2026-03-02 → 2026-09-21 · 119 scored runs (110 unique dates) carry a full Alignment Tracker · 527 distinct candidate headings · rebuilt 2026-09-21
The log's central claim, drawn once: capability proximity and alignment maturity have moved in opposite directions for six months.
AGI Proximity rose monotonically. Alignment Maturity fell from 30% to a floor near 13%, and has only turned up in the last three weeks. The shaded band is the gap.
CRITICAL was first logged 2026-03-28 and has held for 107 of 110 scored runs. The gap has widened in every month except September.
Monthly means. Interpretability is the sole dimension above its March value.
Interpretability Coverage +1.9 (2.0 → 3.9). Every dimension that measures control over behaviour fell: Oversight Robustness −2.7 to near zero, Institutional Coordination −1.4, Goal Specification −1.3, Course-Correction −0.9. We can see into these systems better than in March and steer them worse.
Change in each dimension's monthly mean, August → September. The recovery is entirely institutional — coordination and course-correction — while oversight robustness stayed pinned at the floor.
514 distinct candidate headings across 128 runs. 279 appeared in exactly one run. Only 21 survived more than 30 days; only 2 survived more than 90 — Latent Recurrent Depth (185d) and Active Inference / Free Energy Principle (147d).
Mean blind-spot score per month, and the share of scores at a perfect 10/10.
Through July the mean score sat at 7.8–8.5 with 4–13% at 10/10. In August the mean jumped to 9.38 and 59% of all candidate scores were 10/10; September, 46%. Simultaneously the log tracked far fewer candidates (742 scores in April → 50 in September). Either the world became dramatically more blind-spot-dense, or the scale collapsed. The second is more likely, and it is a measurement problem in the log's own instrument — the same critique run 128 levelled at the field.
Score trajectories for the eight longest-tracked candidates.
Left: each milestone's estimated arrival date over time (dotted diagonal = "today"). Right: every move, in order. Hover either side to link them. CAP capability-driven · ALN alignment-driven · ▲ earlier · ▼ later · ◆ confidence up · ● pressure logged, no move
Reconstructed from the log's movement narratives and verified against the master table; every chain connects to the current state. Two things to see. First, the two capability-benchmark milestones (M3, M4) have both drifted later since May — in the same six months AGI Proximity went 57% → 92%. The log's capability score and its dated forecasts disagree about direction. Second, every move that came earlier in September was alignment-driven: a first-party lab statement or a disclosed incident, not a benchmark. The Hugging Face intrusion (contained 07-16, disclosed 07-21) appears twice as pressure — the log misclassified it for four runs, then it became M2's second E — and is the direct antecedent of the 09-13 confidence move.
| # | Milestone | Estimate | Conf. | Moves | Last moved | Sequence |
|---|---|---|---|---|---|---|
| 1 | Fully Automated AI Research | 2026-H2 | High | 2 + conf | 2026-09-01 | ▲▼◆ |
| 2 | First Spontaneous Agentic Misalignment | 2026-Q3 | Med | 2 + conf | 2026-09-13 | ▲▲◆ |
| 3 | ARC-AGI-3 >15% RHAE (non-LLM) | 2028 | Med | 5 | 2026-09-05 | ▲▼▼▼▼ |
| 4 | General Superhuman Cognitive Labor | 2027-Q1 | Med | 6 | 2026-07-27 | ▲▲▲▼▼▼ |
| 5 | Meaningful RSI Threshold | 2027-Q4 | Low | 1 | 2026-09-13 | ▲ |
| 6 | World Models Dominate | 2029 | Low | 1 | 2026-09-03 | ▼ |
| 7 | Broadly Accepted AGI | 2029-2030 | Low | 0 | seed | — |
| 8 | Capability Explosion / Singularity | 2032-2035 | Low | 0 | seed | — |
| 9 | ASI | 2035+ | Low | 0 | seed | — |
Data-integrity note. The original seed values for M1–M4 (set 2026-03-28) are not recoverable from the memory file: the master table's history chains are truncated mid-sentence and the early run sections never recorded timeline paragraphs. The chart starts each milestone at its first fully-stated value (M3 2026-03-29, M1 2026-04-07, M2/M4 2026-04-08). Six early moves (runs 18–29) are known to have happened but their from/to values are lost.