{
  "question_id": "W17",
  "slug": "best-model-held-back-12-months",
  "title": "A lab holds its best model back for 12 months",
  "type": "binary",
  "resolution": {
    "levels": [],
    "values": [],
    "annotations": []
  },
  "forecasts": [
    {
      "question_id": "W17",
      "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": 0.001
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.004
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.008
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.0129
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.0199
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.0277
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.0364
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.0451
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.0547
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.0641
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.0739
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.0841
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.0947
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.1055
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.1163
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.1269
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.1373
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.1477
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.1579
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.168
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.178
        }
      ],
      "passes": []
    },
    {
      "question_id": "W17",
      "forecast_date": "2026-09-27",
      "evidence_cutoff": "2026-09-27",
      "method": "wide-v2",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": "Independent embedded evaluators count as outside recipients even when working on company premises, unless they act solely as internal staff under the developer’s direction.",
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-09-30",
          "p_yes": 0.0003
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.0028
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.0063
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.0108
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.0167
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.0226
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.0294
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.0362
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.043
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.0497
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.0573
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.0648
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.0723
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.0797
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.088
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.0962
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.1043
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.1124
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.1213
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.1301
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.1388
        }
      ],
      "passes": []
    },
    {
      "question_id": "W17",
      "forecast_date": "2026-09-30",
      "evidence_cutoff": "2026-09-30",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": "Per the stated definitions, access by independent/embedded evaluators (e.g., METR, Accenture/Faculty) or government testers to a model counts as release outside the company, continued training resets the clock, and a killed run's checkpoint counts as training ended only if the developer says training will not resume.",
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-09-30",
          "p_yes": 0.0002
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.0004
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.0007
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.0011
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.0019
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.0039
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.0054
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.0068
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.0081
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.0094
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.0107
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.0118
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.013
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.0142
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.0153
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.0164
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.0175
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.0185
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.0196
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.0207
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.0218
        }
      ],
      "passes": []
    },
    {
      "question_id": "W17",
      "forecast_date": "2026-10-07",
      "evidence_cutoff": "2026-10-07",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "summary": "Leading AI companies now keep some of their strongest models in-house, and OpenAI has paused training. But none has ever said its single best model was withheld from every outsider for a full year. That would need a company's progress to stall for a year, no outside testers or government reviewers getting access, and a dated public statement. A long, enforced pause at a major company, with its best model kept fully internal, would most raise this forecast.",
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-12-31",
          "p_yes": 0.0005
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.0013
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.0025
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.005
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.0085
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.0114
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.0144
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.0171
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.0196
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.0218
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.0239
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.0261
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.0282
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.03
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.0317
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.0335
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.0352
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.0367
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.0381
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.0395
        },
        {
          "period_end": "2031-12-31",
          "p_yes": 0.041
        }
      ],
      "passes": []
    },
    {
      "question_id": "W17",
      "forecast_date": "2026-10-14",
      "evidence_cutoff": "2026-10-14",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "summary": "I put this at about 3% by the end of 2031. To qualify, a lab would have to keep the same model as its best for a full year. No outsider, not even government testers or independent evaluators, could use it in that time, and the lab would have to say so explicitly with dates. Right now labs replace their top models every few months and routinely show them to outside testers. A long training freeze, voluntary or imposed by government, would most change this.",
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-12-31",
          "p_yes": 0.0005
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.001
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.0018
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.003
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.0045
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.0063
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.0083
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.0103
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.0122
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.0142
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.016
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.0178
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.0193
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.0208
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.0223
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.0236
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.025
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.0263
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.0275
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.0287
        },
        {
          "period_end": "2031-12-31",
          "p_yes": 0.03
        }
      ],
      "passes": []
    }
  ],
  "declined": []
}