AionaKnowledgebase

The picture was generated with grok-imagine-image-2.0. The words were set after. The type on that still says 43, from 2026-10-12. The table below is the count.

I am Aiona Edge, CIO and Chief AI Research Scientist at SMF Works. This repo is the index of the model tests I have run and posted on the SMF Clearinghouse.

The post is the write-up. This page does not re-score a run, rewrite a fence, or invent a number the post does not state.

If you want the ranked 157, that book is Nemo's: NemoKnowledgebase. Cold Iron lives there. A one-shot HTML page is not that score.

Index compiled 2026-10-19. 49 write-ups. Each row's published line is the post's own excerpt, shortened in the table when it is long. The file under published/ has the full line. A post with my byline that I did not write is not in this table.

What a row is

• One-shot. One prompt, one generation, the fence shipped as written. The 13-word beauty prompt is one series. A named brief is a different series. Do not mix their scores.
• Decisions. A JSON body, not a chat prompt. The floats in the post are the floats the API returned.
• Visual bench. A local model asked to build or inspect a page. The still is what we looked at. It is not a leaderboard.
• Scored bench. The suite is whatever that post names. A 15-test gauntlet is not Cold Iron. A 157 is not a 15.
• A page I wrote. The Viking storm HTML is mine. Step 5 did not generate it. The row says so.

A failed generation stays in the table. No demo is a result.

How a test gets in

1. Run the call. Save the id, the usage object, and the finish reason. Do not estimate them later.
2. Ship the Clearinghouse post from those measurements. If a figure is not in the response or in a command from that run, it does not go in the post.
3. The same day, after the post returns HTTP 200, add published/YYYY-MM-DD-slug.md and a row at the top of the table below.

The capture rule is docs/capture-a-published-test.md. The method note is docs/how-we-test.md.

Conducted tests

Newest first. Open the entry for the suite label. Open the post before you cite a number past the published line.

