training / docs

Commit History

Today

  1. RESEARCH_LOG: maxk2 final verdict + the three-arm 100K arc closes (mechanisms measured, escalation is the operator's; 1700-rung is the one cross-scale gain) Running

    1a4dbbd
  2. RESEARCH_LOG: maxk2 mid-run correction - class-MAE is scale-floor not collapse-cause; entropy-holding is the decisive axis; post-500 window is the final test

    751c94c

Yesterday

  1. RESEARCH_LOG: maxk1 mid-run arc (instruments, loss-mode shift, staged machinery) at cycle 50

    fca5627
  2. RESEARCH_LOG: maxk1 SF row @199 - 0.188@1320 (3D/8, best ever), first absolute-axis signal under the value-repair

    0d90ea0
  3. RESEARCH_LOG: value-repair trial arc (H-conv-11v works then narrows under entropy collapse; gate decides branch) + escalation shipped default-off + the bracket

    feab628
  4. RESEARCH_LOG: benchwave complete (depth lever revived 0.925@st16, direct h2h 0.675 over prod_v1, king analyses, value blindness) + the new-mandate maxk pivot record

    519ec63
  5. RESEARCH_LOG: value-head blindness measured (danger == equal on the critic; zero losing reads) - third layer of the closed-universe chain complete; phase-choice re-weighting toward reward-shaping/SF-data hybrids

    2fc17d0
  6. RESEARCH_LOG: K3 verdict - in-constraint space CLOSED (rights+punishment still 0/40 castling); four-arm falsification chain complete; escalation provably necessary

    bbec425
  7. RESEARCH_LOG: K2 verdict + pure-self-play class CLOSED (three-arm falsification complete: inputs inert, outcome-curriculum touches resistance not the mechanism, castling needs rights in data); escalation options filed

    86d701e
  8. RESEARCH_LOG: K1 verdict - integration FAIL (ablation: 0.000 deltas; policy ignores exact king channels); C-vs-K mechanism proven (gradient-through-outcome is the missing link); next-lever class = training-process changes (operator's call)

    e58026d
  9. RESEARCH_LOG: absolute-axis diagnosis - king safety absent from policy (0/70 castled, mated at parity); evidenced next intervention: king-safety input channels

    2ee4ac8
  10. RESEARCH_LOG: C1+C2 final verdict - shape (2): tactics converted (probe 0.85 vs 0.663 ceiling), volume axis re-opened internally, absolute SF flat (0.052 vs 0.047); the next question: why 85% gift-taking still loses every SF game

    6f7019a
  11. RESEARCH_LOG: C1 mid-run record (eval band + probe preview + gate machinery) - verdict section to follow at [done]

    9a79974
  12. RESEARCH_LOG: s8-jitter arc complete (eval-mode integrity incident + fix + verdict: free-upgrade default, extrap null) and the C-launch record

    b8acd9f
  13. RESEARCH_LOG: prod_v1 complete - the flat-surface finding (absolute-scale neutrality across volume/objective/train-depth/infer-depth on final), the decoupling decode, and the eliminated-lever table

    agent
    cef9a1f

2 days ago

  1. RESEARCH_LOG: prod_v1 quarter-findings - validated dip-recovery model, ~982@110 absolute anchor, capture-flip move-level analysis, SF/ops lessons

    agent
    7735a06
  2. RESEARCH_LOG: GAE 2x2 sweep verdict - incumbent recipe (st4+MC) wins; lambda late-sag cross-validated at both depths with KL signature; depth-20 training no gain at 40 iters

    agent
    45e2f66
  3. RESEARCH_LOG: gemma decoder v6 verdict (clean null) + engineering facts (sqrt640, zero-row NaN, capture mirror)

    agent
    a65080d
  4. RESEARCH_LOG: week of 2026-11-18 - fast_wide verdict, SF-Elo anchor (~775), test_web v_mae mixture decode, TD(lambda) implementation, free-capture probe

    agent
    0b1de74

3 days ago

  1. RESEARCH_LOG: family ablation verdict (transformer wins both seeds) + param-matched follow-up design

    acb8595
  2. Phase 12: static weight analysis of iter269 - local op undifferentiated, step_emb zero row explains depth collapse, rank4-5 attention dead zone

    f03ec25

  1. log: retract v7-v12 KL conclusion; metric artifact + logp_floor fix

    Verified signature
    9d5bc84
  2. log: repetition fix necessary but not sufficient

    Verified signature
    7e47006
  3. log: v7-v12 stability campaign, 3 bugs fixed, no strength gain

    Verified signature
    f9afc8e

  1. log: repetition-feature bug, greedy baseline, learn bench

    Verified signature
    f14f327
  2. fix eval-on-resume (silently skipped); add learn bench; v4 analysis

    Verified signature
    24c12cb
  3. add learn-phase thread benchmark

    Verified signature
    3f4a221
  4. revert logp_old change (ablation-falsified); keep only_player + kl estimator fixes

    Verified signature
    5d627db
  5. draw-penalty: remove guaranteed-0 repetition optimum

    Verified signature
    35dde2e
  6. parallel self-play (--workers), conversion experiments, whitelist

    Verified signature
    b97f43c

  1. whitelist analysis scripts; fix pid-recycling in launch/job_table

    Verified signature
    2f4aa7f
  2. research log: env correction + thread bench

    Verified signature
    83d8bce

  1. tinychess: full research artifact - self-play RL, ablations, compute scaling, faithfulness, precision, plasticity

    Verified signature
    227f251