Non-Record: Ouroboros — Crawler Architecture Research (1.1364 BPB)#1308

Open

newjordan wants to merge 412 commits intoopenai:mainfrom

newjordan:submission/ouroboros-research

This pull request is big! We're only showing the most recent 250 commits

Commits on Mar 30, 2026

Bandit_Wagon: add run_ablations.sh — BW-01..04 back-to-back at 350s, warmdown=0

Octavian
and
claude
committed
Bandit_Wagon: run_ablations.sh default NPROC=1 (single GPU signal)

Octavian
and
claude
committed
Bandit_Wagon: fix run_ablations.sh env var passing (use env)

Octavian
and
claude
committed
Bandit_Wagon: add 1-GPU winddown wrapper
Octavian
committed
pod_setup.sh: download tokenizer (fineweb_1024_bpe.model) in step 6

Octavian
and
claude
committed
Bandit_Wagon: run_ablations.sh step-based stopping (ABLATION_STEPS=500)

Octavian
and
claude
committed
Add fresh pod bootstrap and single-H100 signal runners
Octavian
committed
Make Rascal runners use portable torchrun default
Octavian
committed
Add Bandit_wagon_5f_ablations: 4F vs 5F direct proxy comparison

Octavian
and
claude
committed
Add Rascal_Turbo race-ready TurboMuon-only variant
Octavian
committed
Bandit_wagon_5f_ablations: 4F+1C confirmed optimal, 5F hypothesis denied

Octavian
and
claude
committed
Add bandit_wagon_XSA: XSA coverage sweep on confirmed 4F+1C config

Octavian
and
claude
committed
Bandit_Wagon: add 8xH100 launcher and checkpoint arch autodetect
Octavian
committed
Add bandit_wagon_crawler_mlp: crawler MLP leaky slope sweep

Octavian
and
claude
committed
bandit_wagon_XSA: XSA=15 (full coverage) wins on BPB AND speed

Octavian
and
claude
committed
Rascal_Turbo: single run.py launcher
Octavian
committed
bandit_wagon_choke: per-loop bottleneck choke sweep for crawler MLP

Octavian
and
claude
committed
bandit_wagon_smear: LoopSmearGate — depth error damping between crawler loops

Octavian
and
claude
committed
Add single-H100 Rascal ablation matrix runner
Octavian
committed
bandit_wagon_tap: per-loop gated encoder tap sweep (7 arms)

Octavian
and
claude
committed
Add sparse skip-gram ngram ablation for single-H100 Rascal
Octavian
committed
Revert unvalidated sparse skip-gram integration from Rascal runner path
Octavian
committed
bandit_wagon_battery: per-loop RoPE scale sweep + mega ablation runner

Octavian
and
claude
committed
Add isolated sparse skip-gram ablation (2200-step single-GPU)
Octavian
committed

Commits on Mar 31, 2026

bandit_wagon_crawler_mlp: log BW3 results — slope insensitive, stay at 0.5

Octavian
and
claude
committed
bandit_wagon_battery: fix MLP.forward to accept optional loop_idx

Octavian
and
claude
committed
Remove ngram sparse ablation files; keep Rascal path ngram-free
Octavian
committed
Add stripped Rascal skip-gram 2200-step calibration runner
Octavian
committed
Add bandit_wagon_choke_shaped experiment (BWCS series)

Octavian
and
claude
committed
Log 2026-03-31 single-H100 RASCAL ablation matrix results
Octavian
committed
Add next calibrated single-GPU RASCAL ablation pack
Octavian
committed
Crawler Leg3 README: add architecture philosophy

Octavian
and
claude
committed
Crawler Leg3 README: add active ablation work section

Octavian
and
claude
committed
Log skip-gram calibration seed444 results
Octavian
committed
Add BWCB ablation: battery scales on pyramid-512 choke

Octavian
and
claude
committed
Add loader-refine single-GPU ablation pack and notes
Octavian
committed
Add BWCD ablation: descending battery on pyramid-512

Octavian
and
claude
committed
BWCD: add BWCD-03 wide-medium-wide bracket (9,3,9)

