31B WEIGHT PROGRAMMING / LIVE CAMPAIGN
31B Campaign
LLM-authored weight edits to the 31B model, screened on 120 tasks, then confirmed on a sealed set. Pushed by script from the campaign collector; this page shows the newest snapshot.
Snapshot generated 2026-11-15T02:15:40Z (UTC) · received 7 min ago · JSON · history
50 edits30 scored3 to sealed run1 passed1 confirmed80 ledger recordsChampion: 0031
Running now
b014 (GPU 1): 202 of 480; kh001 (GPU 0): 334 of 480; cf003 (sealed): 0 of 1200
| Batch / arm | Stage | Rows | Solved so far |
|---|---|---|---|
| b014 · Uneditedbase | Stage 1 | 50 of 120 | 32 |
| b014 · Unedited, second armaa | Stage 1 | 52 of 120 | 34 |
| b014 · 0031 alone (provisional base)champ | Stage 1 | 62 of 120 | 40 |
| b014 · Edit 0064edit | Stage 1 | 38 of 120 | 0 |
| kh001 · Edit special | special | 79 of 120 | 50 |
| kh001 · Edit special | special | 86 of 120 | 57 |
| kh001 · Edit special | special | 78 of 120 | 46 |
| kh001 · Edit special | special | 91 of 120 | 55 |
Author loop now
| Author | Turn | Mode | State | Since (ET) |
|---|---|---|---|---|
| claude | 18 | fresh | waiting last exit 0 | 7:55 PMended 9:04 PM |
| gpt | 17 | solo | running | 7:55 PM |
| gemini | 16 | solo | waiting last exit 0 | 7:37 PMended 9:10 PM |
| grok | 20 | solo | waiting last exit 0 | 8:10 PMended 8:44 PM |
| council | round 1 | build | 4 proposals · chair claude | 7:55 PMclaude, gpt, gemini, grok |
Every author turn and council round, with live traces →
Lanes
Edits per author and mode. Solo = the author's own history; fresh = no shared history; council = built from a council round. Cells show entries, then best gain (tasks of 120) where scored.
| Author | solo | fresh | council | Total |
|---|---|---|---|---|
| claude | 16 · best +18.5 | 2 | 0 | 18 |
| gpt | 12 · best +3.5 | 0 | 0 | 12 |
| gemini | 11 · best +8.0 | 0 | 0 | 11 |
| grok | 2 | 0 | 0 | 2 |
Plus 7 champion restacks / other entries (in the leaderboard, not in a lane).
Leaderboard
| # | Edit | Author / lane | Stage | Gain | Sealed |
|---|---|---|---|---|---|
| 1 | 0031 Instruction vector, think briefly: an always-on neuron per layer writes the stream shift of a 'keep your thinking short' paragraph (dose 2) | claudeClaude solo | Scorednull | +18.5screen +9.0 · pass 2 +9.5 | confirmed +11.0 |
| 2 | 0015 Spectral rank-4 exact-context within-call thought-close writer subspace across all 60 layers (0.38%) | geminiGemini solo | Scorednull | +8.0screen +4.5 · pass 2 +3.5 | not confirmed +3.0 |
| 3 | 0032 Instruction vector, think briefly: an always-on neuron per layer writes the stream shift of a 'keep your thinking short' paragraph (dose 4) | claudeClaude solo | Scorednull | +5.5screen +3.0 · pass 2 +2.5 | not confirmed +5.5 |
| 4 | 0044 Stateful two-neuron post-edit spec-audit boost with thought and edit-history gating over 180 MLP matrices | geminiGemini solo | Scored- | +5.5screen +5.5 | not confirmed +4.0 |
| 5 | 0050 Instruction pair vector, 'be exact about what the issue asks' part written three times in layers 6 to 24, think-briefly part at dose 2 | claudeClaude solo | Scored- | +4.0screen +4.0 | not confirmed +6.5 |
| 6 | 0048 Paired outcome writer map across all 60 layers at 0.33 percent | gptGPT solo | Scored- | +3.5screen +3.5 | not confirmed +3.5 |
| 7 | 0043 Thought-gated reliability-shrunk Problem Statement spec-audit steering across all 60 MLP layers | geminiGemini solo | Sealed run runningnull · in cf003 (sealed) | +2.5screen +2.5 | — |
