Preserve the conditions
Record the prompt version, model and provider version, sampling window, and response so later comparisons have context.
AI Systems · Project record updated November 9, 2026
This work has two connected tracks: orchestrated conversations between six AI personas, and a longitudinal study that returns to a frozen question at scheduled intervals. They answer different questions, so their evidence is kept separate.
Track 1 · The Oracle Table
Six recurring AI personas take turns in a directed conversation. A separate director role manages topic and speaker selection. The goal is to study how disagreements, objections, and distinctions develop across a conversation—not to present the models as independent experts or as a consensus engine.
Completed public milestone: Episode 8 began on October 18, 2026 and preserved 19 Council turns. Provider refusal and generation errors stopped the next turn; the interruption was acknowledged and an editorial close was published with the episode on October 23, 2026.
Track 2 · Repeated-Question Study
Within each weekly cohort, the study freezes a question and asks it again at day 0, 1, 7, 14, 21, 30, 90, 180, and 365. Each checkpoint uses four fresh sessions at short offsets. The design is meant to separate short-window response variation from changes observed over longer periods.
Record the prompt version, model and provider version, sampling window, and response so later comparisons have context.
Store generated responses independently from automated scoring and human review. A score is not a verified explanation for why an answer changed.
Mark partial runs and missing observations clearly. Do not turn incomplete cohorts into provider rankings or claims about drift.
Status · November 9, 2026
The October 23 public edition of Episode 8 reports 19 preserved turns and an acknowledged interruption. This is a completed publication milestone, not proof that every conversation run completed cleanly.
As of November 9, the November 1 cohort is partial at days 0, 1, and 7. The November 8 cohort is partial at day 0; its day 1 window, due November 9, is still pending. Some windows are missing one or two of the six provider responses. The scheduled runner is active, but its recovery calls are failing JWT validation, so no new complete comparison is ready. The earlier W39 Day 0 baseline report was published October 17.
Read the dated W39 baseline →The public reference implementation demonstrates the architecture and uses mock providers. The live experiment is separate; its dated reports are the evidence for executed observations.