Day 85 — the world lives on the disc

“GTA San Andreas on the PS2… I think this is the key to our problems.”

Three things happened today. A benchmark got won. A factory ran. And a video game from 2004 handed us the next two years of architecture.

We beat a 14B (and told the truth about how)

The dry run turned real. The owned coding path — a deterministic composer over card libraries with a node execution sandbox, no LLM in the loop — went up against qwen2.5:14b on a 478-probe gauntlet. Gold computed in the harness, both sides’ code actually run, no model’s opinion in the verdict.

The first pass looked like a sweep. It wasn’t, and we said so before anyone had to ask: the grader was robbing the 14B on whitespace and valid-alternative answers. Fixed against our own interest, the honest number came back a tie — and then a hole surfaced (the composer abstained on a whole family the 14B could do). So we closed it with real capability — prompt-parameter extraction, “count the letter a” pulling the a out of the ask — and re-ran.

478/478. The 14B: 471, its only misses stochastic — code that crashed on an edge case, different ones each run. werbos never once shipped a wrong answer. 178× faster, on a CPU. The honest headline isn’t “we out-reason a 14B” — we don’t, and the phrasing ceiling is real. It’s this: on a large, fair, execution-graded bench, the owned composer holds the line and is never confidently wrong. At scale nobody can wave off.

The factory ran

Two libraries minted on the work laptop, pulled over, cold-checked in ~250-card chunks, and shipped: home (4,555 graded cards, 560 topics — cleaning, plumbing, gardening, life hacks) and health-fitness (5,472, still grading as the day closed). ~10,000 verified cards in a day, each one that survives the gate earning the right to be cited. The recipe box grew too — 50 dishes, web clips and hand-written Obsidian notes, served whole. This is the fuel line, and it’s flowing.

GTA San Andreas gave us the engine

Then the real thing. San Andreas rendered three cities on a PS2’s 32MB of RAM by never holding the world in memory — it lived on the disc, a cursor streamed in only the slice you were looking at, and the draw-distance fog hid the edge. That is werbos. The library is the world on the disc. Retrieval is the cursor. Abstain is the fog. The frontier buys more RAM; we stream.

It sharpened into an architecture with a memory manager: a small base model that understands the ask, mini specialty models distilled from each mature deck that speak the domain (pre-baked, not live — coding at ~50k cards is the ripe one), and the card that holds the truth. Riff for feel, drop to a single cited card for proof. The guardrail that keeps it honest: the mini renders, the card is truth, and every claim it makes must trace to a card or fog out. It can be loose and casual — it just can’t invent. Under-claiming is free; over-claiming is the sin.

And the line the whole thing hangs on, the one that separates this from every RAG system alive: ChatGPT cites ownership — “a source said this,” a finger to point if it’s wrong. werbos cites correctness — “this passed the gate,” the grading pipeline vouching. One deflects responsibility; the other accepts it. It’s the stronger stance, and it’s only possible because we own a graded library. ChatGPT would have to grade the entire internet to copy it.

The whole design is on paper now: docs/streaming-engine.md. Coding is the pilot; the first test costs almost nothing — prove the paraphrase-plus-per-claim-citation loop over the decks werbos already owns, before distilling a single new weight.

And we ran it — first light, no training. Over the thin Cabo+fishing decks the loop buried the roosterfish cards under tourism trivia; over the mature coding sector (41,611 cards) it put the exact right cards at rank #1. So retrieval is a density problem, not a wall — the pilot deck is ready. But the test caught the real linchpin red-handed: the naive gate scored “a 47-foot titanium rod forged on mars” as GROUNDED. Token-overlap isn’t entailment, and a gate that pills nonsense would hand a mini the one thing werbos exists to refuse. So the next build isn’t the GPU — it’s a real gate. The disc is readable; now we earn the right to vouch for what’s on it.

Also shipped

The chat now defaults to our model, not the open-weight one — lead with the owned stack. The /graph page told the truth again: real counts, the live 2M model named, and Security carved into its own realm to match how the router actually thinks. And the deploy caught its own blind spot — the data repo can flash green while sitting stale, so the cards got pulled and verified live by hand.

A benchmark won honestly. A factory humming. An architecture that turns a game console’s oldest trick into the thing nobody else has: fluency without hallucination, at laptop scale, owned. The mainframe’s still unchecked — but today we found the disc it’s printed on.