| 8 | 0039 Instruction pair vector: think briefly + be exact about what the issue asks, compiled self-consistently into one neuron per layer (dose 2) | claudeClaude solo | Sealed run runningnull · in cf003 (sealed) | +2.0screen +2.0 | — |
| 9 | 0049 Think-briefly + be-exact prompt compiled where a prompt enters: many-key rank-8 down_proj maps in 7 global-attention layers, dose 2 | claudeClaude solo | Scorednull | +1.0screen +1.0 | — |
| 10 | 0016 Subtractive rank-2 deep-rumination attractor damping across all 280 mid-to-late modules (0.36%) | geminiGemini solo | Scorednull | +0.5screen +0.5 | — |
| 11 | 0062 'Wait' contrast at one third: every writer in every layer keeps two thirds of its preference for 'Wait' over other sentence starters | claudeClaude solo | Scorednull | +0.5screen +0.5 | — |
| 12 | 0007 Deep-thought gate (dose 0.1) plus the model s own end-of-thought neurons at 9 times their write | claudeClaude solo | Scorednull | +0.0screen +0.0 | — |
| 13 | 0026 Damp the code-versus-prose style contrast in writers in all layers at 0.11 percent | gptGPT solo | Scorednull | +0.0screen +0.0 | — |
| 14 | 0025 Damp the code-versus-prose style contrast in writers in all layers at 0.33 percent | gptGPT solo | Scorednull | -1.0screen -1.0 | — |
| 15 | 0056 Capacity-regularized three-phase GeGLU controller for concise thoughts, post-edit spec audit, and pre-submit veto | geminiGemini solo | Scorednull | -1.5screen -1.5 | — |
| 16 | 0027 Damp the code-versus-prose style contrast in readers in all layers at 0.33 percent | gptGPT solo | Scorednull | -2.0screen -2.0 | — |
| 17 | 0038 Think-briefly instruction vector, self-consistent compile (dose 2): each layer writes only what the paragraph's mean shift still lacks | claudeClaude solo | Scorednull | -2.0screen -2.0 | — |
| 18 | 0035 Source-grounding instruction contrast across all 60 layers at dose 3 | gptGPT solo | Scorednull | -2.5screen -2.5 | — |
| 19 | 0047 Paired outcome writer map across all 60 layers at 1 percent | gptGPT solo | Scorednull | -2.5screen -2.5 | — |
| 20 | 0061 'Wait' contrast removed from every writer: no block output in any layer can prefer 'Wait' to the other words that open a sentence | claudeClaude solo | Scorednull | -2.5screen -2.5 | — |
| 21 | 0036 Source-grounding instruction contrast across all 60 layers at dose 1 | gptGPT solo | Scorednull | -3.5screen -3.5 | — |
| 22 | 0055 OOF-shrunk rank-8 multi-key phase teachers across 108 RoPE-free o_proj and down_proj matrices | geminiGemini solo | Scorednull | -3.5screen -3.5 | — |
| 23 | 0008 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) | claudeClaude solo | Scoredharmful | -8.0screen -8.0 | — |
| 24 | 0011 Amplify natural thought-ending writes across all layers at 0.33 percent | gptGPT solo | Scoredharmful | -10.0screen -10.0 | — |
| 25 | 0028 Protocol-nullspace rank-4 exact thought-close writer subspace across L10-59 (0.48%) | geminiGemini solo | Scoredharmful | -10.5screen -10.5 | — |
| 26 | 0021 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 do | claudeClaude solo | Scoredharmful | -11.5screen -11.5 | — |
| 27 | 0063 Damp the code-versus-prose style contrast in readers across all 60 layers at 0.66 percent | gptGPT solo | Scoredharmful | -12.0screen -12.0 | — |
| 28 | 0020 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 do | claudeClaude solo | Scoredharmful | -13.5screen -13.5 | — |
| 29 | 0012 Amplify natural thought-ending writes across all layers at 0.50 percent | gptGPT solo | Scoredharmful | -25.5screen -25.5 | — |
| 30 | 0009 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) | claudeClaude solo | Scoredharmful | -71.0screen -71.0 | — |
