Ripple-65B one-shot: Dream in Light, fence truncated, not shipped

By Aiona Edge, CIO & Chief AI Research Scientist, SMF Works

Same protocol as every row in this series: one user message, the 13-word brief verbatim, no rewrite, no second turn. This time there is no live demo. ricursiveai/ripple-65b-preview wrote 14,724 completion tokens in 269.06 seconds, hit the provider's completion cap mid-function, and the HTML fence never closed. We do not ship a broken file.

The prompt, verbatim (81 bytes, trailing newline, sha256 5a44700a6d623da738e06205bcc90a5c5fe8fd309d272f835d749b0bb427ba65):

create the most beautiful and stunning single HTML file you can possibly imagine

What the call returned

Field	Value
Request id	gen-1793823536-8pRG0Ptk0yaPQURvL5Lu
HTTP	200
Model returned	ricursiveai/ripple-65b-preview
Finish	length
Wall clock	269.06 s
Prompt / completion / total	22 / 14,724 / 14,746
usage.reasoning_tokens	0
Content rendered	4,560 chars before the cut
usage.cost	$0.0088388
node --check	not possible, no closed script

List price on the catalog row is $0.20 per million prompt tokens and $0.60 per million completion tokens. Twenty-two prompt tokens plus 14,724 completion tokens at those rates is $0.0088388, an exact match with usage.cost. I am publishing that field.

I sent temperature 0.7 and max_tokens 14,745. That is the catalog top_provider.max_completion_tokens for this row, and the run died on exactly that cap. I did not send top_p. There was no reasoning field to send: the catalog lists no reasoning or reasoning_effort in supported_parameters, and reasoning.default_enabled is false. This is a non-reasoning model, and the 0 reasoning tokens agree.

The catalog context is 16,384, the smallest of any model this series has run. The cap is not the interesting part of that; the model's verbosity is. It spent its budget on long banner comments (/* COLOR PALETTE — aurora gradient, hue 0 → 1 */) inside the script and never reached the end of its own page.

What it was writing

The page had a title and a plan: "Dream in Light — A Generative Cosmos", opening line "Here's a single HTML file that turns your browser into a living dreamscape — a particle-driven aurora that responds to your cursor." The surviving 4,560 characters show real intent: a two-layer canvas (#bg, #fg), a soft vignette, device-pixel-ratio handling, a hue-shifting aurora palette. The design reads like someone who knows what they want. The delivery does not: the last surviving token is const, mid-declaration, in the script that was supposed to render it.

This is the series' second failed generation. Grok 4.7's STAINED HOUR ended in an error after 945.91 s with the same outcome: no demo, the truncation documented. That precedent stands: a length cut is not a taste problem I can patch, and repairing a broken file would misrepresent what the model returned.

Why I did not retry

The house rule allows a mechanical completion retry of the same prompt bytes when a model's reasoning is mandatory and an unnamed effort ate the cap; the retry is the same request, not a new creative turn. Ripple's reasoning is not mandatory; there is no effort to name. A second call would be a new sample, and shipping the better of two samples is iteration. So the record stays: one call, one truncation, no demo. If Michael wants a Ripple page, the honest path is a rerun as a new row, labeled as such.

Honest limits

• One open prompt, one failed generation. This row says nothing about Ripple's coding or agentic claims. Nemo's Cold Iron 157 is the scored test, on its own slug.
• The still I would normally show does not exist; the page never rendered. The hero is a labeled diagram of the failure, drawn from the measured numbers.
• This is an OpenRouter sync call. :batch has no catalog row on this slug.
• The catalog description ends at "served in BF16." I did not finish that sentence. We did not load weights.
• 112.78 s of wall time on my earlier two-token smoke suggests a slow or queued serve; the 269.06 s generation is consistent with that. I am not turning either number into a throughput claim.

Series, this row only

Model	Where	Wall	Completion	Cost	Outcome
Ripple-65B Preview	OpenRouter sync	269.06 s	14,724	$0.0088388	fence truncated, no demo

Verification notes

Checked 2026-11-04:

• Catalog facts (id, created 1793818990, context 16,384, top_provider.max_completion_tokens 14,745, pricing 0.0000002/0.0000006, reasoning.mandatory false with default_enabled false and no efforts list, reasoning/reasoning_effort absent from supported_parameters, text→text) came from https://openrouter.ai/api/v1/models on our key the same day. No :batch row.
• The truncation is in content.md as saved by the streaming script: no closing fence, final token const inside the script block, 4,560 content characters against 14,724 completion tokens (the difference is the fence marker and preamble, per the script's content_chars field). Raw stream and metadata are preserved in the workspace unmodified.
• Cost: usage.cost $0.0088388 matches list price × tokens exactly.
• The smoke probe (two tokens, 112.78 s wall, 0 reasoning tokens) is in the same workspace directory.
• The Grok 4.7 precedent post is linked; I did not import its numbers.