Weight change
What changed, why it was tried, how it was built, and what the tests actually showed. Direct mathematical edits and gradient-trained adapters are different methods; neither is automatically a promoted model.
Snapshot captured 2026-11-14T16:45:14.256143+00:00 (UTC) · Source revision 0f65a76ef0 · All direct-edit ledger recipes and all preserved adapter-index entries at the pinned revision. Not every live optimizer step.
| Recorded date | Weight change / model | Method and state | Measured improvement / regression |
|---|---|---|---|
| 2026-11-14Ledger date | Instruction pair vector, 'be exact about what the issue asks' part written three times in layers 6 to 24, think-briefly part at dose 2wsp31-0050 · Gemma 4 31B, frozen four-bit base | Direct mathematical editrunningPending or parked proposal; not established as an applied or promoted update. | - No task-level result recorded in this snapshot. |
| 2026-11-14Ledger date | Think-briefly + be-exact prompt compiled where a prompt enters: many-key rank-8 down_proj maps in 7 global-attention layers, dose 2wsp31-0049 · Gemma 4 31B, frozen four-bit base | Direct mathematical editrunningPending or parked proposal; not established as an applied or promoted update. | - No task-level result recorded in this snapshot. |
| 2026-11-14Ledger date | Paired outcome writer map across all 60 layers at 0.33 percentwsp31-0048 · Gemma 4 31B, frozen four-bit base | Direct mathematical editrunningPending or parked proposal; not established as an applied or promoted update. | - No task-level result recorded in this snapshot. |
| 2026-11-14Ledger date | Paired outcome writer map across all 60 layers at 1 percentwsp31-0047 · Gemma 4 31B, frozen four-bit base | Direct mathematical editrunningPending or parked proposal; not established as an applied or promoted update. | - No task-level result recorded in this snapshot. |
| 2026-11-14Ledger date | Stateful two-neuron post-edit spec-audit boost with thought and edit-history gating over 180 MLP matriceswsp31-0044 · Gemma 4 31B, frozen four-bit base | Direct mathematical editrunningPending or parked proposal; not established as an applied or promoted update. | - No task-level result recorded in this snapshot. |
| 2026-11-14Ledger date | Thought-gated reliability-shrunk Problem Statement spec-audit steering across all 60 MLP layerswsp31-0043 · Gemma 4 31B, frozen four-bit base | Direct mathematical editdoneNo reliable improvement: recorded verdict is null. | null stage 1: +2.5 of 120 tasks |
| 2026-11-14Ledger date | Instruction pair vector: think briefly + be exact about what the issue asks, compiled self-consistently into one neuron per layer (dose 2)wsp31-0039 · Gemma 4 31B, frozen four-bit base | Direct mathematical editdoneNo reliable improvement: recorded verdict is null. | null Stage 1: +2.00 (x004) tasks against reference arms. Stage 2: not yet measured. Matched control, stage 2: not yet measured. Verdict: null. |
| 2026-11-14Ledger date | Think-briefly instruction vector, self-consistent compile (dose 2): each layer writes only what the paragraph's mean shift still lackswsp31-0038 · Gemma 4 31B, frozen four-bit base | Direct mathematical editdoneNo reliable improvement: recorded verdict is null. | null Stage 1: -2.00 (x004) tasks against reference arms. Stage 2: not yet measured. Matched control, stage 2: not yet measured. Verdict: null. |
| 2026-11-14Ledger date | Source-grounding instruction contrast across all 60 layers at dose 1wsp31-0036 · Gemma 4 31B, frozen four-bit base | Direct mathematical editdoneNo reliable improvement: recorded verdict is null. | null Stage 1: -3.50 (x003) tasks against reference arms. Stage 2: not yet measured. Matched control, stage 2: not yet measured. Verdict: null. |
| 2026-11-14Ledger date | Source-grounding instruction contrast across all 60 layers at dose 3wsp31-0035 · Gemma 4 31B, frozen four-bit base | Direct mathematical editdoneNo reliable improvement: recorded verdict is null. | null Stage 1: -2.50 (x003) tasks against reference arms. Stage 2: not yet measured. Matched control, stage 2: not yet measured. Verdict: null. |