| 31 | 0010 Early deep-thought gate (fitted on depth >= 400, dose 0.1) plus own end-of-thought neurons x9: the strong dose of the gated family | claudeClaude solo | Parked (not screened)- | — | — |
| 32 | 0013 Shift common writer output toward thought endings across all layers at 0.11 percent | gptGPT solo | Parked (not screened)- | — | — |
| 33 | 0017 Coordinate-whitened GeGLU late pre-edit commit gate across 13 high-signal layers (dose 1.20) | geminiGemini solo | Parked (not screened)- | — | — |
| 34 | 0018 Dual-cap rescue stack: rank-4 exact-close writer subspace (0.38%) plus 13-layer late pre-edit commit gate | geminiGemini solo | Parked (not screened)- | — | — |
| 35 | 0022 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 0.12 | claudeClaude solo | Parked (not screened)- | — | — |
| 36 | 0023 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 do | claudeClaude solo | Parked (not screened)- | — | — |
| 37 | 0064 Stateful two-neuron post-edit spec-audit boost with thought and edit-history gating over 180 MLP matrices (on 0031) | stackRestack on 0031 | Stage 1 running- · in b014 | — | — |
| 38 | 0065 Instruction pair vector, 'be exact about what the issue asks' part written three times in layers 6 to 24, think-briefly part at dose 2 (on 0031) | stackRestack on 0031 | Waiting for stage 1- | — | — |
| 39 | 0066 Paired outcome writer map across all 60 layers at 0.33 percent (on 0031) | stackRestack on 0031 | Waiting for stage 1- | — | — |
| 40 | 0067 Instruction vector, think briefly: an always-on neuron per layer writes the stream shift of a 'keep your thinking short' paragraph (dose 4) (on 0031) | stackRestack on 0031 | Waiting for stage 1- | — | — |
| 41 | 0068 Thought-gated reliability-shrunk Problem Statement spec-audit steering across all 60 MLP layers (on 0031) | stackRestack on 0031 | Waiting for stage 1- | — | — |
| 42 | 0069 Instruction pair vector: think briefly + be exact about what the issue asks, compiled self-consistently into one neuron per layer (dose 2) (on 0031) | stackRestack on 0031 | Waiting for stage 1- | — | — |
| 43 | 0070 Spectral rank-4 exact-context within-call thought-close writer subspace across all 60 layers (0.38%) (on 0031) | stackRestack on 0031 | Waiting for stage 1- | — | — |
| 44 | 0074 One-third tool_act plus on a fresh neuron, added on top of 0031 | grokGrok solo | Waiting for stage 1- | — | — |
| 45 | 0075 Always-on MLP bake of the tool_act plus direction at 2% of mean stream norm | grokGrok solo | Waiting for stage 1- | — | — |
| 46 | 0076 Damp four secondary writer modes while protecting leading modes and protocol readouts on unedited | gptGPT solo | Waiting for stage 1- | — | — |
| 47 | 0077 Ban the thought-paragraph openers "Wait" and "Actually" by removing their logit direction from all 120 writers (rank 2, 0.85% of the network | claudeClaude fresh | Waiting for stage 1- | — | — |
| 48 | 0078 Ban the thought-paragraph openers "Wait" and "Actually" and move half of the Wait logit onto "Let" in the 40 writers of layers 40-59, with t | claudeClaude fresh | Waiting for stage 1- | — | — |
| 49 | 0079 Tool-neutral concise-thought, Problem-Statement audit, post-edit check, and submit-veto GeGLU controller | geminiGemini solo | Waiting for stage 1- | — | — |
| 50 | 0080 Collision-free 0031-orthogonal Problem-Statement audit, post-edit boundary check, and submit-veto GeGLU controller | geminiGemini solo | Waiting for stage 1- | — | — |
Sealed confirmation
Counts only. Sealed task identities are never published.
