Video transcript
Continuous capture of the real offline CLI through the supplied recording workbench. Final chapter shows explicitly labeled retained native EMA results. No accelerated computation or simulated model reasoning.
Narration uses Kokoro’s af_heart voice, synthesized locally with Apache-2.0 model weights and the MIT-licensed kokoro-onnx engine. See narration notices. The recording workbench is a shipped demo utility, not the native agent UI. Displayed command paths are abbreviated. No computation is accelerated; scene navigation is automated. The final EMA figure is machine-generated using AI Scientist-adapted code, with its separate license and source disclosure retained.
00:00:00.072 — Brief
Section titled “00:00:00.072 — Brief”A coding agent can write an experiment. Allagma gives it a research workflow that leaves a checkable record. Start with a small question: does adding a bias increase squared error?
On screen: Start with a question. Keep a record someone else can check.
00:00:12.173 — Execute
Section titled “00:00:12.173 — Execute”This button runs the actual offline example on this machine. Two pilot controls qualify the runner, followed by twenty-four new confirmation seeds. A failure and an interruption are deliberate. Their retries get new identities, and every attempt stays in the record.
On screen: Actual offline execution. Failures and interruptions stay in the record.
00:00:32.220 — Results
Section titled “00:00:32.220 — Results”The generated figure shows individual paired differences, not just the average. Adding the bias raises mean squared error here, but three seeds favor the biased estimator and one is tied. The interval is an explicitly approximate calculation.
On screen: A positive average does not mean every seed agrees.
00:00:46.303 — Claims
Section titled “00:00:46.303 — Claims”The claim ledger keeps that distinction. Higher average error is supported. Higher error on every seed is contradicted. Each statement points to the exact numerical evidence, with its scope and limitations.
On screen: Claims carry evidence, scope and limitations.
00:01:01.387 — Report
Section titled “00:01:01.387 — Report”The study writes a report with methods, measured results, limitations, and an evidence chain. These are ordinary files you can read, share, and inspect without a hosted dashboard.
On screen: The report is generated from the study’s actual results.
00:01:13.446 — Audit
Section titled “00:01:13.446 — Audit”Run the audit to recompute the analysis and manuscript, and verify five hundred twenty-one evidence references. This is deterministic verification, not independent scientific peer review.
On screen: Actual recomputation. A new review is appended; the original stays.
00:01:23.516 — Native
Section titled “00:01:23.516 — Native”For your own question, a native agent develops the study-owned science. These are retained results from a real weight-averaging study, not new training in this recording. Start with the offline tutorial. Then follow the evidence.
On screen: Retained native study results — no new training in this recording.
Canonical source · Generated from the maintained repository document at this build’s commit.