- Entry date
- 15 April 2026
- Category
- Method
- Lead over the world
- the discipline, not the lead
- Access
- 🔓 Public
A 1-billion-parameter training run crashed out of memory mid-project — and instead of scrubbing the incident from the record, all three response rounds were preserved next to the final fix.
- 26
- council members, seven-gate process, MR-1..20 master requirements
- ~79,196 files / 31GB
- scale, ~279 council sessions recorded (with documented gaps)
- ~4.45
- training loss at step ~1,400, Recipe-C run live at the time
Honest evaluation
Governance held under a real, live OOM crash — all three response rounds preserved on the record, unedited.
What would prove or disprove it further
What would disprove it further: a second incident where the record was quietly cleaned up, or where skipped gates were never revisited. Neither happened here — the one gap that did occur (gates skipped under pressure) is on the record precisely because it was caught and reviewed after the fact, not because it didn't happen.
The evidence — full reasoning behind the verdict
Verdict: proven. The claim under test is whether a heavyweight governance process (26-member council, seven gates, master requirements) holds up under real operational duress, not just on paper during a calm week. It was tested by an actual emergency — a 1-billion-parameter training run crashing out of memory — and all three response rounds (immediate triage, follow-up attempts, final resolution) were preserved on the record, unedited, rather than replaced with a tidied-up summary after the fact.
It landed here because the test wasn't staged. The crash was a real incident during a live Recipe-C training run (loss trending to ~4.45 at step ~1,400), and the response to it is exactly the kind of material most engineering teams clean up before anyone outside sees it. Keeping it as-recorded, plus a later catch-up session that went back and reviewed gates skipped under time pressure rather than pretending they'd all been kept, is the specific evidence for "governance survives duress," not just a claim that it should.
Most crash reports get cleaned up before anyone outside the team sees them. This one didn't — and that choice is most of what makes it worth recording.
What it was
Unification (/unification/) was a meta-project coordinating 22
sub-projects under one constitution: a 26-member council, a formal seven-gate
process (requirements → architecture → design → test → code → integration →
verification), a set of master requirements (MR-1 through MR-20), and a large
patent register audited on a four-column schema (wired-in / prod-flag / benchmark
/ skip-guard). The standing directive was blunt: zero MVP, save every gate
round — not just the final, cleaned-up version.
What we built
ResearchThe flagship sub-project was the Dolphin trainer — a Rust and Python stack covering training, inference, multiplex, autograd, dataloader, and serve crates. Recipe-C training was live at the time, with loss trending down to roughly 4.45 by step ~1,400. Scale across the whole meta-project: roughly 79,196 files, 31GB, and around 279 recorded council sessions (with some documented gaps, honestly flagged as gaps rather than hidden).
What we learned — including the honest negative(s)
- Documentation volume has a real cost. Saving every gate round, every dissent, every incident, produces a genuinely useful audit trail — and a genuinely hard discoverability problem at 31GB with no full-text index.
- The directive wasn't followed perfectly, and that's on the record too. Gates got skipped under time pressure; the honest move was reviewing them after the fact rather than pretending they'd all been kept.
Where it went / status
Active through early May 2026, then folded into the wider synthesis effort. The governance shape it proved out — a council, a gate sequence, and a habit of preserving incidents rather than tidying them — is the direct ancestor of the discipline the later NovaOS synthesis explicitly tried to inherit, while fixing the discoverability problem this project surfaced.
What is still open — kept visible
The honest edges, next to the wins. This is what turns 🔬 into 🟢 — honestly.
- Discoverability is the real cost of 'save every gate round' — with no full-text index, 'which council discussed X?' is genuinely hard to answer across 31GB.
- A later catch-up session had to retroactively review gates that had been skipped under time pressure — the directive wasn't followed perfectly in real time.
- Recipe-C training was live and improving, not a finished, converged model, at the time of the audit.
Proofs & sparks
We demonstrate rather than assert. Each ✅ proof is a visible result with a hard figure.
- Ship the internal dashboard as a public page — Always measured, always current, including the amber — no lab shows its own 3/10 next to its wins.
Where this connects
Sources
- vocabotics project audit — Unification (22 sub-projects, Dolphin trainer, OOM incident), April-May 2026vocabotics internal project history · as of May 2026