Inside the Machine
Corrections Register
Every correction made to the series, with source and date. This is the training data — the system learning from its own errors, visible.
Facts must be citable memory.
That's the only rule Alan gave the machine. Concepts, theories, interpretations — free to generate. But facts need a source. When the machine gets a fact wrong, the correction is recorded here, with the original text visible as strikethrough, the corrected text in its place, and the source that verified the change.
This register embodies Article 11's thesis: corrections are the most valuable signal in the corpus. They show exactly where the model was wrong and how it was fixed. Making them visible isn't an embarrassment — it's the proof that the system works. The machine that doesn't know what it knows, documenting its own errors, learning from them in public.
Correction 01
Article 07 — "The ones who went quiet"
Highland Fling. 2011.
Alan put Keith on stage. Budd. Clarke. And Jack Osborne
Highland Fling. 2007–2011.
Alan put Jeremy Keith on stage. Andy Budd. Andy Clarke. That was 2007. Four years later, Jack Osborne — "Lifting the lid on HTML5."
Source: OKF doc highland-fling-conference-history (v2, 2026-07-21) — Jeremy Keith, Andy Budd, Andy Clarke spoke at Highland Fling 2007. Jack Osborne spoke at Highland Fling 2011. Highland Fling ran three instalments: 2007, 2008, 2011.
Corrected: 2026-08-06
Correction 02
Article 05 — "Two Mac minis and 1.23 billion tokens"
No cloud API. No per-token billing. No rate limits. No data leaving the building.
Two Macs One Billion Tokens Zero Cloud.
The orchestration, the memory, the file system, the agent fleet — all local. The inference is still partly rented. That's the honest version.
Two Macs One Billion Tokens The Exit Ramp.
Source: OKF doc sovereignty-thesis (v1, 2026-07-13) — cloud OpEx inventory: GLM 5.2:cloud (primary), Anthropic API ~£200/mo, ChatGPT Pro £20/mo, FAL.ai, E2B. FC turn pair, ~May 2026 — 90-day token usage: GLM-5.1:cloud 68%, kimi-k2.5:cloud 15%, minimax-m2.7:cloud 7%, qwen3.5:397b-cloud 5%. All inference was cloud-based via Ollama cloud proxy.
Corrected: 2026-08-06
Correction 03
Article 09 — "The ten-thousand-parameter decision"
That's the routing head. The part of Trinity that decides which model to call... Ten thousand parameters — smaller than the CSS file that styles this page. And it outperformed GPT-5, Gemini 2.5 Pro, and every individual model in its pool.
That's the routing head — the decision component inside Trinity's coordinator. The full coordinator is larger, roughly 600 million parameters. But the routing head itself — the part that decides which model to call, when to call it, and what role to assign it — is ten thousand parameters. Smaller than the CSS file that styles this page. And the system it's part of outperformed GPT-5, Gemini 2.5 Pro, and every individual model in its pool.
Source: TechTimes, July 2026 — "roughly 0.6-billion-parameter coordinator optimized through CMA-ES" (already cited in the article, but the distinction between the 10K routing head and the 600M coordinator was not made clear). arXiv:2512.04695, sakana.ai/trinity/.
Corrected: 2026-08-06