The scoreboard
The same 200 hand-written analyst questions, the same dossier path, the same judge. The set and every miss are on the 200.
updated by hand as results land · nothing here is a forecast
Why this is the moonshot
Eleven training runs since 23 October taught us one thing above the rest: the score on the 200 never moved with another family of training data, at either size. It moved with the size of the model that reads the graph: 170, 185, 195. The remaining misses are reasoning, dates and long dossiers, and more of the same pairs cannot fix them. A frontier open model that reads at 195 untrained changes what training is for. It is no longer the first year of work, it is the last five percent: our dialect, our roles, our conversation.
Tonight's untrained 20b at 185 is the first proof. It equals the 14B we spent a week training, on a smaller card, faster.
Four tiers, one card
The card running day and night with no cap. Continuous agents, private-model pilots inside a customer's wall, a channel bot that never touches a third party, the big model teaching the small ones, synthetic data at scale.
Claude. Judgement and voice: the Daily's final pass, adversarial review, synthesis, the analyst's take, anything under the byline, code changes. The big boss.
gpt-oss on our card, as NIKO full time. Every grounded, structured, high-volume job across the estate: reading the graph, drafting from facts, proposing read-only queries, triage, detection rules, the agents' structured output. Zero cost per token.
Deterministic code and the small models. The 34-million-parameter Cortex for surprise scores and ranking, the 3B on CPU as the fallback reader. Free, always on.
What moves to the card, and what never does
Moves
- reading
- the Cortex reader, the graph's briefs and reports, explain-a-change, ask-the-graph
- drafting
- first drafts of the Daily and Weekly, SITREPs, briefings, comms templates, architecture passes
- queries
- plain-English to read-only query proposals, behind the same double guard
- extraction
- entities and relationships from feed text, reviewed by a human before the graph
- agents
- triage, rules, the scheduled ops agents with tools off and schema-checked output
- bots
- the Telegram, Slack and Teams agent, answering without a third party
Never moves
- voice
- the Daily's voice pass and verify; the ScottG voice
- judgement
- the Scientist's adversarial review, Deep Read synthesis, the analyst's take
- byline
- anything published under a name
- code
- code changes, and every big-red-button decision: live products, auth, infrastructure, deletes
- numbers
- no model, large or small, is ever the source of a number. Numbers come from the graph.
Two front doors, one engine
Ask
Public, free, the funnel. A plain answer box over the whole graph: conversational, allowed to paraphrase, sources folded under the answer, rate-limited, signal only. Built now, public in 2027.
Investigate
The console. Forensic: quotes, never paraphrases, every query shown, receipts, second look, hunts, briefings, drill-down into every name. Live for testers.
The order of work
- Run 4's verdict. The 14B plus the multi-turn family. Its gate is the dialogue set, because that is what multi-turn is for.
- gpt-oss-20b untrained on the 200. Done: 185. The 20b is the student unless the 120B spike changes the picture.
- The training spike. One 96 GB card: the 20b in 16-bit, the 120B in 4-bit. Seconds per step, memory, and whether an adapter serves over the quantised base without a lossy merge.
- Run 5: train gpt-oss on the clean set plus the fixed roles. Gate: not below the untrained score, dialogues at or above 85 of 90, roles at or above their baselines, invented at or below 0.6 percent.
- Local-first routing, one app at a time, each job class with its own small hand set.
- The card comes home. Rented by the week first; owned when the numbers and the price agree.
If we started again
- Start with the frontier open model reading the graph. Fine-tuning is the last five percent, not the first year.
- Design the graph for the reader. Dossiers as the primary artefact, evidence on edges, the nightly export is the index.
- Tools as the spine. Parameterised, read-only, with provenance. Never natural language to query language.
- Agentic by default. Every investigation a streamed tool trace, every hunt a loop on our own card.
- One product. Ask bar, explorer, dossier. Everything else is a tool the model calls.
- Two readers only. The big one for quality, a distilled student for cost, same sets, same gate.
- Our own writing as training data from day one. The model drafts the paper, Claude edits.
- Evaluation first. The hand sets and the gate before the model, a public eval page from week one, an error bar on every bar.
- The card early. Rented a month, then owned, inside the wall.
- Nightly distillation from the big model into the student, so the model always knows yesterday.
- The moonshot. The big model as a swarm of analyst personas on one card, their work streaming in public, every claim through the prediction registry.
- What does not change. The graph, the receipts, cast iron or it does not ship.
The story line
The deck
Twelve slides in plain language for anyone: the library, the librarian, the recipe, the apprentice, the newsroom and the receipt, with the scorecard. Free to share.