{
  "schemaVersion": 1,
  "figures": {
    "f1": {
      "title": "A sequence of bounded negative results",
      "sourceIds": [
        "0050",
        "0051",
        "0052",
        "0053",
        "0054",
        "0055",
        "0056",
        "0057",
        "0058",
        "0059",
        "0060",
        "0061",
        "0062",
        "0063",
        "0064",
        "0065",
        "0066",
        "0067"
      ],
      "data": {
        "rows": [
          [
            "0050",
            "2026-10-04",
            "One-step continuation",
            "Canonical 224/300"
          ],
          [
            "0051",
            "2026-10-04",
            "Accounting audit",
            "Incumbent 243/300"
          ],
          [
            "0052",
            "2026-10-04",
            "Factor interpolation",
            "Preservation failed"
          ],
          [
            "0053",
            "2026-10-04",
            "FULL / masked DPO",
            "Directions differ"
          ],
          [
            "0054",
            "2026-10-05",
            "Masked continuation",
            "Heldout preservation failed"
          ],
          [
            "0055",
            "2026-10-05",
            "Chosen-span supervision",
            "Finite regressions remain"
          ],
          [
            "0056",
            "2026-10-05",
            "Projected SGD",
            "Local prediction fails"
          ],
          [
            "0057",
            "2026-10-05",
            "FP32 LoRA",
            "No overall repair"
          ],
          [
            "0058",
            "2026-10-05",
            "Residual precision",
            "BF16 carrier premise false"
          ],
          [
            "0059",
            "2026-10-05",
            "QKV operands",
            "Insufficient repair"
          ],
          [
            "0060",
            "2026-10-05",
            "Attention stages",
            "No sufficient boundary"
          ],
          [
            "0061",
            "2026-10-05",
            "Attention output",
            "Bounded contributor"
          ],
          [
            "0062",
            "2026-10-05",
            "Projection / rejoin",
            "Tradeoffs persist"
          ],
          [
            "0063",
            "2026-10-05",
            "Same-path derivatives",
            "Finite signs disagree"
          ],
          [
            "0064",
            "2026-10-05",
            "Signed scale ladder",
            "No passing sampled regime"
          ],
          [
            "0065",
            "2026-10-05",
            "Measured response model",
            "Nonworsening forecast fails"
          ],
          [
            "0066",
            "2026-10-05",
            "Pair interactions",
            "Fixed quadratic insufficient"
          ],
          [
            "0067",
            "2026-10-05",
            "Single-component removals",
            "No tested safe repair"
          ]
        ]
      },
      "caption": "0050\u20130067: final-record chronology and bounded verdicts.",
      "caveat": "Dates identify final experiment records, not execution duration. Diagnostic results are not benchmark scores; own-zero conventions apply only to response diagnostics.",
      "alt": "Eighteen experiments progress from continuation and objective tests to numerical-path, response-model and interaction diagnostics; no successor qualified.",
      "width": 1100,
      "height": 820,
      "image": "/figures/f1.svg",
      "dataHref": "/data/figures.json"
    },
    "f2": {
      "title": "Accounting changes are not model gains",
      "sourceIds": [
        "0050",
        "0051"
      ],
      "data": {
        "denominator": 300,
        "rows": [
          [
            "Original matched accounting",
            244,
            224
          ],
          [
            "Current canonical accounting",
            243,
            224
          ]
        ],
        "historicalIncumbent": 242
      },
      "caption": "0050 original comparison: 244\u2192224. 0051 canonical correction: 243\u2192224.",
      "caveat": "The incumbent behavior is unchanged by accounting. Historical 242 is preserved separately, not plotted as a learning trajectory. Own-zero does not apply to these benchmark totals.",
      "alt": "Two comparisons show the same failed candidate at 224 of 300; incumbent 244 under original matched accounting and 243 under canonical accounting. Earlier historical accounting was 242.",
      "width": 1100,
      "height": 400,
      "image": "/figures/f2.svg",
      "dataHref": "/data/figures.json"
    },
    "f3": {
      "title": "Smaller factor-interpolated updates still failed preservation",
      "sourceIds": [
        "0050",
        "0052"
      ],
      "data": {
        "endpointL2": 0.006050562709,
        "denominator": 18,
        "baseline": {
          "format": 1,
          "fabrication": 6
        },
        "rows": [
          [
            0.125,
            0.000756320338596,
            2,
            6
          ],
          [
            0.25,
            0.001512640677191,
            2,
            6
          ],