Octavian
and
claude
committed
Record mega ablation results in BWB HYPOTHESIS.md

Octavian
and
claude
committed
Add race-ready Rascal final submission package with loader_cache4 launcher
Octavian
committed
BWCB results: ascending battery hurts pyramid, all configs worse

Octavian
and
claude
committed
BWCB Run B (4 shards): 1,2,4+pyramid beats pyramid alone by -0.00210

Octavian
and
claude
committed
Enforce FA3 preflight and CUDA runtime path in final Rascal launcher
Octavian
committed
BWCD results: 9,1,1+pyramid wins at -0.01193 vs pyramid alone

Octavian
and
claude
committed
BWCD complete: BWCD-03 (9,3,9) final — quant_gap +0.0062, worst of group

Octavian
and
claude
committed
Add Bandit_Wagon_III: pyramid-512 + 9,1,1 battery production runner

Octavian
and
claude
committed
Rascal: record 2026-03-31 TTT sweep regression (seed 444)
Octavian
committed
Add bandit_wagon_cannon (BWE): per-loop output calibration ablation

Octavian
and
claude
committed
Log Run C results: 1-shard pod with different val data — not directly comparable

Octavian
and
claude
committed
Rascal: add rascal_master config copies
Octavian
committed
BW3 run.sh: clean competition runner with preflight guards

Octavian
and
claude
committed
Rascal_Master: SOTA-exact race script — fix COPRIME_MAX_LOADED_SHARDS default to 1

Octavian
and
claude
committed
BW3 seed=444: 1.20684 int6_sw_bpb, 10.07MB — pyramid-512 + 9,1,1 results

Octavian
and
claude
committed
BW3 run.sh: auto-save checkpoint after run

Octavian
and
claude
committed
Add Bandit_Wagon_IV: 9,1,1 battery without pyramid choke

Octavian
and
claude
committed
BW4 seed=444: 1.18731 int6_sw_bpb — beats Leg 3 SOTA with battery only

Octavian
and
claude
committed
Bandit cannon: log seed444 proxy results (TTT bust)
Octavian
committed
BW4: add gate_fullgraph.sh — Tier 1 COMPILE_FULLGRAPH=1 test

Octavian
and
claude
committed
BW4 gate_fullgraph: fix broken ablation — revert TORCHDYNAMO_OPTIMIZE_DDP to 0

Octavian
and
claude
committed
Add Bandit_Wagon_V: BW4 + COMPILE_FULLGRAPH=1 (Tier 1 speed win)

Octavian
and
claude
committed
Add master progress checklist

Octavian
and
claude
committed
BW5 seed=444 results: 1.18672 int6_sw_bpb

Octavian
and
claude
committed
Bandit_Wagon_V_Cannon: single GPU cannon gate on BW5 base

Octavian
and
claude
committed
Add QK_SLOT_Ablation: single-GPU cross-correlation harness

Octavian
and
claude
committed
Lab cleanup: archive old BW experiments, add LAB_PROTOCOL.md

Octavian
and
claude
committed
BW5 seed=300: 1.18758 — does not individually confirm

Octavian
and
claude
committed
Add one-shot 8x quick AB runner for Rascal GPTQ stream vs insta
Octavian
committed
Make 8x quick AB runner self-contained with FA3 preflight and no rg dependency
Octavian
committed
Lock Rascal baseline launcher to record trainer and add one-shot baseline script
Octavian
committed
Bandit_Wagon_V_Cannon: cannon gate results — does not promote

Octavian
and
claude
committed
Add one-shot cu124 baseline runner that reuses custom FA3 module
Octavian
committed
Make cu124+custom-FA3 runner auto-detect non-venv base python
Octavian
committed
Harden cu124 custom-FA3 runner with python-path and filesystem auto-discovery
Octavian
committed
Extend custom FA3 detection to conda env pythons and .so module files
Octavian
committed
Add --no-deps FA3 wheel fallback to cu124 baseline runner
Octavian
committed
Add pyramid/cannon gate scripts — 1gpu + 8gpu for each hypothesis

Octavian
and
claude
committed
Add QK_SLOT_Ablation STATUS.md — current position log

Octavian
and
claude
committed
Fix gate scripts: stop swallowing output with command substitution

Octavian
and
claude
committed
Add sota_now.sh — clean single-file cu124 baseline runner