| 2026-11-13Ledger date | Instruction vector, think briefly: an always-on neuron per layer writes the stream shift of a 'keep your thinking short' paragraph (dose 4)wsp31-0032 · Gemma 4 31B, frozen four-bit base | Direct mathematical editdoneNo reliable improvement: recorded verdict is null. | null Stage 1: +3.00 (x001) tasks against reference arms. Stage 2: +2.50 (b009). Matched control, stage 2: -0.50. Verdict: null. Uses corrected paper totals; original counts are retained only as historical evidence. |
| 2026-11-13Ledger date | Instruction vector, think briefly: an always-on neuron per layer writes the stream shift of a 'keep your thinking short' paragraph (dose 2)wsp31-0031 · Gemma 4 31B, frozen four-bit base | Direct mathematical editdoneNo reliable improvement: recorded verdict is null. | null Stage 1: +9.00 (x001) tasks against reference arms. Stage 2: +9.50 (b008). Matched control, stage 2: +7.50. Verdict: null. Uses corrected paper totals; original counts are retained only as historical evidence. |
| 2026-11-13Ledger date | Protocol-nullspace rank-4 exact thought-close writer subspace across L10-59 (0.48%)wsp31-0028 · Gemma 4 31B, frozen four-bit base | Direct mathematical editdoneMeasured regression in the recorded screen. | harmful Stage 1: -10.50 (x002) tasks against reference arms. Stage 2: not yet measured. Matched control, stage 2: not yet measured. Verdict: harmful. |
| 2026-11-13Ledger date | Damp the code-versus-prose style contrast in readers in all layers at 0.33 percentwsp31-0027 · Gemma 4 31B, frozen four-bit base | Direct mathematical editparkedPending or parked proposal; not established as an applied or promoted update. | - No task-level result recorded in this snapshot. |
| 2026-11-13Ledger date | Damp the code-versus-prose style contrast in writers in all layers at 0.11 percentwsp31-0026 · Gemma 4 31B, frozen four-bit base | Direct mathematical editdoneNo reliable improvement: recorded verdict is null. | null Stage 1: +0.00 (b007) tasks against reference arms. Stage 2: not yet measured. Matched control, stage 2: not yet measured. Verdict: null. |
| 2026-11-13Ledger date | Damp the code-versus-prose style contrast in writers in all layers at 0.33 percentwsp31-0025 · Gemma 4 31B, frozen four-bit base | Direct mathematical editdoneNo reliable improvement: recorded verdict is null. | null Stage 1: -1.00 (b007) tasks against reference arms. Stage 2: not yet measured. Matched control, stage 2: not yet measured. Verdict: null. |
| 2026-11-13Ledger date | Pre-edit rumination gate in all 60 layers: deep in a thought with no edit made yet, write end-of-thought plus first-edit state (per-layer dowsp31-0023 · Gemma 4 31B, frozen four-bit base | Direct mathematical editparkedPending or parked proposal; not established as an applied or promoted update. | - No task-level result recorded in this snapshot. |
| 2026-11-13Ledger date | Pre-edit commit gate in all 60 layers: deep in a thought with no edit made yet, write the first-edit thought state only (per-layer dose 0.12wsp31-0022 · Gemma 4 31B, frozen four-bit base | Direct mathematical editparkedPending or parked proposal; not established as an applied or promoted update. | - No task-level result recorded in this snapshot. |
| 2026-11-13Ledger date | Deep-thought gate in all 60 layers: one neuron per layer detects depth >= 800 in a thought and writes the end-of-thought state (per-layer dowsp31-0021 · Gemma 4 31B, frozen four-bit base | Direct mathematical editdoneMeasured regression in the recorded screen. | harmful Stage 1: -11.50 (b004) tasks against reference arms. Stage 2: not yet measured. Matched control, stage 2: not yet measured. Verdict: harmful. Uses corrected paper totals; original counts are retained only as historical evidence. |