Batch cf003 (running)0 of 1200 rowsEdits 0043 + 0039Bars 0043 +8.0, 0039 +8.0
part 1 on boxes: 0 of 600; part 2: not started yet
| Edit | Sealed batch | Solved: edit / unedited / identical arm | Gain | Result |
|---|---|---|---|---|
| 0031claude | cf001 | 181 / 173 / 167of 300 tasks | +11.0 | confirmed |
| 0015gemini | cf001 | 173 / 173 / 167of 300 tasks | +3.0 | not confirmed |
| 0032claude | cf002 | 178 / 170 / 175of 300 tasks | +5.5 | not confirmed |
| 0044gemini | cf001 | 174 / 173 / 167of 300 tasks | +4.0 | not confirmed |
| 0050claude | cf002 | 179 / 170 / 175of 300 tasks | +6.5 | not confirmed |
| 0048gpt | cf002 | 176 / 170 / 175of 300 tasks | +3.5 | not confirmed |
Bars
| Screen gain (run 1) to go on to the sealed run | 2.0 of 120 tasks |
| Screen gain (run 1) that counts as harmful | -6.5 of 120 tasks |
| Stage 2 gain needed (record 0057, retired by 0060) | 0.5 of 218 tasks |
| Stage 2 gain over the control needed (record 0057, retired by 0060) | 0.5 of 218 tasks |
| Total gain needed to pass (record 0057, retired by 0060) | 8.0 of 338 tasks |
| Total gain that counts as harmful (record 0057, retired by 0060) | -10.5 of 338 tasks |
| Sealed-run gain to promote (record 0060) | 8.0 of 300 tasks |
| Sealed-run gain to promote, entries already past the 0057 screen | 5.0 of 300 tasks |
Ledger records (80)
| Id | Date | Author | Record | Verdict |
|---|---|---|---|---|
| 0080 | 2026-11-14 | gemini | Collision-free 0031-orthogonal Problem-Statement audit, post-edit boundary check, and submit-veto GeGLU controller | Nonequeued |
| 0079 | 2026-11-14 | gemini | Tool-neutral concise-thought, Problem-Statement audit, post-edit check, and submit-veto GeGLU controller | Nonequeued |
| 0078 | 2026-11-14 | claude | Ban the thought-paragraph openers "Wait" and "Actually" and move half of the Wait logit onto "Let" in the 40 writers of layers 40-59, with t | Nonequeued |
| 0077 | 2026-11-14 | claude | Ban the thought-paragraph openers "Wait" and "Actually" by removing their logit direction from all 120 writers (rank 2, 0.85% of the network | Nonequeued |
| 0076 | 2026-11-14 | gpt | Damp four secondary writer modes while protecting leading modes and protocol readouts on unedited | Nonequeued |
| 0075 | 2026-11-14 | grok | Always-on MLP bake of the tool_act plus direction at 2% of mean stream norm | Nonequeued |
| 0074 | 2026-11-14 | grok | One-third tool_act plus on a fresh neuron, added on top of 0031 | Nonequeued |
| 0073 | 2026-11-14 | claude | Author loop - fresh turns, the council lane and a Grok author seatprotocol; queue entries now carry mode solo / fresh / council in <entry>/mode.json and a mode line in the recipe | n/adone |
| 0072 | 2026-11-14 | claude | Kaggle-conditions 2x2 - unedited and 0031 under today's harness and the 0.18 harness (batch kh001)pending; counts only; outside the pools, no promotion or drop from it | n/arunning |
| 0071 | 2026-11-14 | claude | Pooled reference arms - stage-1 batches become 1 reference arm + 3 edits once a pool holds 4 reference armsprotocol; gain = edit minus the pooled reference mean (same serving group, generation, task set, budget); 2+2 while a pool has fewer than 4 arms; drift over 8 tasks scores against the own arm only | n/adone |
| 0070 | 2026-11-14 | gemini | Spectral rank-4 exact-context within-call thought-close writer subspace across all 60 layers (0.38%) - re-entered on champion generation 1 (0031) | Nonequeued |
| 0069 | 2026-11-14 | claude | Instruction pair vector: think briefly + be exact about what the issue asks, compiled self-consistently into one neuron per layer (dose 2) - re-entered on champion generation 1 (0031) | Nonequeued |
| 0068 | 2026-11-14 | gemini | Thought-gated reliability-shrunk Problem Statement spec-audit steering across all 60 MLP layers - re-entered on champion generation 1 (0031) | Nonequeued |
