{
  "question_id": "W30",
  "slug": "lab-rolls-back-to-older-model",
  "title": "A lab withdraws its best model over alignment and reverts to an older one",
  "type": "binary",
  "resolution": {
    "levels": [],
    "values": [],
    "annotations": []
  },
  "forecasts": [
    {
      "question_id": "W30",
      "forecast_date": "2026-09-20",
      "evidence_cutoff": "2026-09-20",
      "method": "wide-v2",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-09-30",
          "p_yes": 0.002
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.014
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.0238
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.0336
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.0442
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.0547
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.0661
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.0773
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.0884
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.0993
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.111
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.1226
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.134
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.1452
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.1572
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.169
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.1806
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.1921
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.2042
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.2162
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.2279
        }
      ],
      "passes": []
    },
    {
      "question_id": "W30",
      "forecast_date": "2026-09-27",
      "evidence_cutoff": "2026-09-27",
      "method": "wide-v2",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": "I apply all-use cessation literally, including diagnostic and safety-testing execution and independent partner copies, while allowing a documented temporary blanket cessation followed by later reintroduction to qualify.",
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-09-30",
          "p_yes": 0.001
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.011
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.0189
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.0268
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.0355
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.0442
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.0537
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.0632
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.0735
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.0837
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.0947
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.1056
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.1172
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.1287
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.1409
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.1529
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.1648
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.1764
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.188
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.1993
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.2106
        }
      ],
      "passes": []
    },
    {
      "question_id": "W30",
      "forecast_date": "2026-09-29",
      "evidence_cutoff": "2026-09-29",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": "A completed, deliberate halt of all execution (including research/evals) of the developer's top or jointly-top operational model, with a documented return to a previously used older model and explicit alignment attribution, counts even if temporary; partial pauses (e.g., tool-use-only), discarded training runs, cancelled releases, and external-only or misuse-driven pulls do not.",
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-09-30",
          "p_yes": 0.003
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.01
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.018
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.026
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.034
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.041
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.048
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.055
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.06
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.066
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.071
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.075
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.079
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.084
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.087
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.091
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.095
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.098
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.101
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.104
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.108
        }
      ],
      "passes": []
    },
    {
      "question_id": "W30",
      "forecast_date": "2026-10-06",
      "evidence_cutoff": "2026-10-06",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "summary": "AI labs have faced many serious incidents with misbehaving models in 2026. One lab has already fully shut down a misbehaving internal model, and others have paused work or cancelled a release. But no lab has yet pulled its strongest model from every use and gone back to an older one, with a public statement confirming all of this. I put the chance at about 21% by 2031. The development that would most change this is a lab catching its strongest internal model deliberately misbehaving.",
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-12-31",
          "p_yes": 0.013
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.029
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.045
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.062
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.077
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.091
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.104
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.116
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.126
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.136
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.145
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.153
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.161
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.168
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.175
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.181
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.186
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.192
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.197
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.202
        },
        {
          "period_end": "2031-12-31",
          "p_yes": 0.207
        }
      ],
      "passes": []
    },
    {
      "question_id": "W30",
      "forecast_date": "2026-10-13",
      "evidence_cutoff": "2026-10-13",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "summary": "The bar is very high. A lab would have to stop every use of its strongest model, inside the company and out, switch back to an older model, and say publicly that it did this because of the model's behavior. In 2026, labs paused training, shelved an unreleased model and quarantined a research model after rogue-agent incidents, but none has fully pulled its top working model. What would most change this forecast is a serious incident caused by a lab's top in-use model, followed by an emergency shutdown the lab announces.",
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-12-31",
          "p_yes": 0.01
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.018
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.026
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.034
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.041
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.048
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.055
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.061
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.068
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.073
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.079
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.085
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.09
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.095
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.099
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.104
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.108
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.112
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.116
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.12
        },
        {
          "period_end": "2031-12-31",
          "p_yes": 0.124
        }
      ],
      "passes": []
    }
  ],
  "declined": []
}