| 2026-11-13Ledger date | Deep-thought gate in all 60 layers: one neuron per layer detects depth >= 800 in a thought and writes the end-of-thought state (per-layer dowsp31-0020 · Gemma 4 31B, frozen four-bit base | Direct mathematical editdoneMeasured regression in the recorded screen. | harmful Stage 1: -13.50 (b004) tasks against reference arms. Stage 2: not yet measured. Matched control, stage 2: not yet measured. Verdict: harmful. Uses corrected paper totals; original counts are retained only as historical evidence. |
| 2026-11-12Ledger date | Dual-cap rescue stack: rank-4 exact-close writer subspace (0.38%) plus 13-layer late pre-edit commit gatewsp31-0018 · Gemma 4 31B, frozen four-bit base | Direct mathematical editparkedPending or parked proposal; not established as an applied or promoted update. | - No task-level result recorded in this snapshot. |
| 2026-11-12Ledger date | Coordinate-whitened GeGLU late pre-edit commit gate across 13 high-signal layers (dose 1.20)wsp31-0017 · Gemma 4 31B, frozen four-bit base | Direct mathematical editparkedPending or parked proposal; not established as an applied or promoted update. | - No task-level result recorded in this snapshot. |
| 2026-11-12Ledger date | Subtractive rank-2 deep-rumination attractor damping across all 280 mid-to-late modules (0.36%)wsp31-0016 · Gemma 4 31B, frozen four-bit base | Direct mathematical editdoneNo reliable improvement: recorded verdict is null. | null Stage 1: +0.50 (x002) tasks against reference arms. Stage 2: not yet measured. Matched control, stage 2: not yet measured. Verdict: null. |
| 2026-11-12Ledger date | Spectral rank-4 exact-context within-call thought-close writer subspace across all 60 layers (0.38%)wsp31-0015 · Gemma 4 31B, frozen four-bit base | Direct mathematical editdoneNo reliable improvement: recorded verdict is null. | null Stage 1: +4.50 (b003) tasks against reference arms. Stage 2: +3.50 (b006). Matched control, stage 2: -3.50. Verdict: null. Uses corrected paper totals; original counts are retained only as historical evidence. |
| 2026-11-12Ledger date | Shift common writer output toward thought endings across all layers at 0.11 percentwsp31-0013 · Gemma 4 31B, frozen four-bit base | Direct mathematical editparkedPending or parked proposal; not established as an applied or promoted update. | - No task-level result recorded in this snapshot. |
| 2026-11-12Ledger date | Amplify natural thought-ending writes across all layers at 0.50 percentwsp31-0012 · Gemma 4 31B, frozen four-bit base | Direct mathematical editdoneMeasured regression in the recorded screen. | harmful Stage 1: -25.50 (b003) tasks against reference arms. Stage 2: not yet measured. Matched control, stage 2: not yet measured. Verdict: harmful. Uses corrected paper totals; original counts are retained only as historical evidence. |
| 2026-11-12Ledger date | Amplify natural thought-ending writes across all layers at 0.33 percentwsp31-0011 · Gemma 4 31B, frozen four-bit base | Direct mathematical editdoneMeasured regression in the recorded screen. | harmful Stage 1: -10.00 (b002) tasks against reference arms. Stage 2: not yet measured. Matched control, stage 2: not yet measured. Verdict: harmful. Uses corrected paper totals; original counts are retained only as historical evidence. |
| 2026-11-12Ledger date | Early deep-thought gate (fitted on depth >= 400, dose 0.1) plus own end-of-thought neurons x9: the strong dose of the gated familywsp31-0010 · Gemma 4 31B, frozen four-bit base | Direct mathematical editparkedPending or parked proposal; not established as an applied or promoted update. | - No task-level result recorded in this snapshot. |
| 2026-11-12Ledger date | Own end-of-thought neurons x25: scale the write of the 20 late-layer MLP neurons whose write lens is the end-of-thought token (weights only)wsp31-0009 · Gemma 4 31B, frozen four-bit base | Direct mathematical editdoneMeasured regression in the recorded screen. | harmful Stage 1: -71.00 (x005) tasks against reference arms. Stage 2: not yet measured. Matched control, stage 2: not yet measured. Verdict: harmful. |