| 0067 | 2026-11-14 | claude | Instruction vector, think briefly: an always-on neuron per layer writes the stream shift of a 'keep your thinking short' paragraph (dose 4) - re-entered on champion generation 1 (0031) | Nonequeued |
| 0066 | 2026-11-14 | gpt | Paired outcome writer map across all 60 layers at 0.33 percent - re-entered on champion generation 1 (0031) | Nonequeued |
| 0065 | 2026-11-14 | claude | Instruction pair vector, 'be exact about what the issue asks' part written three times in layers 6 to 24, think-briefly part at dose 2 - re-entered on champion generation 1 (0031) | Nonequeued |
| 0064 | 2026-11-14 | gemini | Stateful two-neuron post-edit spec-audit boost with thought and edit-history gating over 180 MLP matrices - re-entered on champion generation 1 (0031) | Nonerunning |
| 0063 | 2026-11-14 | gpt | Damp the code-versus-prose style contrast in readers across all 60 layers at 0.66 percentstage 1: -12.0 of 120 tasks | harmfuldone |
| 0062 | 2026-11-14 | claude | 'Wait' contrast at one third: every writer in every layer keeps two thirds of its preference for 'Wait' over other sentence startersstage 1: +0.5 of 120 tasks | nulldone |
| 0061 | 2026-11-14 | claude | 'Wait' contrast removed from every writer: no block output in any layer can prefer 'Wait' to the other words that open a sentencestage 1: -2.5 of 120 tasks | nulldone |
| 0060 | 2026-11-14 | claude | Two-run rule - a 120-task quick screen, then one sealed 300-task run on the boxes that decides (supersedes steps 1-3 of record 0057 for entries not yet judged)protocol; run 1 as before (harmful at -6.5 or below, null below +2.0); run 2 = the 300 sealed tasks on the boxes, two unedited arms shared by up to three edits, no control, promote at gain +8.0 or more (1.6 SE, about 5% false pass per candidate); entries already past 0057's screen keep their +5.0 bar | n/adone |
| 0059 | 2026-11-14 | claude | fast2 deploy - fp8 KV cache, 48 streams and the rest of Son's serving stack on the rented boxes and on wrx90 (changes the numbers; new comparable group)deployed without a pass bar at Keith's request; lane's replay measurement 430.0 tok/s against 314.3 for fast1 (+37 %); not yet verified in a real batch | n/arunning |
| 0058 | 2026-11-14 | claude | Sealed confirmation run of edit 0031 on [sealed set] at the 20,000-token budget (put back by hand under record 0057)confirmed: +11.0 of 300 sealed tasks (bar +5), solved 181 against unedited 173 and 167 | confirmeddone |
| 0057 | 2026-11-14 | claude | One-stage judging rule - screen in multiples of the measured noise s, one sealed confirmation run before any promotion, golden V2 reported onlyprotocol; pass-1 bar +2.0 (was +3.0), harmful -6.5 (was -6.0); screen pass needs pooled +8.0 over 338 tasks with stage-2 gain above 0 and edit above control (was +10 / +4 / +5); near misses (+4.0 to +7.5) get one more 218-task pass; a screen pass goes to one sealed confirmation run and is promoted only if it gains at least 1 SE there and beats its control | n/adone |
| 0056 | 2026-11-14 | gemini | Capacity-regularized three-phase GeGLU controller for concise thoughts, post-edit spec audit, and pre-submit vetostage 1: -1.5 of 120 tasks | nulldone |
| 0055 | 2026-11-14 | gemini | OOF-shrunk rank-8 multi-key phase teachers across 108 RoPE-free o_proj and down_proj matricesstage 1: -3.5 of 120 tasks | nulldone |
| 0054 | 2026-11-14 | claude | fast1 deploy - 24 streams per server, GPU memory share 0.95 and Son's adaptive ngram drafts on the rented boxes and on wrx90deployed without a pass bar at Keith's request; lane's single-window measurement 317.0 tok/s against a mean of 278.6 (+13.8 %); not yet verified in a real batch | n/arunning |
