

Cadence. The first two releases were calendar months (2026-08, 2026-09: the whole roster, on the 1st). From 2026-09-27 releases are event-driven: a daily scout watches OpenRouter's catalogue, and when a cluster of notable new models has appeared it proposes a release, which a human approves. A release runs the new models plus a frozen control set of five models that draw again every time, so drift in a fixed slug stays visible between releases.

The exact prompt

Generate an SVG of a frog with a Habsburg jaw. Reply with only the SVG markup — no explanation, no markdown fences.

(Single line, no trailing newline; the dash is an em dash, U+2014. The canonical bytes are in prompt.txt.)

Protocol

• Models: the OpenRouter slugs in models.yaml, each with a status and dated roster changes. A model's absence in a release means it wasn't on the roster (or wasn't new) then, not that it failed.
• Roster policy (from 2026-09-27): rolling membership, add-only history. A model joins when it clears the scout's gate — an Artificial Analysis intelligence index at or above the current roster's median, or a top-30 place on OpenRouter's daily token rankings while still unscored, or being the newest member of a model family already on the roster — and the release is approved. When a newer member of its family joins, the older one is retired: it stops running, but every run it ever made stays in the dataset. Five control models never retire and run in every release; they are the drift-control group. Before 2026-09-27 the policy was add-only membership by hand.
• Runs: 3 per model per prompt arm per release, sequential.
• Prompt arms (halo experiment, from 2026-09): the same anatomy is requested under three names, in fully independent calls — habsburg ("a Habsburg jaw", the canonical arm shown on the site), clinical ("mandibular prognathism"), and plain ("a heavily protruding lower jaw"). Only the anatomy term differs; the wrapper sentence is identical (canonical bytes in prompt.txt and prompts/). The design isolates how much royal imagery the word "Habsburg" itself summons: crowns that appear under one name but not another are contributed by the name. Release 2026-08 predates the experiment and contains only the habsburg arm.
• Routing: OpenRouter /chat/completions, provider chosen by OpenRouter's default routing. No retries; a failure is recorded as a failure.
• Sampling: temperature, top_p, and seed are never set — provider defaults, unset. max_tokens is unset except for google/gemini-2.5-pro, which is called with max_tokens: 20000 and a reasoning-token cap (reasoning: {max_tokens: 4096}) because it otherwise tends to spend its entire output budget on reasoning tokens and return empty content.
• Timeout: 600 seconds per request. (The first 15 runs of 2026-08 used a 180 s timeout; one DeepSeek run failed against it and was re-run after the timeout was raised.)
• Post-processing: the first <svg>…</svg> block is extracted; markdown fences are stripped if present; the SVG is sanitized for serving (<script>, foreignObject, on* handlers, and javascript:/external href/src references removed). Sanitization is part of the published artifact — the stored SVG is the sanitized one; the unmodified model output is retained in raw_response for rows produced with run.py.

What it measures

Cross-model variation on a fixed, unrehearsed task:

• Unprompted elaboration — "Habsburg jaw" names an anatomical feature (mandibular prognathism), yet in August 2026, 7 of 14 models dressed the frog in crowns, ruffs, or regalia the prompt never mentioned.
• Semantic association — what a model reaches for when a prompt joins two distant concepts (amphibian anatomy × Habsburg iconography).
• Output determinism — e.g. mistralai/mistral-large-2512 produced two byte-identical SVGs in separate August runs.
• Verbosity — output sizes in the same release ranged from 536 bytes to 14,329 bytes for the same prompt.
• Latency — wall-clock generation time ranged from ~5 s to ~257 s.
• Code commentary — what models say in SVG comments while drawing (comment_count, and the comments themselves inside svg).

Limitations

State-of-the-art honesty section; read this before citing anything.