Octavian
and
claude
committed
Fix sota_now.sh to use real submission file; vault the correct source

Octavian
and
claude
committed
Relax CUDA check to 12.x — cu128 pod is valid

Octavian
and
claude
committed
Fix pod_setup.sh: auto-detect WORKSPACE from script location

Octavian
and
claude
committed
Fix inverted awk stack parity check in sota_now.sh

Octavian
and
claude
committed
Quarantine racecar lab confusion; fix records/ with vault file

Octavian
and
claude
committed
Fix pod_setup.sh: dataset shard count crash under set -euo pipefail

Octavian
and
claude
committed
Add two-track lab structure: neural/ and crawler/

Octavian
and
claude
committed
BWVC 8GPU gate results: scalar cannon passes speed gate

Octavian
and
claude
committed
Add RESULTS.md stub for Bandit_Wagon_V_Pyramid

Octavian
and
claude
committed
Junkyard: move all legacy/inactive dirs off root surface

Octavian
and
claude
committed
Add folder-based CLAUDE.md agent protocols

Octavian
and
claude
committed
BWVP 1GPU gate results: pyramid STRONG PASS on quality

Octavian
and
claude
committed
Move active BW5 gates into crawler/ track; fix train_gpt.py symlink

Octavian
and
claude
committed
Fix PyramidCannon gate_1gpu.sh usage comment path

Octavian
and
claude
committed
Add H→A→R cycle to protocol; scaffold all 3 files in new_leg.sh

Octavian
and
claude
committed
Add QK_GAIN_SLOT_Gate ablation experiment

Octavian
and
claude
committed
Add per-track science boards + auto-stub in new_leg.sh

Octavian
and
claude
committed
Add submissions/ PR zone — validate script, protocol, templates

Octavian
and
claude
committed
Move pod_setup.sh to scripts/ — accessible from repo root

Octavian
and
claude
committed
BWVPC 1GPU gate: pyramid+cannon PASSES

Octavian
and
claude
committed
Fix smoke test threshold to scale with nproc

Octavian
and
claude
committed
Fix gate_8gpu.sh: update path comment and step_avg pass criterion

Octavian
and
claude
committed
Record smoke test result: 739ms/step is healthy on 1xH100

Octavian
and
claude
committed
BWVPC 8GPU gate: DOES NOT PROMOTE — pyramid+cannon fails

Octavian
and
claude
committed
Update SCIENCE.md: close pyramid/cannon screw, reflect actual completed results
Octavian
committed
Two crawler legs: BW5_Cannon full run + BW6_Skipgram gate
Octavian
committed
Add BW6_Skipgram gate_8gpu.sh — 2000-step 8xH100 A/B gate
Octavian
committed
Add PIPELINE.md — full ranked hypothesis queue, both tracks
Octavian
committed
BW5_Cannon full run: DOES NOT PROMOTE — +0.00020 vs BW5 at 600s/8034 steps
Octavian
committed
Housekeeping: archive closed crawler legs, move QK_Gain_SLOT to neural track
Octavian
committed
BW6_Skipgram gate: null result — trigram neutral on crawler, −140KB size artifact
Octavian
committed
Archive BW6_Skipgram — null result, closed
Octavian
committed
Update PIPELINE.md: close cannon/skipgram, fix paths, update neural trigram note
Octavian
committed
Fix SLOT backward crash + record run 1 results

Octavian
and
claude
committed
Add BW7 MegaGate: 8-arm ablation on 4xGPU

Octavian
and
claude
committed
Add BW7 MegaGate pod_setup.sh — fresh pod one-shot launcher

Octavian
and
claude
committed
Relax flash_attn preflight in BW7 MegaGate — warn not abort

Octavian
and
claude
committed
Fix torchrun path: fallback to python3 -m torch.distributed.run

Octavian
and
claude
committed
Update neural/SCIENCE.md with competitive intelligence + hypothesis roadmap

Octavian
and
claude
committed
Add Arch+Sched Sweep: 6-case 4×GPU ablation (rope_32, bigram_4096, qat_{early,late}, swa_dense)

Octavian
and
claude
committed
Expand sweep to 9 cases: add gptq, bigram_3072, warmdown_4k