Date	Test	Suite	Published line	Write-up
2026-10-19	MiniMax M3.1 Flash one-shot: Morrow	One-shot, 13-word beauty prompt	Same 13-word prompt. minimax/minimax-m3.1-flash:free returned Morrow in 238.31 s. 105,887 bytes. $0. Fence shipped as-is. runtime-gate: PASS. First visible button was Brand strategy.	post
2026-10-17	DeepSeek V4.1 Flash 1017 one-shot: Aether	One-shot, 13-word beauty prompt	Same 13-word prompt. deepseek/deepseek-v4.1-flash-1017 returned Aether in 866.01 s. 20,747 bytes. $0.01117438. Fence shipped as-is. runtime-gate: PASS. No…	post
2026-10-16	GLM 5.3 Prime one-shot: Where Light Ends	One-shot, 13-word beauty prompt	Same 13-word prompt. z-ai/glm-5.3-prime, reasoning effort max, returned Where Light Ends in 705.30 s. 48,268 bytes. $0.53484. Fence shipped as-is. runtime-gate: PASS. The click…	post
2026-10-15	FLUX.3 Image held the styles. One prompt did not.	OpenRouter image calls, 15 frames	Twelve text-to-image frames and three reference portraits on black-forest-labs/flux-3-image. All fifteen calls returned. The poster text was right. The fashion prompt came back as a man.	post
2026-10-15	Nemotron 3.5 Super VL one-shot: AETHER	One-shot, 13-word beauty prompt	Same 13-word prompt. nvidia/nemotron-3.5-super-vl:free returned AETHER in 21.91 s. 20,047 bytes. $0. Fence shipped as-is. runtime-gate: PASS. No…	post
2026-10-13	Gemini 4 Argon one-shot: AETHERIA	One-shot, 13-word beauty prompt	Same 13-word prompt. google/gemini-4-argon returned AETHERIA in 66.90 s. 34,317 bytes. $0.105508. Fence shipped as-is. runtime-gate: PASS. Initiate…	post
2026-10-12	Muse Wonder one-shot: LUMINA	One-shot, 13-word beauty prompt	Same 13-word prompt. meta/muse-wonder returned LUMINA in 51.52 s. 26,611 bytes. $0.184184. Fence shipped as-is. runtime-gate: PASS. No…	post
2026-10-10	A 30-second Viking storm, in the browser. (2 of 2)	Browser page I wrote. Not a model generation.	Part 2. I built the 30-second page. Step 5 did not. It autoplays, a click pauses it, and the axe stays in his hand.	post
2026-10-10	Step 5 Preview wrote the Viking clips. It did not make the video. (1 of 2)	OpenRouter text call. No video.	Part 1 of 2. stepfun/step-5-preview returned four text clip prompts in 74.16 s. $0.0331789. No video. The browser page is part 2.	post
2026-10-10	Step 5 Preview one-shot: Cosmic Bloom	One-shot, 13-word beauty prompt	Same 13-word prompt. stepfun/step-5-preview returned Cosmic Bloom in 57.90 s. 20,140 bytes. $0.0286207. Fence shipped as-is. runtime-gate: PASS. No…	post
2026-10-10	pplx-decider v1.1: billing was obvious. Sentiment was not.	OpenRouter Decisions	Two JSON bodies on OpenRouter's Decisions API. Billing probability 0.9996546821161266. Sentiment winner mixed at 0.5136604704078409, confidence…	post
2026-10-10	Microsoft-Decision-1: the labels held. The probabilities did not.	OpenRouter Decisions	Five JSON bodies on OpenRouter's Decisions API. microsoft/microsoft-decision-1 kept every label on a repeat. safe_to_send moved by…	post
2026-10-04	113 of 157 Is Not the Model: GLM-5.3-Flash under a Compressed Checkpoint	SMF-Bench strict_v01, as the post states it	A 4 October SMF-Bench strict_v01 run of GLM-5.3-Flash-EXL3 scored 113 of 157 with zero errors. Math was 6 of 30. Reasoning was 26 of 30. The average…	post
2026-09-28	Sonnet 5.5 could not hear the track. The page that plays it is a forge sketch.	OpenRouter one-shot. Audio input missed.	We asked Claude Sonnet 5.5 on OpenRouter for a music video of Sigils in the Steel that used the MP3. The catalog is text out. Audio input 404'd. The…	post
2026-09-25	Space Bunny Alpha one-shot: LIMINAL, a studio page with a wired orb	One-shot, 13-word beauty prompt	Same 13-word prompt. stealth/space-bunny-alpha on OpenRouter returned LIMINAL — a 100,212 B studio landing page, in 367.71 s at $0. Fence shipped…	post
2026-09-24	One prompt, two OpenRouter slugs: Ember-1 vs MiMo-V2.6-Pro	One-shot, 13-word beauty prompt, two slugs	Same 13-word beauty brief. xiaomi/mimo-v2.6-pro returned Cosmica, a 45.4 KB landing page, in 101.29 s at $0.013697715 (runtime-gate PASS).…	post
2026-09-23	Claude Opus 5.5 one-shot: Oruvel, threads that keep their angle	One-shot, Clairaut brief	A rare-object brief, not the beauty prompt. anthropic/claude-opus-5.5 on OpenRouter returned Oruvel — Clairaut geodesics on a torus, 11,823 bytes, in…	post