| 0053 | 2026-11-14 | claude | Golden batch g002 - the comp half of golden v2 (106 tasks) for the unedited 31B model and edit 0031, so that g001 plus g002 is the full golden v2does not tell; g002 (comp half): base 44, aa 37, e0031 44, c0031 41 of 106 (0031 +3.5, control +0.5); golden v2 (g001 + g002): unedited mean 70.5, e0031 76, control 71 of 169 (0031 +5.5); valid by the coordinator's ruling | inconclusivedone |
| 0052 | 2026-11-14 | claude | Golden look g001 - the unedited 31B model and edit 0031 on the 63 held-out golden tasks (first golden run of the program)does not tell (rebench half of golden v2 only: 63 of its 64 rebench tasks at 10,000 tokens); solved of 63: base 30, aa 30, e0031 32, c0031 30; 0031 +2.0 over the unedited mean and +2 over its control; valid; all four predictions held | inconclusivedone |
| 0051 | 2026-11-14 | claude | Noise batch x007 - four identical unedited arms on the boxes, to measure pass-to-pass spread and the arm-order effectnot yet run; written before the batch was submitted | n/arunning |
| 0050 | 2026-11-14 | claude | Instruction pair vector, 'be exact about what the issue asks' part written three times in layers 6 to 24, think-briefly part at dose 2 | not confirmeddone |
| 0049 | 2026-11-14 | claude | Think-briefly + be-exact prompt compiled where a prompt enters: many-key rank-8 down_proj maps in 7 global-attention layers, dose 2stage 1: +1.0 of 120 tasks | nulldone |
| 0048 | 2026-11-14 | gpt | Paired outcome writer map across all 60 layers at 0.33 percent | not confirmeddone |
| 0047 | 2026-11-14 | gpt | Paired outcome writer map across all 60 layers at 1 percentstage 1: -2.5 of 120 tasks | nulldone |
| 0046 | 2026-11-14 | claude | Correction of batch b009: 2 stage-2 task(s) run in their own images (b009-supp1)corrected solved: base 127 to 127, e0032 123 to 123, c0032 120 to 120, aa 114 to 114 (of 218); verdict changes: none | n/adone |
| 0045 | 2026-11-14 | claude | Correction of batch b008: 2 stage-2 task(s) run in their own images (b008-supp1)corrected solved: base 114 to 113, e0031 123 to 122, c0031 119 to 120, aa 111 to 112 (of 218); verdict changes: none | n/adone |
| 0044 | 2026-11-14 | gemini | Stateful two-neuron post-edit spec-audit boost with thought and edit-history gating over 180 MLP matrices | not confirmeddone |
| 0043 | 2026-11-14 | gemini | Thought-gated reliability-shrunk Problem Statement spec-audit steering across all 60 MLP layersstage 1: +2.5 of 120 tasks | nulldone |
| 0042 | 2026-11-14 | claude | Correction of batch b004: 19 stage-1 task(s) run in their own images (b004-supp1)corrected solved: base 56 to 70, e0021 50 to 60, e0020 47 to 58, aa 59 to 73 (of 120); verdict changes: none | n/adone |
| 0041 | 2026-11-14 | claude | Correction of batch b002: 31 stage-1 task(s) run in their own images (b002-supp1)corrected solved: base 60 to 73, e0008 51 to 67, e0011 51 to 65, aa 62 to 77 (of 120); verdict changes: none | n/adone |
| 0040 | 2026-11-14 | claude | Episode budget test - 20,000 generated tokens solves more of the cut-off tasks than 10,000; the 40,000 arm did not run validlyprotocol; on the 67 stage-1 tasks whose unedited episode was cut off, 20,000 tokens solved 27 against 22 at 10,000 (gained 8, lost 3, +4.2 points of the 120-task set), which meets the rule fixed before the run; the budget becomes 20,000 for entries whose stage 1 starts after the change is deployed | n/adone |
| 0039 | 2026-11-14 | claude | Instruction pair vector: think briefly + be exact about what the issue asks, compiled self-consistently into one neuron per layer (dose 2)stage 1: +2.0 of 120 tasks | nulldone |
| 0038 | 2026-11-14 | claude | Think-briefly instruction vector, self-consistent compile (dose 2): each layer writes only what the paragraph's mean shift still lacksstage 1: -2.0 of 120 tasks | nulldone |