Octavian
and
claude
committed

Commits on Apr 1, 2026

Sweep: gptq case reuses baseline checkpoint (SKIP_TRAIN=1 + LOAD_CHECKPOINT)

Octavian
and
claude
committed
Add SLOT legality analysis to neural/SCIENCE.md

Octavian
and
claude
committed
Add BW8_Tap: shared encoder tap dim=32 — strongest MegaGate signal

Octavian
and
claude
committed
Add QK_Gain_SLOT_Legal: context-only SLOT (legal causality-safe variant)

Octavian
and
claude
committed
Add BW9_Anchor gate + update SCIENCE.md with MegaGate results

Octavian
and
claude
committed
Fix gate.sh: remove quotes from inline env var assignments
Octavian
committed
Add BW10_GPTQ — loop-aware GPTQ gate on BW8 baseline

Octavian
and
claude
committed
Add gptq_full case: full training run with SKIP_GPTQ=0 (not post_only)
Octavian
committed
Log Arch+Sched sweep results (seed 444, 4×GPU): all 9 cases dead

Octavian
and
claude
committed
Add Rascal_III_SLOT leg: context-only Legal SLOT on Rascal II base

Octavian
and
claude
committed
Add BW11_5Flat — 5F+1C depth revisit on BW8 baseline

Octavian
and
claude
committed
Rascal_III_SLOT run.sh: minimal racer, exact SOTA env + SLOT_ENABLED=1

Octavian
and
claude
committed
BW10_GPTQ: gate PASS — −0.00486 int6_sw, step time clean

Octavian
and
claude
committed
BW10_GPTQ: add production run.sh (8×H100, 600s, LOOP_AWARE_GPTQ=1)

Octavian
and
claude
committed
BW11_5Flat: add production run.sh (8×H100, 600s, NUM_FLAT_LAYERS=5)

Octavian
and
claude
committed
Add RASCAL_WINDOWN_TESTING — 4-arm legal window strategy suite

Octavian
and
claude
committed
BW10_GPTQ: full run PROMOTES — 1.18292670 BPB, new champion

Octavian
and
claude
committed
Rascal_III_SLOT: surgical SLOT via hook, no model class changes

Octavian
and
claude
committed
BW11_5Flat: full run PROMOTES — 1.17651313 BPB, new champion

Octavian
and
claude
committed
Scaffold Crawler II submission — seed=444 pre-filled, seed=300 pending

Octavian
and
claude
committed
Rename submission: Crawler II → Nightcrawler

Octavian
and
claude
committed
Nightcrawler: fill seed=300 results — 1.17490448 BPB, mean 1.1757

Octavian
and
claude
committed
Nightcrawler: add seed logs; fix validate.sh set-e/arithmetic bug

Octavian
and
claude
committed
Nightcrawler: add seed=4 (1.17676091 BPB), update mean to 1.1761

Octavian
and
claude
committed
Rascal_III_SLOT: log both full runs — signal confirmed, size blocked

Octavian
and
claude
committed
Rascal_III_SLOT: enforce cu124 — reject cu128 at launch

Octavian
and
claude
committed
Enforce strict Neural SOTA stack parity on pod setup/run
Octavian
committed
Accept custom hopper FA3 path while keeping strict SOTA gating
Octavian
committed
Prefer working FA3 provider and fail clearly on ABI mismatch
Octavian
committed
Disable FA3 wheel installs by default for custom-head pods
Octavian
committed
Restore custom FA3 discovery in pod_setup.sh

Octavian
and
claude
committed
Lock pod_setup.sh snapshot to vault (2026-04-01, cu124, custom FA3)

Octavian
and
claude
committed
Re-enable FA3 wheel install by default (restore cu124 Hopper FA3)

Octavian
and
claude
committed
Update vault pod_setup snapshot to e7a5944 (FA3 wheel default restored)

Octavian
and
claude
committed
Fix FA3 install: replace dead pytorch WHL URL with GitHub releases

Octavian
and
claude
committed
Add FRESHPOD_OF_BELAIR.sh — original working pod setup (fa91fda)

Octavian
and
claude
committed
Restore vault pod_setup to fa91fda original (FRESHPOD_OF_BELAIR)

