{
  "question_id": "W31",
  "slug": "lab-pays-capability-cost-for-readable-reasoning",
  "title": "A lab publishes the capability it gave up to keep reasoning readable",
  "type": "binary",
  "resolution": {
    "levels": [],
    "values": [],
    "annotations": []
  },
  "forecasts": [
    {
      "question_id": "W31",
      "forecast_date": "2026-09-20",
      "evidence_cutoff": "2026-09-20",
      "method": "wide-v2",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": "Readability-motivated language or training constraints can qualify without rejecting a latent architecture, but the publication must establish a measured sacrifice for the implemented frontier-model choice rather than only a small-model ablation or a monitorability improvement.",
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-09-30",
          "p_yes": 0.01
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.0447
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.0876
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.1378
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.1896
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.2382
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.2839
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.3269
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.3672
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.4052
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.4349
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.4632
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.49
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.5155
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.5349
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.5535
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.5714
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.5885
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.6009
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.6128
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.6245
        }
      ],
      "passes": []
    },
    {
      "question_id": "W31",
      "forecast_date": "2026-09-27",
      "evidence_cutoff": "2026-09-27",
      "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": 0.005
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.0398
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.083
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.1289
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.1768
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.2262
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.2726
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.3163
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.3573
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.3958
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.4261
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.4548
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.482
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.5079
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.5276
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.5465
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.5646
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.582
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.5967
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.6108
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.6244
        }
      ],
      "passes": []
    },
    {
      "question_id": "W31",
      "forecast_date": "2026-09-30",
      "evidence_cutoff": "2026-09-30",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": "Choices framed by the developer as keeping reasoning legibly verbalized in text count; that covers architecture limits, training constraints, checkpoint or model retention, and inference configuration, including monitorability-preserving choices explicitly tied to readable CoT. Only official developer publications count, not executive social posts or press interviews. A numerical capability loss against an evaluated alternative for a frontier model is required.",
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-09-30",
          "p_yes": 0.004
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.042
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.078
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.11
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.137
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.162
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.185
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.206
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.226
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.244
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.261
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.277
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.291
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.305
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.317
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.329
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.341
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.352
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.362
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.373
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.382
        }
      ],
      "passes": []
    },
    {
      "question_id": "W31",
      "forecast_date": "2026-10-07",
      "evidence_cutoff": "2026-10-07",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "summary": "No artificial intelligence company has yet published a figure showing how much performance it sacrificed to keep a top model's reasoning in readable language; labs usually describe such safety choices without numbers. The issue became prominent after OpenAI's newest model moved some reasoning out of view, so a quantified disclosure within five years is plausible but less likely than not. A lab rejecting a hidden-reasoning design and stating the cost would most change this.",
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-12-31",
          "p_yes": 0.02
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.046
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.072
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.097
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.122
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.143
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.163
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.183
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.203
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.22
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.236
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.252
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.268
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.281
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.294
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.307
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.319
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.33
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.341
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.351
        },
        {
          "period_end": "2031-12-31",
          "p_yes": 0.362
        }
      ],
      "passes": []
    },
    {
      "question_id": "W31",
      "forecast_date": "2026-10-14",
      "evidence_cutoff": "2026-10-14",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "summary": "AI labs are under heavy pressure to keep their models' reasoning readable, and OpenAI staff have said informally that they limited a new technique to keep it so. Still, no lab has officially published how much performance it gave up for that reason, and labs rarely admit lost capability. I put the chance at about 3% by end-2026, 14% by end-2027 and 32% by end-2031. The biggest swing factor is whether OpenAI's promised write-up on its architecture or monitoring approach includes such numbers.",
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-12-31",
          "p_yes": 0.032
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.068
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.096
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.118
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.138
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.155
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.171
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.186
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.2
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.213
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.224
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.236
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.247
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.257
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.267
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.277
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.285
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.294
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.302
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.309
        },
        {
          "period_end": "2031-12-31",
          "p_yes": 0.316
        }
      ],
      "passes": []
    }
  ],
  "declined": []
}