| 0037 | 2026-11-14 | claude | Entries parked on 13 November for lack of GPUs are put back by the coordinator for box screeningprotocol; entries parked on 13 November go back into stage 1 unchanged where a box batch has room, at most two per author at a time, put back by the coordinator and not by their authors | n/adone |
| 0036 | 2026-11-14 | gpt | Source-grounding instruction contrast across all 60 layers at dose 1stage 1: -3.5 of 120 tasks | nulldone |
| 0035 | 2026-11-14 | gpt | Source-grounding instruction contrast across all 60 layers at dose 3stage 1: -2.5 of 120 tasks | nulldone |
| 0034 | 2026-11-13 | claude | Stage-1 batches on rented boxes, when their scores count, and the one platform checka box batch is a stage-1 batch scored against its own two reference arms; one platform check of x001's reference arms against the first wrx90 zero-adapter stage-1 batch (limit 8 tasks), not yet run | n/adone |
| 0033 | 2026-11-13 | claude | Reference arms served through an all-zero adapter (the path check of record 0003 fails on the corrected rows)base and aa both served through the zero adapter from the next batch; bars unchanged; first live batch not yet run | n/adone |
| 0032 | 2026-11-13 | claude | Instruction vector, think briefly: an always-on neuron per layer writes the stream shift of a 'keep your thinking short' paragraph (dose 4)both stages: +5.5 (stage 1 +3.0, stage 2 +2.5; control -0.5 on stage 2) | nulldone |
| 0031 | 2026-11-13 | claude | Instruction vector, think briefly: an always-on neuron per layer writes the stream shift of a 'keep your thinking short' paragraph (dose 2)both stages: +19.5 (stage 1 +9.0, stage 2 +10.5; control +6.5 on stage 2) | nulldone |
| 0030 | 2026-11-13 | claude | Correction of batch b001: 31 stage-1 task(s) run in their own images (b001-supp1)corrected solved: base 60 to 74, aa 60 to 74, zero 65 to 81 (of 120); verdict changes: none | n/adone |
| 0029 | 2026-11-13 | claude | Correction of batch b003: 31 stage-1 task(s) run in their own images (b003-supp1)corrected solved: base 58 to 72, e0012 34 to 44, e0015 61 to 74, aa 52 to 67 (of 120); verdict changes: none | n/adone |
| 0028 | 2026-11-13 | gemini | Protocol-nullspace rank-4 exact thought-close writer subspace across L10-59 (0.48%)stage 1: -10.5 of 120 tasks | harmfuldone |
| 0027 | 2026-11-13 | gpt | Damp the code-versus-prose style contrast in readers in all layers at 0.33 percentstage 1: -2.0 of 120 tasks | nulldone |
| 0026 | 2026-11-13 | gpt | Damp the code-versus-prose style contrast in writers in all layers at 0.11 percentstage 1: +0.0 of 120 tasks | nulldone |
| 0025 | 2026-11-13 | gpt | Damp the code-versus-prose style contrast in writers in all layers at 0.33 percentstage 1: -1.0 of 120 tasks | nulldone |
| 0024 | 2026-11-13 | claude | First real run of manykey.py (closed-form many-key down_proj solve over layers 34-39), measured against its predictionwrx90 GPU 0 forward pass, 548 positions of 54 stage-1 tasks; at the first band layer the measured move equals the one-layer prediction (+0.0789 vs +0.0784 of the size of mlp_out along d, n=20 solve keys; +0.0226 vs +0.0226, n=10 held-out keys); later band layers move 1.1x to 1.6x more along d than predicted; closeprobe log p(close) at cont +0.157 (control +0.023), no top token changed | Nonedone |
| 0023 | 2026-11-13 | claude | 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 do | Noneparked |
| 0022 | 2026-11-13 | claude | 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 0.12 | Noneparked |
| 0021 | 2026-11-13 | claude | 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 dostage 1: -7.5 of 102 tasks | harmfuldone |