Octavian
and
claude
committed
Rascal_III_SLOT: strip agent-injected torch/CUDA version gate

Octavian
and
claude
committed
Add install_fa3_cu124.sh — FA3 wheel installer for cu124/torch2.4

Octavian
and
claude
committed
Add setup_pod_cu130.sh — install torch cu130, verify FA3

Octavian
and
claude
committed
Harden pod_setup.sh + restore run.sh version gate

Octavian
and
claude
committed
Vault-lock pod_setup (cu124, sympy pin, hard assertions)

Octavian
and
claude
committed
pod_setup: add source-build fallback for FA3 when wheel 404s

Octavian
and
claude
committed
pod_setup: add wheel to pip deps (needed for FA3 source build)

Octavian
and
claude
committed
pod_setup: add 2.8.3 wheels (cp312 support), fix cross-device fallback

Octavian
and
claude
committed
Add fresh pod wrapper for Rascal III SLOT
Octavian
committed
Rascal III SLOT copy: fix SLOT batch gating and pack int6 to 6-bit
Octavian
committed
Restore pod_setup.sh to pre-storm version (1ea0901)

Octavian
and
claude
committed
Restore all run scripts to pre-storm versions (9332369)

Octavian
and
claude
committed
Remove storm debris scripts

Octavian
and
claude
committed
Relax CUDA gate in Rascal III SLOT run.sh: 12.4* → 12.*

Octavian
and
claude
committed
Add RASCAL_III_SLOT_F: SLOT + true 6-bit int6 packing

Octavian
and
claude
committed
Rascal III SLOT: add seed300 run log (2026-04-01 17:20)
Octavian
committed
pod_setup: smart FA3 install — Dao-AILab wheels + system discovery

Octavian
and
claude
committed
Add setup_cu124_fa3_venv.sh — automated cu124+FA3 bridge setup

Octavian
and
claude
committed
setup_cu124_fa3_venv: symlink .so instead of copy, add disk cleanup

Octavian
and
claude
committed
crawler: add BW12 2k interaction ablation sequence
Octavian
committed
crawler: record BW12 results and add BW13 4x interaction series
Octavian
committed
pod_setup: enforce true FA3 runtime (remove FA2 wheel fallback)
Octavian
committed
scripts: add fast trimmed-kernel FA3 installer; remove broken cu124 bridge
Octavian
committed
setup: vault fast FA3 installer; remove best-effort FA3 bootstrap path
Octavian
committed
fa3: canonical quick build uses --no-build-isolation
Octavian
committed
rascal_slot: disable torch._dynamo optimize_ddp by default to avoid inductor NaN compile failure
Octavian
committed
revert: restore Rascal III SLOT train_gpt_slot.py exactly (no agent edits)
Octavian
committed
hotfix: add sitecustomize to disable dynamo optimize_ddp and suppress compile crashes
Octavian
committed
crawler: record BW13 interaction ablation results
Octavian
committed
fa3: revert installer flags to historical 4-flag baseline
Octavian
committed
crawler: add BW14 big-swing 2k ablation sequence
Octavian
committed
fa3: add portable cu124 wheel build/install scripts
Octavian
committed

Commits on Apr 2, 2026

neural: add 2k same-checkpoint SLOT H2H probe
Octavian
committed
neural: relax h2h runner stack gate
Octavian
committed
crawler: unify all ablations into single 4x runner
Octavian
committed
neural: add rascal ii 2k comparison runner
Octavian
committed
scripts: add locked cu124 fa3 pod installer
Octavian
committed
scripts: add clean pre-storm pod_setup copy (1ea0901)

Octavian
and
claude
committed
crawler: add BW16 depth sweep for NUM_FLAT_LAYERS 6-11
Octavian
committed
vault: extract user-tested SOTA file from 99b790d (d70ec518, 2468 lines)

Octavian
and
claude
committed
crawler: add BWX latest run (8F + quant sweep)
Octavian
committed
crawler: switch BWX latest baseline to 9F
Octavian
committed
neural: add Lucky — SOTA sequential + SLOT eval

Octavian
and
claude
committed
neural: bake Lucky defaults — seed=444, SLOT=1, SKIP_GPTQ=1