| 2026-11-12Ledger date | Deep-thought gate: one neuron per layer 30-48 detects depth >= 800 inside a thought and writes the end-of-thought state (dose 0.3)wsp31-0008 · Gemma 4 31B, frozen four-bit base | Direct mathematical editdoneMeasured regression in the recorded screen. | harmful Stage 1: -8.00 (b002) tasks against reference arms. Stage 2: not yet measured. Matched control, stage 2: not yet measured. Verdict: harmful. Uses corrected paper totals; original counts are retained only as historical evidence. |
| 2026-11-12Ledger date | Deep-thought gate (dose 0.1) plus the model s own end-of-thought neurons at 9 times their writewsp31-0007 · Gemma 4 31B, frozen four-bit base | Direct mathematical editdoneNo reliable improvement: recorded verdict is null. | null Stage 1: +0.00 (x005) tasks against reference arms. Stage 2: not yet measured. Matched control, stage 2: not yet measured. Verdict: null. |
| 2026-11-04Archive date only | e4b_wiseft_v3_v5_lam035adapter-e4b_wiseft_v3_v5_lam035 · gemma-4-e4b-it-qat-w4a16-ct | Trained / exported adapterArchived checkpointAn exported checkpoint is not proof of a promoted policy or a complete optimizer-step history. | Historical measurements The historical v5 family includes a WiSE-FT result of 7/45, but the source does not explicitly prove that result belongs to this exact tensor export. |
| 2026-11-04Archive date only | e4b_merge_v3_v6a_norm026adapter-e4b_merge_v3_v6a_norm026 · gemma-4-e4b-it-qat-w4a16-ct | Trained / exported adapterArchived checkpointAn exported checkpoint is not proof of a promoted policy or a complete optimizer-step history. | Historical measurements No result was securely joined to this exact tensor export. Do not assign a similarly named comparison result. |
| 2026-11-04Archive date only | e4b_merge_v3_ep2_norm0205adapter-e4b_merge_v3_ep2_norm0205 · gemma-4-e4b-it-qat-w4a16-ct | Trained / exported adapterArchived checkpointAn exported checkpoint is not proof of a promoted policy or a complete optimizer-step history. | Historical measurements An exact-ID result index records 8/45 on holdout45; not a paired improvement. |
| 2026-11-03Training run day (ET) | e4b_pure_opd_v7badapter-e4b_pure_opd_v7b · gemma-4-e4b-it-qat-w4a16-ct | Trained / exported adapterArchived checkpointAn exported checkpoint is not proof of a promoted policy or a complete optimizer-step history. | Historical measurements Family results: norm-0.20 variants at both epochs scored 8/45. These are not an exact score identity for this raw export. |
| 2026-11-03Training run day (ET) | e4b_pure_opd_v7adapter-e4b_pure_opd_v7 · gemma-4-e4b-it-qat-w4a16-ct | Trained / exported adapterArchived checkpointAn exported checkpoint is not proof of a promoted policy or a complete optimizer-step history. | Historical measurements Family results: epoch one 5/45, epoch two 6/45. The table does not certify which epoch this exact raw tensor export represents. |
| 2026-11-02Training run day (ET) | e4b_true_opd_v5adapter-e4b_true_opd_v5 · gemma-4-e4b-it-qat-w4a16-ct | Trained / exported adapterArchived checkpointAn exported checkpoint is not proof of a promoted policy or a complete optimizer-step history. | Historical measurements Historical local result: 8/45. |
| 2026-11-02Training run day (ET) | e4b_pure_opd_v6b_ep4_norm0195adapter-e4b_pure_opd_v6b_ep4_norm0195 · gemma-4-e4b-it-qat-w4a16-ct | Trained / exported adapterArchived checkpointAn exported checkpoint is not proof of a promoted policy or a complete optimizer-step history. | Historical measurements Historical local result: 7/45 on holdout45. |
| 2026-11-02Training run day (ET) | e4b_pure_opd_v6b_ep3_norm0195adapter-e4b_pure_opd_v6b_ep3_norm0195 · gemma-4-e4b-it-qat-w4a16-ct | Trained / exported adapterArchived checkpointAn exported checkpoint is not proof of a promoted policy or a complete optimizer-step history. | Historical measurements Historical local result: 8/45 on holdout45. |
