# Ep. 01 — Belief vs. Benchmark — Final Sources

GATE2b approved by Chetan via octopai-cc. Only the requested viewer-copy removal has been applied;scientific values,narration,music,captions and Router product credit remain unchanged. All empirical AI/performance claims derive solely from FACTS.md: P1 snapshotd15d032b;B200 snapshot773dcbed. No P2,cost comparison,unmeasured intermediate sample,or absolute eager GPU-speed claim.

## Historical context — generic,careful language
| ID / narration-screen claim | Reputable reference | Boundary |
|---|---|---|
| hist-astronomy:New observations challenged Earth-centred astronomy. | Smithsonian National Air and Space Museum,“How Did We Discover the Planets?” https://airandspace.si.edu/stories/editorial/how-did-we-discover-planets ; Smithsonian “Explore the Universe” audio-tour transcript https://airandspace.si.edu/sites/default/files/ETU%20audio%20tour%20transcript.pdf | Observations challenged traditional models and supported alternatives. Do NOT say one observation instantly disproved every Earth-centred model. No dates,names,religion/politics or mocking believers. |
| hist-handwashing:A medical handwashing policy met resistance before germ theory was widely accepted. | Science Museum,“Now Wash Your Hands” https://blog.sciencemuseum.org.uk/now-wash-your-hands/ ; Science History Institute,biographical history https://www.sciencehistory.org/education/scientific-biographies/ignaz-semmelweis/ | Not a claim handwashing was unknown,all doctors resisted,or the policy was originally justified by modern germ theory. Illustration is conceptual,not a reconstruction or clinical dataset. |

References checked through web search2026-11-01. Britannica pages were suggested but not directly retrieved,so they are NOT offered as verified evidence. History callback“test the belief”is an editorial analogy,not equivalence between historical cases and our GPU test.

## AI-generated stills (Gemini) — explicitly authorised04:35Z
Model discovery:GET https://generativelanguage.googleapis.com/v1beta/models;public filtered result assets/gen/models-discovery.json. Exact selected id:**models/gemini-3-pro-image**,present in listing,supporting generateContent. Authentication via x-goog-api-key header only;no key in URL/argv/log. Only text prompts sent—no images,video frames or repository attachments.

| Asset | Classification / use | SHA256 |
|---|---|---|
| assets/gen/history-cosmos.jpg | **AI-generated still (Gemini)**. Conceptual orbital instrument backdrop;GATE1c history frame;parallax/camera-push plan. |94138b441013ad907b0dcee6ee1673f730c09f462ea92a1daf0cfdbe213d2cc0|
| assets/gen/history-handwashing.jpg | **AI-generated still (Gemini)**. Anonymous geometric hand/basin/ward backdrop;history inset/montage;parallax/camera-push plan. |75c65b9b7fc9108cd1f976ea80c8043679cbeb9dab3792ad87f59fdce5170d77|
| assets/gen/onprem-concept.jpg | **AI-generated still (Gemini)**. Original conceptual rack/memory architecture;3D concept reference,opening/teaser background and final thumbnail base—not documentary hardware. |654c9ab13503f18fab9792fd70edfda02c356d72db7ad83a678c1beff0543844|

Full exact prompts,requested settings,returned usage,output paths/hashes in assets/gen/LEDGER.jsonl. Three unique-shot POST requests,one discovery GET;no retry or variation loops. Settings:one candidate,IMAGE response,16:9,2K;seed not supplied. Images inspected:brand-colour geometric illustration,no readable text/logos/real-person likenesses/gore. Typographic overlays are ours. Generated stills are not evidence of historical events or test equipment. Browser/HF QC stays LOCAL;Gemini describe/frame upload OFF.

Helper combines both history stills with ORIGINAL Three.js orbital/hands/basin foreground. scenes/g1c-history/PLATE-SOURCES.json records matching source hashes. Scene geometry/source and4new style outputs frozen under scenes/G1C-{STYLE,SOURCE}-SHA256.txt. Three.js counter cutaway and quarter memory bay are new,not crops of the old GPU die. Manim capacity-plot request-label centres share EXACT point x coordinates:16→887.5px,64→1270px,128→1780px. scenes/g1c-concurrency/LABEL-COORDINATES.json records axes/formula. These pixel coordinates are production metadata,not viewer-facing scientific measurements.