Octavian
and
claude
committed
neural: Lucky — switch to coprime loader (matches March 30 SOTA)

Octavian
and
claude
committed
crawler: harden BWX contender runner with window-first selection
Octavian
committed
neural: Lucky sequential+SLOT seed444 log — 1.10514 BPB (size busted)

Octavian
and
claude
committed
crawler: add submission-ready Bandit Wagon X 9F full-run pack
Octavian
committed
pod_setup: switch to competition dataset source (willdepueoai/parameter-golf)

Octavian
and
claude
committed
crawler: add BW17 DGX-spark cadence longform interaction suite
Octavian
committed
crawler: drop local BW17 smoke artifacts
Octavian
committed
Lucky: bypass torch.save for int6 artifact — raw bytes + JSON header

Octavian
and
claude
committed
Lucky: remove "embed" from int6 categories — restore original quantization

Octavian
and
claude
committed
bw17: harden cadence sweep resume + stride controls
Octavian
committed
bw17: add single-command resume launcher from failed arm
Octavian
committed
Add LC4 8x launcher script
Octavian
committed
Fix LC4 launcher pod paths
Octavian
committed
Make LC4 launcher use pod env fallbacks
Octavian
committed
Probe torch-capable python in LC4 launcher
Octavian
committed
Add RECOVERY_TEST_1 LC4 trainer copy
Octavian
committed
Add lucky_slot recovery trainer
Octavian
committed
crawler: add 24h audit table from BW12-14, BWX9F, isolated 10F
Octavian
committed
Revert lucky_slot export to safe serializer
Octavian
committed
Replace lucky_slot trainer with Lucky
Octavian
committed
Add lucky_slot seed300 SLOT run log
Octavian
committed
Make Lucky SLOT post-export and window-isolated
Octavian
committed
Bake 30 percent SLOT defaults into Lucky
Octavian
committed
Allow partial dataset pulls in pod setup
Octavian
committed
Add 2k rascal signal hunt experiments
Octavian
committed
Log 4gpu 2k build sweep results
Octavian
committed
Add Spark SLOT size hunt runners
Octavian
committed
Add Lucky II contender with legal SLOT
Octavian
committed
Split QK4 contender from warmdown kit
Octavian
committed
Add Lucky III: Lucky II + brotli byte-shuffle compression
Octavian
committed
Add brotli to pod setup
Octavian
committed

Commits on Apr 3, 2026

Lucky III: revert QK4 to baseline 1.5, keep brotli+SLOT
Octavian
committed
Add BW20_Brotli_2k: crawler compression gate (zstd → brotli)

Octavian
and
claude
committed
Add SLOT_brotli: baseline + SLOT + brotli byte-shuffle
Octavian
committed
Add BW XI: BWX 9F + brotli + GPTQ production run

Octavian
and
claude
committed
Fix coprime_shards_per_batch default to 1 (match safepoint)
Octavian
committed
Fix loader_mode default to coprime (match safepoint)
Octavian
committed
Enable SLOT by default
Octavian
committed
Log SLOT_brotli seed 300 breakthrough result
Octavian
committed
BW XI: stack all confirmed gains — loop-aware GPTQ, QK4, loops=2, brotli

Octavian
and
claude
committed
Add Lucky IV: per-sample SLOT delta + 24 steps

Octavian
and
claude
committed
Add Helix: dual-stream co-firing architecture (crawler + flat interleaved)

Octavian
and
claude
committed
Lucky IV: revert per-sample delta, keep only SLOT_STEPS=24

Octavian
and
claude
committed
Add Helix DGX Spark micro test suite (16 arms)

Octavian
and
claude
committed
Lucky IV seed 300: 1.09600210 BPB — new SOTA

Octavian
and
claude
committed
Helix: add Marco-Polo cross-attention mode + 7 new test arms

Octavian
and
claude
committed
Log Helix micro test findings — dim=64 is the breakout signal

Octavian
and
claude
committed
Helix micro suite 2: push dim ceiling, marco-polo, depth scaling (22 arms)

Octavian
and
claude
committed
Add Ouroboros crawler research (non-record) — 1.1364 BPB, 15.05MB

Octavian
and
claude
committed