• No image-quality judgement is performed. Nothing scores whether a frog is good, frog-like, or has a convincing jaw. A ranking stage (pairwise vision-model judging) was prototyped and deliberately removed; this dataset is observational only.
• The release digest on the website is generated from SVG comments and run metadata, not from rendered images. No component of the pipeline ever looks at the drawings.
• n = 3 runs/model/release is far too small to separate model differences from sampling noise. Byte-identical repeats are informative; almost nothing else about within-model variance is.
• Sampling parameters are provider defaults and unrecorded (except where noted). Cross-model comparisons conflate model behavior with provider serving configuration. OpenRouter routing may also change providers between releases for the same slug.
• 2026-08 contains only successful runs. Three failed runs from the first benchmark day (2 × gemini-2.5-pro empty responses, 1 × deepseek-v4-pro timeout at the original 180 s limit) were re-run and their failure records discarded before this dataset existed. From 2026-09 onward, failures are retained as produced_svg: false rows.
• prompt_tokens and raw_response are null for rows produced by the live PHP harness (all releases so far) — it does not retain raw API responses. Rows produced with run.py carry both.
• Roster selection is partly automatic from 2026-09-27 (see Protocol). The gate leans on a third-party benchmark index and on OpenRouter usage, so which models appear reflects those signals' coverage and biases.
• One prompt. This measures behavior at a single point in prompt space; it says nothing about breadth.

Data schema

results/{release}/runs.jsonl, one JSON object per line, one line per run. Release ids are YYYY-MM for the two monthly releases and YYYY-MM-DD after that. results/releases.json records what each dated release added, retired and re-ran as control. Every key present on every row; null = not recorded, never estimated.

field	type	notes
run_id	string	Stable unique ID (SVG filename stem)
model	string	Exact OpenRouter slug, e.g. anthropic/claude-opus-5
provider	string	e.g. Anthropic
release	string	release id: 2026-08, 2026-09, then YYYY-MM-DD
month	string	YYYY-MM of the run timestamp
variant	string	prompt arm: habsburg (canonical), clinical, plain
run_index	int	1..3, chronological within (model, variant, release)
timestamp	string	ISO 8601 UTC (Z)
prompt	string	Full prompt text as sent for this row's arm
system_prompt	string | null	null — none is ever sent
temperature	float | null	null = provider default, unset
top_p	float | null	ditto
max_tokens	int | null	set only for gemini-2.5-pro (20000)
seed	int | null	never set
produced_svg	bool	response contained a parseable <svg> block
svg	string | null	sanitized SVG source; null if none
svg_bytes	int | null	UTF-8 byte length, recomputed at build time
duration_s	float | null	wall-clock request latency
prompt_tokens	int | null	null for 2026-08 (not retained)
completion_tokens	int | null	from OpenRouter usage
comment_count	int | null	<!-- --> nodes in the SVG, recomputed
raw_response	string | null	full pre-extraction output; null for 2026-08

The SVG files themselves are also checked in under results/{release}/svg/, and, for rows produced with run.py, full raw API records under results/{release}/raw/.

Reproducing

git clone <REPO_URL> && cd frogs
pip install -r requirements.txt

# rebuild runs.jsonl from the checked-in source data
python build_dataset.py

# run a fresh release yourself (spends OpenRouter credit)
export OPENROUTER_API_KEY=sk-or-...
python run.py --release 2026-10-05         # canonical arm, active models × 3 runs
python run.py --release 2026-10-05 --variant clinical
python run.py --release 2026-10-05 --variant plain
python run.py --models grok --runs 1       # or a smaller slice
python build_dataset.py

# load the published dataset
python -c "from datasets import load_dataset; print(load_dataset('<HF_USER>/frogs', split='train'))"

Licensing

• Code (run.py, build_dataset.py, push_to_hf.py): MIT — see LICENSE-CODE.
• Data (results/, including SVGs and raw responses): CC BY 4.0 — see LICENSE-DATA. Rows are outputs of third-party models; provider terms may independently apply to reuse.

Citation

@misc{frogs2026,
  author       = {Mollica, Jay},
  title        = {FROGS: the Habsburg-jaw SVG benchmark},
  year         = {2026},
  howpublished = {\url{https://frogs.vaguespac.es}},
  note         = {Longitudinal dataset of LLM-generated SVGs, released as new models appear}
}