Log

The project began on 19 September 2026. Design records are append-only: each revision is written above the previous one and says what it replaces.

Timeline

Fig. 4The project as a worldline. It diverged on 24 September, when the goal moved from predicting a person's survey answers to a replica that carries on as them. One tick per day.

Entries

Date	What happened
19–20 Sep	The interview harness: subscription model clients, frozen configurations, and an evaluator account that alone can open the held-out answers.
22–23 Sep	An extraction assay, an elicitation design and a null model. Remote sessions that run without any model.
24 Sep	The goal moves: from predicting a person's answers to a replica that carries on as them.
25–26 Sep	The General Social Survey pilot: three nights of local inference, 36,810 forecasts.
27–29 Sep	Personal-updating studies and a reading programme on consciousness and connectomics, then a pause to consolidate.
4 Oct	The drift battery is scored, and architecture v0 is written down.
5 Oct	The base specification and the research on change and voice. M1 is designed.
6–7 Oct	The first engine trains on 400M tokens, then fails its first real test. The corpus is rebuilt so that reading is necessary.
9–11 Oct	Revisions 1.3.1 to 1.3.9 of the acceptance gate, each reviewed independently. Pretraining moves to rented GPUs.

M1 revisions

Revision	Date	What changed
1.0–1.1	5 Oct	The original design and the decisions after the first build review.
1.2–1.2.8	6–7 Oct	Argument and voice families, engine decisions, smoke drivers; 1.2.8 diagnoses the failed first test.
1.3	7 Oct	A corpus in which reading is necessary.
1.3.1–1.3.2	9 Oct	Person conditioning, and the gate on the full recipe.
1.3.3–1.3.5	10 Oct	Decisions after the build reviews; the implementation's open failures.
1.3.6–1.3.9	10–11 Oct	Fixes for every way the reviewers found a check could pass without its capability.