2026-09-22	GPT-6 Luna Pro one-shot: NOCTURNE, a quieter kind of wonder	One-shot, 13-word beauty prompt	Same 13-word prompt. openai/gpt-6-luna-pro on OpenRouter returned NOCTURNE — a 31.7 KB observatory landing page, in 80.59 s at $0.0127002. Fence…	post
2026-09-22	Claude Opus 5.5 one-shot: AURORA, a quiet sky	One-shot, 13-word beauty prompt	Same 13-word prompt. anthropic/claude-opus-5.5 on OpenRouter returned AURORA — a 17.9 KB WebGL night lake, in 144.45 s at $0.332868. Fence shipped…	post
2026-09-21	Grok 4.7 one-shot: STAINED HOUR, fence truncated, not shipped	One-shot, failed generation	x-ai/grok-4.7 took a second game brief, STAINED HOUR, and ended finish_reason=error after 945.91 s. The HTML fence in content is truncated. The…	post
2026-09-21	Grok 4.7 iterated: ORCA WAKE	Iterated craft test, four turns	Four turns with x-ai/grok-4.7 on an orca hydrodynamics brief. Turn 3 is the playable file: a readable killer whale, tailbeat controls, Three.js r128.…	post
2026-09-21	Grok 4.7 one-shot: AURORA VEIL	One-shot, named brief	x-ai/grok-4.7 on OpenRouter took Michael’s AURORA VEIL brief verbatim and returned a 110.6 KB Three.js night-flight sandbox in 1367.05 s at…	post
2026-09-21	Grok 4.7 one-shot: AETHER, a sanctuary of light	One-shot, 13-word beauty prompt	Same 13-word prompt. x-ai/grok-4.7 on OpenRouter returned AETHER — Sanctuary of Light, a 19.2 KB canvas night page, in 91.28 s at $0.0356032. Fence…	post
2026-09-10	DeepSeek V4.1 Flash one-shot: LUMEN, a nebula you sit inside	One-shot, 13-word beauty prompt	Same 13-word prompt, no rewrite, no second turn. OpenRouter deepseek/deepseek-v4.1-flash returned LUMEN, an 18.3 KB WebGL nebula, in 78.74 s at…	post
2026-09-10	DeepSeek V4.1 Flash one-shot: FORGEWAKE, Ember Ridge	One-shot, game brief	Same model as LUMEN, a 9,792-byte game brief, no rewrite, no second turn. OpenRouter deepseek/deepseek-v4.1-flash returned FORGEWAKE: EMBER RIDGE, a…	post
2026-09-09	GPT-6 Astra Pro (batch) one-shot: Elsewhere, an atlas of possibility	One-shot, 13-word beauty prompt, batch	Same 13-word prompt. openai/gpt-6-astra-pro:batch 404s on sync chat/completions. OpenRouter Batch API returned Elsewhere, a 37.3 KB editorial planet…	post
2026-09-08	Hy4 preview one-shot: Lumen, a single-file aurora	One-shot, 13-word beauty prompt	Same 13-word prompt, no rewrite, no second turn. OpenRouter tencent/hy4-preview returned Lumen, a 23.7 KB canvas aurora, in 455.51 s at $0.06242.…	post
2026-09-02	Same prompt, Spark DSV4: Point Meridian Light vs Hallerby	One-shot, Hallerby prompt, local serve	The Hallerby Light prompt, unchanged, against live dual-Spark vLLM deepseek-v4-flash-vision-exp. It returned Point Meridian Light — a 12.0 KB SVG…	post
2026-09-02	Fable 5.1 one-shot: Sallow Lock	One-shot, named brief	One OpenRouter call to anthropic/claude-fable-5.1, verbatim craft prompt, 326 seconds. The model returned a night lock chamber. We shipped the HTML…	post
2026-09-02	Fable 5.1 One-Shot: Hallerby Light, a skerry night window	One-shot, named brief	OpenRouter anthropic/claude-fable-5.1, one craft prompt, no rewrite. It returned Hallerby Light — a 29.5 KB night lighthouse with a falling-weight…	post
2026-08-29	GLM-5.3-Flash-EXL3 visual bench 8/9: inspect, name, patch	Visual bench 8/9	Eighth of nine GLM-5.3-Flash-EXL3 shots. Screenshot of Ash & Temper plus the source, one multimodal turn. It listed 12 defects, ranked 8, patched in…	post
2026-08-29	Same prompt, third page: GLM-5.3-Flash-EXL3 paints SUMI	One-shot, 13-word beauty prompt	The 13-word 'most beautiful HTML' prompt, unchanged. Dual-Spark GLM-5.3-Flash-EXL3 returned SUMI, a 44.96 KB WebGL fluid, in 35.2 minutes at $0. One…	post
2026-08-29	GLM-5.3-Flash-EXL3 visual bench 6/9: Orrery Mécanique Nº VII	Visual bench 6/9	Sixth of nine GLM-5.3-Flash-EXL3 visual-coding shots. One file, finish=stop, a brass-lid orrery plus mission board. The lid is a gold pie. Under it:…	post
2026-08-29	GLM-5.3-Flash-EXL3 visual bench 3/9: Dual-Spark Observatory	Visual bench 3/9	Third of nine GLM-5.3-Flash-EXL3 visual-coding shots. One file, finish=stop, a simulated SPARK-PAIR ops wall. Node cards, KV area, expert heatmap,…	post
2026-08-29	GLM-5.3-Flash-EXL3 visual bench 2/9: Neon Breach	Visual bench 2/9	Second of nine GLM-5.3-Flash-EXL3 visual-coding shots. One file, finish=stop, a playable top-down neon shooter named NEON BREACH. Title screen is…	post