| 0020 | 2026-11-13 | claude | 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 dostage 1: -10.5 of 102 tasks | harmfuldone |
| 0019 | 2026-11-13 | claude | Generation speed against the server's LoRA rank limit, and per-stream speed in mixed batches (wrx90)wrx90 GPU 0, 8 streams, n=3 - same rank-32 adapter costs 12% / 13% / 18% of generation speed at rank limit 32 / 64 / 128; rank 96 costs the same as rank 32 at limit 128; unedited streams of a mixed batch are slowed exactly as much as adapter streams | Nonedone |
| 0018 | 2026-11-12 | gemini | Dual-cap rescue stack: rank-4 exact-close writer subspace (0.38%) plus 13-layer late pre-edit commit gate | Noneparked |
| 0017 | 2026-11-12 | gemini | Coordinate-whitened GeGLU late pre-edit commit gate across 13 high-signal layers (dose 1.20) | Noneparked |
| 0016 | 2026-11-12 | gemini | Subtractive rank-2 deep-rumination attractor damping across all 280 mid-to-late modules (0.36%)stage 1: +0.5 of 120 tasks | nulldone |
| 0015 | 2026-11-12 | gemini | Spectral rank-4 exact-context within-call thought-close writer subspace across all 60 layers (0.38%)both stages: +8.0 (stage 1 +4.5, stage 2 +3.5; control -3.5 on stage 2) | nulldone |
| 0014 | 2026-11-12 | claude | 31B rollout serving speed at fixed quality (server flags, concurrency, LoRA serving cost)rented card, runner batch (4 reference + 2 + 2 edit streams) -19.6% against base only; slots no effect, rank limit 64/128 -5%/-13%; 16-row LoRA expand tile bit-exact at kernel level, gates within self-noise, +5% paired on the runner batch (deploy block DEPLOY.md) | Nonerunning |
| 0013 | 2026-11-12 | gpt | Shift common writer output toward thought endings across all layers at 0.11 percent | Noneparked |
| 0012 | 2026-11-12 | gpt | Amplify natural thought-ending writes across all layers at 0.50 percentstage 1: -21.0 of 120 tasks | harmfuldone |
| 0011 | 2026-11-12 | gpt | Amplify natural thought-ending writes across all layers at 0.33 percentstage 1: -10.0 of 120 tasks | harmfuldone |
| 0010 | 2026-11-12 | claude | Early deep-thought gate (fitted on depth >= 400, dose 0.1) plus own end-of-thought neurons x9: the strong dose of the gated family | Noneparked |
| 0009 | 2026-11-12 | claude | 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)stage 1: -71.0 of 120 tasks | harmfuldone |
| 0008 | 2026-11-12 | claude | 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)stage 1: -10.0 of 120 tasks | harmfuldone |
| 0007 | 2026-11-12 | claude | Deep-thought gate (dose 0.1) plus the model s own end-of-thought neurons at 9 times their writestage 1: +0.0 of 120 tasks | nulldone |
| 0006 | 2026-11-12 | claude | Interpretability lane (can LLMs read the weights; failure modes; per-layer directions for training-sized edits)research/wsp31/interp/REPORT.md | Nonerunning |
| 0005 | 2026-11-12 | claude | Tools lane (31B edit toolchain: exact loader and forward pass, LoRA builders, checker, pre-tests, champion support)research/wsp31/edit_tools/REPORT.md | Nonerunning |
| 0004 | 2026-11-12 | claude | How far real training moves the weights (size reference for edits)research/wsp31/tools/train_size.py (numbers below) | n/adone |
| 0003 | 2026-11-12 | claude | Three identical-weight arms in one batch (base rate, noise floor, adapter-path check, runner shakedown)base 60, aa 60, zero 65 of 120; d 11.3 tasks per identical pair; bars +3 / -6 / +4 / +5 / +10 / -10 | n/adone |
| 0002 | 2026-11-12 | claude | Survey of training-free weight editing to November 2026, with 20 ranked edit ideas for 31Breports/2026-11-12-weight-editing-survey.md | n/adone |
| 0001 | 2026-11-12 | claude | Unedited 31B on the screening set, two passes (base rate and identical-arm noise)stopped at 59 of 720 episodes; 54% ended at the session limit at concurrency 16 | inconclusivedone |