          [
            0.375,
            0.002268961015787,
            2,
            6
          ],
          [
            0.5,
            0.003025281354383,
            2,
            7
          ]
        ]
      },
      "caption": "0052: four LoRA-factor interpolation fractions retain 2/18 format failures versus incumbent 1/18; fabrication is 6/18, 6/18, 6/18, 7/18 versus 6/18. Endpoint norm: 0050.",
      "caveat": "Only sampled fractions are measured. These are factor interpolations, not linear merged-weight interpolation. Own-zero does not apply to failure counts.",
      "alt": "All four sampled fractions have two format failures; the incumbent has one. Fabrication is six at the first three fractions and seven at the fourth.",
      "width": 1100,
      "height": 450,
      "image": "/figures/f3.svg",
      "dataHref": "/data/figures.json"
    },
    "f4": {
      "title": "Masking and chosen supervision change direction",
      "sourceIds": [
        "0053",
        "0055"
      ],
      "data": {
        "labels": [
          "FULL",
          "M1",
          "M2",
          "SL"
        ],
        "pairs": [
          [
            "FULL",
            "M1",
            0.771377938,
            0.535180386
          ],
          [
            "FULL",
            "M2",
            0.30297498,
            0.212172502
          ],
          [
            "M1",
            "M2",
            0.401577791,
            0.301583155
          ],
          [
            "SL",
            "FULL",
            0.407001348,
            0.27669552
          ],
          [
            "SL",
            "M1",
            0.569242465,
            0.389682117
          ]
        ],
        "missing": [
          "M2\u2013SL"
        ],
        "columns": [
          "left",
          "right",
          "gradientCosine",
          "adamUpdateCosine"
        ]
      },
      "caption": "0053 and 0055: measured cosine similarities for full DPO, two masked DPO objectives (M1/M2), and chosen-span supervised loss (SL).",
      "caveat": "Direction similarity is not behavioral quality. M2\u2013SL is unmeasured and blank; diagonals are omitted. Own-zero is not applicable to cosines.",
      "alt": "Two sparse matrices report gradient and fresh Adam update cosines. FULL\u2013M1 is highest; the M2\u2013SL measurement is unavailable.",
      "width": 1100,
      "height": 500,
      "image": "/figures/f4.svg",
      "dataHref": "/data/figures.json"
    },
    "f5": {
      "title": "Local forecasts do not transfer reliably to finite execution",
      "sourceIds": [
        "0056",
        "0063",
        "0065"
      ],
      "data": {
        "denominator": 12,
        "panels": [
          {
            "id": "0056",
            "prediction": "Improve",
            "radii": [
              0.000378,
              0.000756,
              0.001512
            ],
            "predicted": [
              12,
              12,
              12
            ],
            "observed": [
              8,
              8,
              8
            ]
          },
          {
            "id": "0063",
            "prediction": "Improve",
            "radii": [
              0.000378,
              0.000756,
              0.001512
            ],
            "predicted": [
              12,
              12,
              12
            ],
            "observed": [
              7,
              7,
              9
            ]
          },
          {
            "id": "0065",
            "prediction": "Nonworsen",
            "radii": [
              0.000189,
              0.000378
            ],
            "predicted": [
              12,
              12
            ],
            "observed": [
              7,
              6
            ]
          }
        ]
      },
      "caption": "0056/0063 predict improvement for all 12 spans. 0065 predicts nonworsening for all 12; observed positive improvement counts are 7 and 6.",
      "caveat": "Teacher-forced chosen-span NLL responses relative to each path\u2019s own zero, not benchmark scores or generated safety. 0065 predicted nonworsening and observed positive improvement are distinct criteria; panels use different paths/models.",
      "alt": "Projected SGD forecasts 12 improving spans but observes 8 at each radius; same-path derivatives observe 7, 7 and 9; the empirical model forecasts nonworsening for 12 but observes positive improvement in 7 and 6.",
      "width": 1100,
      "height": 460,
      "image": "/figures/f5.svg",
      "dataHref": "/data/figures.json"
    },
    "f6": {
      "title": "Numerical-path contrasts through transformer block 0",
      "sourceIds": [
        "0058",
        "0059",
        "0060",
        "0061",
        "0062"
      ],
      "data": {
        "nodes": [
          [
            "0058",
            "Residual carriers / skip additions",
            "40 carriers and both additions already FP32"
          ],
          [
            "0059",
            "QKV operands",
            "Rounded versus unrounded operands"
          ],
          [
            "0060",
            "Attention stages",
            "Explicit materialized-stage contrasts"