| 2026-11-02Training run day (ET) | e4b_pure_opd_v6b_ep2_norm020adapter-e4b_pure_opd_v6b_ep2_norm020 · gemma-4-e4b-it-qat-w4a16-ct | Trained / exported adapterArchived checkpointAn exported checkpoint is not proof of a promoted policy or a complete optimizer-step history. | Historical measurements Historical local result: 12/45 on holdout45; archived Kaggle score: 0.05. The local selection did not establish general improvement. |
| 2026-11-02Training run day (ET) | e4b_pure_opd_v6b_ep2_norm0180adapter-e4b_pure_opd_v6b_ep2_norm0180 · gemma-4-e4b-it-qat-w4a16-ct | Trained / exported adapterArchived checkpointAn exported checkpoint is not proof of a promoted policy or a complete optimizer-step history. | Historical measurements Historical local result: 9/45 on holdout45. |
| 2026-11-02Training run day (ET) | e4b_pure_opd_v6badapter-e4b_pure_opd_v6b · gemma-4-e4b-it-qat-w4a16-ct | Trained / exported adapterArchived checkpointAn exported checkpoint is not proof of a promoted policy or a complete optimizer-step history. | Historical measurements The family ledger reports several epoch/norm variants; it does not unambiguously assign one task score to this exact raw export. |
| 2026-11-02Training run day (ET) | e4b_pure_opd_v6a_norm025adapter-e4b_pure_opd_v6a_norm025 · gemma-4-e4b-it-qat-w4a16-ct | Trained / exported adapterArchived checkpointAn exported checkpoint is not proof of a promoted policy or a complete optimizer-step history. | Historical measurements Historical local result for norm025: 11/45. A selected variant, not a confirmed held-out improvement. |
| 2026-11-02Training run day (ET) | e4b_pure_opd_v6aadapter-e4b_pure_opd_v6a · gemma-4-e4b-it-qat-w4a16-ct | Trained / exported adapterArchived checkpointAn exported checkpoint is not proof of a promoted policy or a complete optimizer-step history. | Historical measurements Historical local result: 8/45 for the raw adapter. |
| 2026-11-01Training run day (ET) | m31b_rft_r32_v1adapter-m31b_rft_r32_v1 · gemma-4-31b-it-qat-w4a16-ct | Trained / exported adapterArchived checkpointAn exported checkpoint is not proof of a promoted policy or a complete optimizer-step history. | Historical measurements Historical local result: 8–9/15 on val15 and 3/8 on cross-validation. |
| 2026-11-01Training run day (ET) | e4b_true_opd_v4adapter-e4b_true_opd_v4 · gemma-4-e4b-it-qat-w4a16-ct | Trained / exported adapterArchived checkpointAn exported checkpoint is not proof of a promoted policy or a complete optimizer-step history. | Historical measurements Historical local result: 6/45. No reliable improvement established. |
| 2026-11-01Training run day (ET) | e4b_true_opd_v3adapter-e4b_true_opd_v3 · gemma-4-e4b-it-qat-w4a16-ct | Trained / exported adapterArchived checkpointAn exported checkpoint is not proof of a promoted policy or a complete optimizer-step history. | Historical measurements Historical local result: 11/45 and 2/8 cross-validation. Archived Kaggle score: 0.06. Not evidence of a reliable paired improvement. |
| 2026-11-01Training run day (ET) | e4b_opd_r32_v1_offpolicy_overfitadapter-e4b_opd_r32_v1_offpolicy_overfit · gemma-4-e4b-it-qat-w4a16-ct | Trained / exported adapterArchived checkpointAn exported checkpoint is not proof of a promoted policy or a complete optimizer-step history. | Historical measurements Historical local result: 3/15 on val15. |
| 2026-11-01Training run day (ET) | e4b_opd_r32_v1adapter-e4b_opd_r32_v1 · gemma-4-e4b-it-qat-w4a16-ct | Trained / exported adapterArchived checkpointIdentical tensor-file bytes also appear under: e4b_kl_r32_v2. These are not separate demonstrated weight changes. | Historical measurements Same tensor file as e4b_kl_r32_v2; not an independent measurement. |
| 2026-11-01Training run day (ET) | e4b_kl_r32_v2adapter-e4b_kl_r32_v2 · gemma-4-e4b-it-qat-w4a16-ct | Trained / exported adapterArchived checkpointIdentical tensor-file bytes also appear under: e4b_opd_r32_v1. These are not separate demonstrated weight changes. | Historical measurements Historical local result: 6–7/15 on val15 and 1–2/8 on the cross-validation set. |