Skip to content
dormouse
Search
Ctrl
K
Cancel
GitHub
Select theme
Dark
Light
Auto
Menu
architecture
dormouse bf16 — all-bf16, no fp32 waste
The system map
design-minimal: the floor of dormouse (2026-09-25)
f16 plan
dormouse glossary
The 28 crates of vendor/dormouse-fused: what each one is for
Mixture arms: the anchors, acted on
The model walk
Multi-GPU
dormouse post-training — self-evolving + EGGROLL
TK design on cubecl
Architecture
archive
Verdict audit, 2026-09-25
BROKEN or KNOWN-WRONG
Сводка находок: что сломано, что подтверждено, что не проверено
fused/ grad-coverage verification (phase 1 item 1)
Retracted claims
The fused RMSNorm CUDA kernel: it lowers now, and there is a second defect behind it
What is in here
Archive
protocols
A/B protocol (2026-09-27, instrument corrected 2026-09-28) — how every mechanism gets judged
Cross-process reproducibility of model init, by measurement
ORACLE — what is actually being verified in burn-kda and burn-gdn2
Oracle tiers
Verification
Protocols
research
papers
Provenance (sha256)
Papers & specs
attnres
code snapshots
dspark
engram
gdn kda
jepa
muon plus
output gate silu vs sigmoid
spec flashkda
spec gdn2 official
spec kimi k3
spectral reference
tsct
decisions
Decisions
do we need jepa
class b 2026 10 01
1b readiness plan
notes
Notes
2026-09-07 · minicpm5 dormouse direction
2026-09-21 · per gb sota
2026-09-25 · nasa burn rust practices
2026-09-27 · adaptive depth safe
2026-09-27 · adopt vs port
2026-09-27 · cublas integration poc
2026-09-27 · domain anchors
2026-09-27 · engram reenable
2026-09-27 · kda sota ceiling
2026-09-27 · oracle audit
2026-09-27 · pkm engram deepseek
2026-09-27 · pytorch baseline
2026-09-27 · scale ladder
2026-09-29 · res attn
2026-09-29 · res factor
2026-09-29 · res optimizer
2026-09-29 · res simplify
2026-09-29 · res tracks3
2026-09-30 · alphaxiv answers batch 1
2026-09-30 · attnres gate falsification
2026-09-30 · attnres integration
2026-09-30 · cross process repro
2026-09-30 · state carry defect
2026-10-01 · alphaxiv answers batch 3
2026-10-01 · alphaxiv answers batch 4
2026-10-01 · spectral inventory
2026-10-07 · technology super research
2026-10-11 · 022 reads unwritten memory
2026-10-11 · alodlm fang et al
2026-10-11 · deepseek v41 flash kv compression
2026-10-11 · recurrence equivalence
2026-10-15 · 1b tech verification
2026-10-15 · agentic post training
2026-10-15 · data strategy
2026-10-15 · mimo rl notes
2026-10-15 · torch spike
2026-10-15 · tsct super kernel
2026-10-16 · budget savers
2026-10-16 · data mixing
2026-10-16 · jepa dspark alternatives
2026-10-16 · jepa into dormouse
2026-10-16 · long context 1m recipe
2026-10-16 · muon orthogonalization cost
2026-10-16 · new tech scan
2026-10-16 · one button pipeline
2026-10-16 · sf muon
2026-10-16 · token skipping
2026-10-17 · muon variants
2026-10-17 · olmo core ideas
2026-10-18 · act4
2026-10-18 · compile moe
2026-10-18 · data capacity
2026-10-18 · patch size
2026-10-18 · ram offload
2026-10-18 · self play
2026-10-18 · throughput leftovers
Research
reviews
Code audit: 6/10, and the two holes that matter
Findings
The review loop: rounds and scores
Review round 1: 7/10, the fixes verified and what they broke
Review round 2: 6/10, the regressions the fixes introduced
Review round 3: 7/10, the final score of the cycle
Night report: the loop closed, the fleet landed, the base is the wall
Reviews
hygiene 2026 10 01
wgpu smoke 2026 10 02
graph trainer 2026 10 01
verify tails 2026 09 30
upstream issue autotune candidate coin flip
unguard 2026 10 02
tsct2 design 2026 10 04
tsct spec 2026 10 03
tsct diag 2026 10 01
trio prereg 2026 10 04
speedland 2026 10 03
spectral stack audit 2026 10 01
spectral audit 2026 10 01
situ oom 2026 10 02
situ 2026 09 30
rope 2026 09 30
rmsnorm live 2026 10 03
research digest 2026 10 02
research agent round2 2026 10 02
repo excellence 2026 10 01
posttrain prep 2026 10 03
patching research 2026 10 02
optimization map 2026 10 02
olmo integration 2026 10 04
night verdicts 2026 10 02
night 2026 10 02
multimodal chapter 2026 10 02
msa reentry 2026 10 02
moe routing 2026 10 01
moe dev 2026 10 03
mhc 2026 09 30
maxopt2 2026 10 02
littlebit design 2026 10 15
launch atlas 2026 10 02
kdafix 2026 10 02
kda oracle ci 2026 10 02
kda gradflow 2026 09 30
kda adjoint 2026 10 01
jepa3 2026 10 03
integration 2026 10 02
future byte 2026 09 30
fused library ci 2026 10 01
fidelity audit 2026 10 02
f16 tensor core 2026 10 01
doc coverage wave2 2026 10 01
doc coverage 2026 10 01
dispatch guard 2026 10 02
dedup optimizer 2026 10 01
dead pub audit 2026 10 01
cuda graph 2026 09 30
crate deletion 2026 10 02
config seams 2026 10 01
ci quality 2026 10 01
byteflow rate 2026 10 02
byteflow patcher 2026 10 02
byteflow ab 2026 10 02
breadth t3 2026 09 30
bpbaudit 2026 10 04
bf compat 2026 10 03
adjoint tune 2026 10 03
ab wave 2026 10 01
2026-10-21 · battery
2026-10-19 · model scorecard
2026-10-19 · infra audit
2026-10-17 · port night report
2026-10-17 · port audit
2026-10-16 · b1 profile
2026-10-09 · bug list 022
2026-10-09 · 022 autotune corruption
2026-09-30 · kda formula audit
2026-09-29 · muon tsct review
2026-09-29 · gdn kda review
2026-09-29 · engram jepa dspark review
2026-09-29 · attnres review
start-here
The agent rulebook
Contributing
The docs map
License
notes
Notes: architecture
Notes: the goal
Working memory
Working plan
Working notes
Program plan
Quick start
Status
What this is
Start here
tooling
A/B testing
Build time: which command, and why it costs what it costs
CLI flags
Data
Determinism
The .dmexp file
First PR
Precision
Presets
Reproduce
First run
Tooling
check_doc_refs
dist_gate
mk_wordmark
wt
adr
ADR
ADR-0001
ADR-0002
ADR-0003
ADR-0004
ADR-0005
ADR-0006
ADR-0007
ADR-0008
ADR-0009
ADR-0010
ADR-0011
ADR-0012
ADR-0013
ADR-0014
ADR-0015
ADR-0016
ADR-0017
ADR-0018
ADR-0019
ADR-0020
ADR-0021
ADR-0022
ADR-0023
dormouse
GitHub
Select theme
Dark
Light
Auto
License
License
Section titled “License”
MIT — see
LICENSE
.