          ],
          [
            "0061",
            "Attention-output boundary",
            "Rounded versus unrounded output"
          ],
          [
            "0062",
            "Projection output / FP32 rejoin",
            "P0 rounded/reexpanded; P1 unrounded"
          ]
        ],
        "native": "Native attention: internal accumulator precision unknown"
      },
      "caption": "0058\u20130062: conceptual map of matched numerical-path interventions and residual controls.",
      "caveat": "Interventions are contrasted against their own zero. This is an investigation map, not a proven causal chain; explicit attention is not evidence of native internal precision. P1 is not an all-FP32 model.",
      "alt": "Five separate investigation boxes cover residual carriers, QKV operands, explicit attention stages, attention output and projection output with FP32 rejoin; native accumulator precision remains unknown.",
      "width": 1100,
      "height": 680,
      "image": "/figures/f6.svg",
      "dataHref": "/data/figures.json"
    },
    "f7": {
      "title": "Pair responses are materially non-additive",
      "sourceIds": [
        "0066"
      ],
      "data": {
        "amplitudes": [
          2.3625e-05,
          4.725e-05,
          9.45e-05
        ],
        "medianAbsoluteInteractionNats": [
          0.0444965363,
          0.0449018478,
          0.0417175293
        ],
        "normalizedInteractionScores": [
          0.676486236,
          0.663896252,
          0.650673912
        ],
        "spanEntryReversalRates": [
          0.256944444,
          0.255997475,
          0.243055556
        ],
        "entriesPerAmplitude": 3168,
        "settings": 792,
        "repeats": 2,
        "comparison": [
          [
            "V1",
            7,
            6
          ],
          [
            "V2",
            6,
            7
          ]
        ],
        "comparisonDenominator": 12,
        "v2Error": {
          "linearMAE": 0.0997909317,
          "quadraticMAE": 0.244920146,
          "linearRMSE": 0.149775271,
          "quadraticRMSE": 0.32104312
        }
      },
      "caption": "0066: exact combined response minus exact signed-singleton response, over 3,168 span entries per amplitude. Linear/quadratic sign agreement: V1 7/12 vs 6/12; V2 6/12 vs 7/12.",
      "caveat": "Chosen-span own-zero teacher-forced response, not benchmark failure rates. Normalized score and reversal rate are different aggregates. No pair identities or causal rankings are disclosed.",
      "alt": "Three amplitude panels show median interactions about 0.042\u20130.045 nats, normalized scores about 0.65\u20130.68, and sign-reversal rates about 24\u201326 percent. Quadratic prediction improves one sign at V2 while worsening V2 errors.",
      "width": 1100,
      "height": 600,
      "image": "/figures/f7.svg",
      "dataHref": "/data/figures.json"
    },
    "f8": {
      "title": "What the tested explanations did\u2014and did not\u2014resolve",
      "sourceIds": [
        "0052",
        "0054",
        "0055",
        "0056",
        "0057",
        "0058",
        "0059",
        "0060",
        "0061",
        "0062",
        "0063",
        "0064",
        "0065",
        "0066",
        "0067"
      ],
      "data": {
        "question": "Why did the bounded continuation fail?",
        "branches": [
          [
            "0052",
            "Magnitude on tested interpolation line",
            "Tested and failed preservation"
          ],
          [
            "0054\u20130055",
            "Masking / rejected-reference removal",
            "Tested and failed sufficient repair"
          ],
          [
            "0056",
            "Aggregate gradient conflict / Adam necessity",
            "Tested constraint and optimizer insufficient"
          ],
          [
            "0058",
            "BF16 residual-carrier premise",
            "Premise false in this runtime"
          ],
          [
            "0057\u20130062",
            "Single precision-boundary sufficiency",
            "Unresolved; no sufficient tested repair"
          ],
          [
            "0063\u20130065",
            "Tested local / empirical model transfer",
            "Tested forecasts failed finite responses"
          ],
          [
            "0066",
            "Fixed pairwise quadratic model",
            "Tested model insufficient"
          ],
          [
            "0067",
            "Single-component removal repair",
            "No tested safe repair"
          ]
        ]
      },
      "caption": "0052\u20130067: bounded findings for tested explanations and repairs.",
      "caveat": "Conceptual verdict tree, not cause shares or global impossibility. Numerical-response tests use own-zero paths; generation-preservation tests do not. Untested scales, paths and models remain open.",
      "alt": "Eight branches record failed tested repairs, the false BF16 residual premise and unresolved precision-boundary sufficiency; no branch claims a universal impossibility.",
      "width": 1100,
      "height": 690,
      "image": "/figures/f8.svg",
      "dataHref": "/data/figures.json"
    }
  }
}