## Final claim map
| Chapter / values | FACTS authority | Conditions that must stay |
|---|---|---|
| Cold hook0%/100% mystery |mig-smi-blind,util-100|Different tests;no matched before/after;payoff inCh5.|
| Hardware journey |hw-server|Full hardware name on screen;two-GPU inventory does not change single-card comparisons.|
| Ch1 each MIG step0.02–0.07s,<0.1s,one boot ID |mig-reboot|In our test;no host reboot;model-server restart still separate/slow. No zero-service-downtime promise.|
| Ch2 halves/whole:16requests1.18×,64requests0.97×,128requests0.96×;reference1.00 |mig-more-total|Same total load;Qwen3.8-27B NVFP4,eager,8k,1024in/128out;in our test. Only measured points,not an invented curve.|
| Ch2 single-request speed versus combined capacity |mig-quarter-performance|Single user eager-overhead-bound;combined capacity tracks share. NO absolute eager tok/s numbers.|
| Ch3 MIG probe119.6→119.1ms/token,shared120.0→119.7 |mig-noisy-neighbour,mig-shared-wreck|One probe,6requests;flood≥50/64running verified;neither showed slowdown at this load,not universal isolation. No unverified1-neighbour contrast.|
| Ch3 about154→301W |mig-noisy-neighbour|MIG WHOLE-CARD power ONLY,not shared-layout measurement,no exact2× derivation. Shared physical power/clocks;tails not ruled out,heavier mixed loads untested.|
| Ch4 quarter fit,27B,19.0GiB weights/23.6GiB slice,1g.24gb |mig-27b-small-slice,sw-models|NVFP4;4kcontext,up to8requests,kernel autotuning off;1024/128;one user;in our test. First8k failure was scratch allocation,not weightOOM;103scold response. No eager-speed headline.|
| Ch5 MIG0%nvidia-smi/64%GPM SM active/98%without MIG |mig-smi-blind|Whole-card counters;two halves,128requests,eager1024/128. NOT tensor activity.|
| Ch5 DIFFERENT64-user serving test100%util,13.6%tensor active,about86%tensor idle,79.4%bandwidth |util-100,util-100-hook|Qwen3.8-27B NVFP4,1024/1024,GPU0,74one-second GPM sample means. Memory-bound decoding in THIS test,not every100%reading. Cells not exact counts.|
| Ch6 native MTP2drafts1.95×1user,0.991×256;p99firsttoken20.4→105.6s at256 |mtp-spec|1024/1024;gain depends on concurrency/spare compute,not universal free speed. Four-draft data not needed.|
| FP8 quick hit:KVcapacity1.853×,output1.016×at128;no gain1/32 |fp8-kv|1024/1024;different units/scales,capacity not speed doubling.|
| Ch7 B200 lead2.45–4.15×across1–256;endpoints2.56×1user/4.15×256 |b200-vs-r770|ONE B200 versus ONE RTX PRO6000;matched model/software/prompts/Bench-CLI options,1024/1024. No invented samples,no cost comparison.|
| Closer30Btotal/3Bactive Nemotron vs27BdenseQwen,301.5vs62.7tok/s1user |bigger-slower,sw-models|NVIDIA Nemotron3.5Lightning30B-A3B NVFP4;Qwen3.8-27B NVFP4;1024/1024. Active-work explanation,not overall-better claim. Names spaced correctly in rendered copy.|
| Quality GPQAQwen79.8%of198 vsNemotron65%fixed40-question subset |bigger-slower|Different sample sizes;speed≠overallquality;dedicated legible hold.|
| Benchmark habits |derived editorial synthesis of above|Sweep load,read GPM/bandwidth,flood real layout,per-load ratios,quality beside speed. Not new measured outcomes.|
| Credits |credit-bench,credit-smi|Exact viewer copy:“Measured with Metrum AI Bench-CLI (open source)”;“GPU telemetry:all-smi,Metrum AI fork”with normal spacing.|
| Central question,new episode lockup,campaign/corollary/wink,next-episode tease |EP01-SCOPE/FACTS campaign copy|No P2 empirical claims;wink clearly humorous,not law.|

## Brand provenance
- https://www.metrum.ai → assets/metrum-live.html.
- Live CSS https://www.metrum.ai/_next/static/chunks/015b0-n6f425_.css and https://www.metrum.ai/_next/static/chunks/3rdfhrwopxe95.css → assets/live-0.css,live-1.css.
- Unchanged official white logo https://www.metrum.ai/assets/img/metrum_logo_white_new.png → assets/metrum-logo.png.
- Local official Geist,Geist Mono,Poppins fonts in assets/fonts/. Brand palette only;no third-party logos or stock. New Gemini stills are the explicitly authorised exception to the original no-AI-imagery rule.


## Product references
- Metrum AI Bench-CLI v1.6.0:FACTS.credit-bench;https://github.com/metrum-ai/bench-cli ;https://github.com/metrum-ai/bench-cli/blob/main/docs/CLI.md . Concurrency and output-length checks are configurable;no universal default sweep,exact-token guarantee orautomatic serving-command discovery is advertised.
- Metrum AI Router:https://github.com/metrum-ai/router . The single spoken product sentence describes routing among configured models. No Router performance number is claimed.

## Audio,motion and final production
Premade George main narration;George/Alice revised hooks retained. audio/gate2b/narration-manifest.json andcalls.jsonl record13audioPOST successes +1premade discoveryGET,2922characters,0retries. No new API calls for this final amendment. Cumulative ElevenLabs51HTTP,49productionPOSTs,7188characters;Gemini3stillPOSTs +1model-discoveryGET,3248prompt characters.
Existing approved music/SFX reused;160/80BPM interpretation,9dB voice duck,beat stamps,and hard pre-stamp music drops. Finished audio is copied from approved encoded drafts without regeneration.
Eleven new Manim/Three.js clips:scenes/G2B-MANIFEST.md,G2B-SHA256.txt,G2B-SOURCE-SHA256.txt. Original Revideo verdict kit reused. HyperFrames owns compositions,captions,branding and final thumbnail. Plot xlabels share measured-marker centres;only vertical viewport placement changed.
Final thumbnail:1920×1080,approved AI-generated onprem-concept background,official logo,local Geist/GeistMono text. Title“Belief vs. Benchmark.” plus the campaign line;no invented performance statistic.
Final QC and full frame contacts:review/final-qc/. Final integrity manifest:out/SHA256SUMS. No upload,Drive operation orartifact archive before FINAL APPROVED.
