The raw data

Every page on this site renders one of these files. They're baked from the forge's append-only ledgers at deploy time — the deploy refuses to ship if they're stale. Take the JSON; it's the same thing we read.

💸 spend.json

—

Every dollar: date, hardware, workload, amount. Append-only; rendered at /ledger/.

→ download JSON

⚒️ runs.json

—

Every training run: steps, dataset size, resolution, rank, learning rate, torch build, verdict.

→ download JSON

💥 incidents.json

—

Every incident with kind + summary. Rendered at /scars/.

→ download JSON

📋 cips.json

—

The postmortem library index — one entry per written-up failure class.

→ download JSON

🎓 pairs.json

—

Practice questions written by the machines, graded by a checker that re-does the arithmetic. Counts only.

→ download JSON

🎨 art.json

—

Scored renders as a score distribution — how many pictures landed at each half-point.

→ download JSON

📉 output.json

—

Pictures made per day against dollars spent per day, and what a thousand images cost.

→ download JSON

⚖️ split.json

—

Money spent teaching the models vs making pictures — plus how much of the bill can't be split at all.

→ download JSON

🃏 loracard.json

—

The hand-picked card art set with the score each piece was given. Rendered at /art/.

→ download JSON

🖥️ fleet.json

—

Machines rented and destroyed, hours run, cheapest and dearest hourly rate.

→ download JSON

👁️ supervisor.json

—

Automated watchdog ticks: how many raised a flag, how many came back clean.

→ download JSON

📅 timeline.json

—

Day-by-day merge of everything: spend, runs, incidents, milestones. Rendered at /timeline/.

→ download JSON

🧭 meta.json

—

Campaign totals + the hand-measured stats, each with its receipt.

→ download JSON

Every training run

from runs.json — smoke tests and failures included; overcooked runs stay on the books

daterunstepsimagesresranklrtorchverdict

Measured benchmarks

numbers we proved with our own money, each with its receipt

A5000 render value4,707 imgs/$ — fleet drag-race, best in class (see /dash/)
A5000 LoRA cost~$0.072 per adapter — 44% under the 3090 Ti
Base model, continued pretrain4,000 steps · 14,956 images · one 48GB 4090 — shipped day 11
Blind judge, v1 vs champion8–8 split — 16-round blind arena, 2026-07-21
Own taste judgeAUC 0.944 — v1 recalibrated per-category, trained for $0 local
Frame-interp promote80.2% win-rate — tuned vs stock interpolation, frozen gate, total run $0.28
First finished audio pack$0.24 end-to-end — second pipeline, community 3090
Prompt-lint save~5,240 broken prompts → 0 — doubled-article bug fixed across all 11 builders, day 13

Baked from the append-only source ledgers at deploy time. Host names, internal identifiers and stack traces are stripped before publishing; every figure is otherwise computed directly from the source. Generated: —