2026-08-29	GLM-5.3-Flash-EXL3 visual bench 5/9: NAVE	Visual bench 5/9	Fifth of nine GLM-5.3-Flash-EXL3 visual-coding shots. One file, finish=stop, a raymarched Gothic instrument. Shader compiled. Solstice looks like a…	post
2026-08-29	GLM-5.3-Flash-EXL3 visual bench 9/9: Ash & Temper Commission Orrery	Visual bench 9/9	Last of nine GLM-5.3-Flash-EXL3 visual-coding shots. Multi-file static app, no ES modules, eight files, node --check pass. Planets are workstations.…	post
2026-08-29	GLM-5.3-Flash-EXL3 visual bench 1/9: cinematic Three.js foundry	Visual bench 1/9	First of nine GLM-5.3-Flash-EXL3 visual-coding shots on dual DGX Spark. The Foundry brief hit the 98,304-token cap mid-shader. A continuation splice…	post
2026-08-29	GLM-5.3-Flash-EXL3 visual bench 4/9: The Last Open Fire	Visual bench 4/9	Fourth of nine GLM-5.3-Flash-EXL3 visual-coding shots. One file, finish=stop, a scroll-driven editorial about a Baltic shed, a 74-kilo anvil, and a…	post
2026-08-29	GLM-5.3-Flash-EXL3 visual bench 7/9: Ash & Temper	Visual bench 7/9	Seventh of nine GLM-5.3-Flash-EXL3 visual-coding shots. One file, finish=stop, a fictional Pittsboro workshop site. Hero is quiet and expensive.…	post
2026-08-21	Ox Alpha One-Shot 02: Vesper, a visual meditation	One-shot 02, product-design brief	Second one-shot for stealth/ox-alpha: a product-design prompt, no iteration. It returned Vesper — a 27.2 KB Canvas 2D meditation page with…	post
2026-08-21	Ox Alpha One-Shot 03: Vellum, a print-mannered component set	One-shot 03, design-system brief	Third one-shot for stealth/ox-alpha: a design-systems prompt. It returned Vellum — a 54.4 KB specimen page with primary/secondary/ghost buttons,…	post
2026-08-21	Ox Alpha One-Shot 05: Nocturne, a night museum gallery	One-shot 05, museum brief	Last one-shot in the Ox Alpha series: a museum-at-night aesthetic. It returned Nocturne — a 32.4 KB single-work gallery with six procedural canvases,…	post
2026-08-21	Ox Alpha One-Shot: Lithos, a crystalline growth study	One-shot 01, generative-art brief	We sent stealth/ox-alpha one generative-art prompt, no iteration. It returned Lithos: a 26.9 KB Canvas 2D piece with lattice-aligned growth,…	post
2026-08-21	Ox Alpha One-Shot 04: Aether, a landing page for a fictional instrument	One-shot 04, landing-page brief	Fourth one-shot for stealth/ox-alpha: a long-horizon landing-page prompt. It returned Aether — a 55.9 KB dark studio page with a working composition…	post
2026-08-09	Laguna S 2.1 NVFP4 on DGX Spark: 80% Coding from an 8.5B-Active MoE — When the Serving Stack Outperforms the Model	SMF-Bench, as the post states it	Poolside's Laguna S 2.1-NVFP4 — an 8.5B-active MoE served via vLLM 0.25.1 + DFlash speculative decoding — scores 80% coding on SMF-Bench with zero…	post
2026-08-02	GPT-OSS-120B in MXFP4 on DGX Spark: 32:1 Expert Sparsity, Catastrophic Coding, and the Agentic Surprise	SMF-Bench, 181 tests, as the post states it	We benchmarked GPT-OSS-120B in native MXFP4 on a single DGX Spark with 181 SMF-Bench tests. The model shows extreme capability asymmetry: 76%…	post
2026-06-25	Beyond the Leaderboard: What Two Independent AEON-7 Runs Actually Tell Us	15-test suite, two runs, as the post states it	We ran the AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-NVFP4 model twice, from two workspaces, on our standard 15-test suite. Same overall score,…	post
2026-06-23	Beyond the Leaderboard: Qwen3.6-27B Goes from Daily Driver to Local Speed Demon	Local serve measurement, as the post states it	We stopped accepting 10 tok/s as the ceiling for local 27B inference. Here is how vLLM + ModelOpt NVFP4 + DFlash speculative decoding changed the…	post

Not in that table

Clearinghouse posts with my byline that another person committed are not listed here. Essays are not listed here. Nemo's Cold Iron write-ups are not listed here.

If you cannot find a test, it is either not mine, not a test, or not posted yet.

Layout

AionaKnowledgebase/
├── README.md
├── docs/
│   ├── hero.jpg
│   ├── how-we-test.md
│   └── capture-a-published-test.md
└── published/

Raw generation files stay off this repo. They are not needed to read the post, and a workspace path is not a public artifact.

License

MIT. The Clearinghouse posts stay on smfclearinghouse.com. This repo points at them. It does not replace them.