# internal_octopai.mp4 — transcript (INTERNAL ONLY)

Runtime 7:19.60 · 1920×1080 master · narration voice Roger (ElevenLabs)

| # | in | out | narration | claim ids |
|---|---|---|---|---|
| 1.1 | 00:00:00,000 | 00:00:04,911 | Before a single slide existed, two agents spent days arguing with silicon. | see deliver/QC.md |
| 1.2 | 00:00:06,300 | 00:00:26,545 | A failed attempt costs exactly as much as a successful one. Forty-eight core-seconds against forty-seven, measured. So why does the whole industry price AI by the token? Because tokens are easy to count, and accepted work is not. We spent a week working out what an agent actually costs. | see deliver/QC.md |
| 1.3 | 00:00:27,467 | 00:00:37,916 | A hundred individual agents, spread across machines, working in parallel. Eleven leads, each owning one workstream, with their own subagents. They coordinate with each other. | see deliver/QC.md |
| 2.1 | 00:00:39,567 | 00:01:14,989 | The work ran as a constellation, not a pipeline. Phosphoros went deep on the NPU inside a Dell Pro laptop. Hestia went deep on the Strix Halo box. Enodia owned the router hierarchy and the containment demo. Hephaestus the Instinct core. Astraea the evals and the final QC. Plutus priced every hop. Lynceus recorded what actually ran. Thaumas built the show experience. Daedalus the brand pack and the three.js wireframe kit. Apollo the films. And Calliope held the narrative, the scripts and the claim register. | see deliver/QC.md |
| 2.2 | 00:01:15,400 | 00:01:24,073 | More than a hundred agents at peak: eleven leads and eighty-nine named subagents, plus their own. An estimate, not a claim about concurrency. | see deliver/QC.md |
| 2.3 | 00:01:25,400 | 00:01:35,666 | Then the swarm assembled the proposal. One story, one claim register, one cost model, a deck and the films — a hundred agents working in parallel on the same problem. | see deliver/QC.md |
| 3.1 | 00:01:39,267 | 00:02:01,393 | The NPU specialist went first, and its first job was not to make the NPU look good — it was to prove the NPU did anything at all. Busy time grew in a straight line with prompt length: 0.377 milliseconds per prompt token, R² 0.991, over twenty-five requests. Every other path read exactly zero. | see deliver/QC.md |
| 3.2 | 00:02:02,233 | 00:02:19,448 | Then the models. A thirty-five-billion-parameter mixture-of-experts ran on the NPU at 13.5 tokens a second and scored 58 out of 60 on a fixed quality suite — above both cloud references. Whisper ran a clip in 2.74 seconds against 4.33 on the CPU. | see deliver/QC.md |
| 3.3 | 00:02:20,600 | 00:02:34,210 | And then the Router itself ran on the laptop: all forty-eight test prompts routed exactly as labelled, the full set in 2.9 minutes, down from 22.1. The edge of the hierarchy, working. | see deliver/QC.md |
| 4.1 | 00:02:35,100 | 00:02:42,623 | That is the finding, not a defect. One device, one request at a time — and it changes the recommended split. | see deliver/QC.md |
| 4.2 | 00:02:45,867 | 00:03:07,496 | Wait, really? The same decision, three configurations, one machine: 0.40 seconds alone, 39 seconds sharing the NPU, 0.45 seconds with the chat on the iGPU. Fifty out of fifty decisions still correct, and the chat only nine percent slower. Preliminary, one device, labelled that way. | see deliver/QC.md |
| 4.3 | 00:03:12,100 | 00:03:23,751 | It changed the recommended split, it changed a slide, and it became a scene in all the films. It also went upstream as an issue — a limitation you can measure is worth more than a claim you cannot. | see deliver/QC.md |
| 5.1 | 00:03:28,367 | 00:03:51,041 | The Halo specialist went second. One box, 128 gigabytes of unified memory — and memory decides what fits: a 122-billion-parameter model fits only because the memory is shared. Sixteen interactive chat streams at 15.8 tokens a second each. Against an NVIDIA GB10 it leads by 1.3 times for one user, and ties at sixteen. | see deliver/QC.md |
| 5.2 | 00:03:51,867 | 00:04:16,918 | Then the surprise: software moved the numbers more than silicon did. One host upgrade sped the mixture-of-experts model up 45 to 64 percent. And one n-gram trick looked like a big win — until the quality gate caught it: hard code fell from 29 of 40 to 6 of 40, with 25 new syntax errors. A speed-only benchmark would have called that an improvement. | see deliver/QC.md |
| 5.3 | 00:04:17,700 | 00:04:25,484 | None of it counts until the telemetry is honest. GPU power here was reporting socket power. Two wrong metrics became none. | see deliver/QC.md |
| 6.1 | 00:04:26,100 | 00:04:40,990 | Then the swarm assembled the media itself. A deterministic wireframe kit — laptops, racks, particle streams, the agent loop, a router gate — rendered frame by frame, and every frame audited for colour and legibility. | see deliver/QC.md |
| 6.2 | 00:04:45,400 | 00:04:56,345 | A hundred agents, in parallel, on one problem: the analysis, the charts, the diagrams, the films, the deck — assembled and cross-checked. Nothing was written twice. | see deliver/QC.md |
| 6.3 | 00:04:58,967 | 00:05:11,218 | Which is why provenance was the hard part. A prior result, a current capture and an illustration must never blur. Every frame carries its label, its source and its timecode. | see deliver/QC.md |
| 6.4 | 00:05:15,100 | 00:05:23,407 | And the discipline: every number on screen carries a claim ID, and every claim that could not be sourced was cut, not softened. | see deliver/QC.md |
| 7.1 | 00:05:32,800 | 00:05:49,753 | The cost question took the longest to settle. The unit was settled by the pilot's own evidence: forty-eight core-seconds for a failed attempt, forty-seven for one that passed. If a failure costs the same as a success, then the only honest denominator is accepted jobs. | see deliver/QC.md |
| 7.2 | 00:05:55,000 | 00:06:23,865 | So they built one model, and everything else reads from it: the receipt for a single job, the three-scenario comparison, the break-even occupancy, the payback, and the four levers — retries, escalation share, cached input, occupancy. Every cost number in the deck and the films comes out of that one script, and it refuses to print if it stops matching the plan page. And one thing never appears in any of it: what it cost us to make this. | see deliver/QC.md |
| 8.1 | 00:06:27,333 | 00:06:49,328 | Lessons, internal only. Test the harness before the model — three of our own failures came from test setup, not the models. Label the scope, or the number will travel further than its evidence. Instrumentation gaps are findings, not embarrassments; the pilot's six invalid sustained studies are more useful than a smooth curve would have been. | see deliver/QC.md |
| 8.2 | 00:06:50,833 | 00:07:05,122 | And here is the part that is still a plan. The next campaign will be built by a swarm that routes its own tokens through the Metrum AI Router, with every hop priced in the ledger. The solution, built on the system it describes. | see deliver/QC.md |
| 9.1 | 00:07:05,800 | 00:07:14,002 | Internal only. What ships is the deck, the films and their sources. What stays here is how they were made — and who said no. | see deliver/QC.md |
