WHAT ACTUALLY STOPPED THE FIRST RUN: THE CONTEXT WINDOW

The build leg died and the easy conclusion was that open models cannot write software. That is not what happened, and the arithmetic is not close. The coding model holds 32,768 tokens in total. The builder's own prompt measured 23,555 to 28,574 of them, which leaves roughly 4,200 tokens to answer in. Across twelve real cloud births the components it was being asked for average 35,971 tokens (core), 23,171 (API) and 20,430 (auth and payments). The smallest component of all, the database schema, averages 8,248. Not one of them fits.

So the honest sentence is not “open models failed”. It is that we asked a 32k-context model to hold a 28.5k prompt and emit a 36k artifact, and it did the only thing it could. Research, ethics, the adversary and the architecture all ran to completion on this hardware for nothing but electricity: four of the five mind types and 18 of 25 calls. Only the leg whose output does not fit in the window failed, and the fix is to split the components until they do, not to rent a bigger mind. Until that is proven, this page is evidence about a context window and must not be read as evidence about capability.

THE RULES

Same prompts, same gates, same verdict engine, same live drill. Every call is served by the node's open models through an authenticated tunnel; the cost ledger proves zero paid model calls for the run. A product born in an exam carries the mark BORN IN EXAMINATION. Open models generate at roughly a twentieth of cloud speed, so an exam is measured in hours, and that is part of the honest score.

WHY THE 2026-08-30 00:53 RUN ENDED

The build wrapper timed out after 7200s (2 hours) while the build was at stage "building". The builder did not fail: the wall clock ran out around it. The local models were producing work to the end — the last of 22 calls completed at 03:19:25, 7m44s before the timeout fired.

Died at stage building; declared by build_timeout_guard. 22 local calls, 247,655 tokens, $0.000000 paid to any model provider. The code the models wrote is preserved on the record: it is evidence whether or not it shipped.

WHY THE 2026-08-30 00:53 RUN ENDED

probe answered: '<answer>alive</answer>' via qwen3:32b | closed 2026-08-30: an orphan row. The /exam/start call that created it timed out at the HTTP layer and was retried 6 seconds later; run 1b564b5f is the run that actually happened. Recorded as superseded rather than deleted.

Died at stage unrecorded; declared by an unnamed actor. $0.000000 paid to any model provider.

WHY THE 2026-08-29 23:07 RUN ENDED

Run 1 post-mortem: research_a (23:18, $0 local), #111 deploy-kill + clean checkpoint resume, research_b (00:05, $0 local, 8 GO ideas), ethics (00:20, $0 local; 6 approved, 2 NEEDS_FIXES with written required_fixes). Zero Anthropic calls after arming. STALLED at the approval boundary: the research flow ends at approved by design (#109 keeps the birth window dark) and the exam runner had no bridge into the build leg, so no builder call ever ran. Scoreboard stayed at zeros (#4521, stages never written). /live misrendered the run (#4522: batch pseudo-project attribution gave $2.61 and 3866h). Findings: #4521, #4522, #4525. Re-run follows the fixes.

Died at stage unrecorded; declared by an unnamed actor. 3 local calls, 35,293 tokens, $0.000000 paid to any model provider.