This website requires JavaScript.
156ea71879
fix(fleet): every LLM root unit restarts on a voluntary exit; parity seam gates it
Frank
2026-10-13 22:04:48 -07:00
ab80c1366d
refact(deployer): one MainPID gate for all three LLM endpoints
Frank
2026-10-13 22:02:48 -07:00
c5a816e6a8
refact(deployer): the LLM handlers gate through the shared library
Frank
2026-10-13 21:58:13 -07:00
cb25796a6b
refact(frank-archon): serve gates run on the shared library; the serve seam is behavioural
Frank
2026-10-13 21:44:38 -07:00
f31e7c5ed8
refact(frank-zealot): serve gates run on the shared gate library
Frank
2026-10-13 21:41:18 -07:00
028afdeb58
refact(frank-dragoon): serve gates run on the shared gate library
Frank
2026-10-13 21:40:02 -07:00
acf2a5cee5
feat(scripts): one artifact-gate library for the LLM endpoints
Frank
2026-10-13 21:38:50 -07:00
a9d45e7012
frank-archon: the slab is a threshold, and decode is where slots cost
Frank
2026-10-13 20:51:32 -07:00
2581b8d4b6
docs(frank-archon): the slab is a slope, and the ceiling was mine not the engine's
Frank
2026-10-13 20:33:02 -07:00
09a0251d2d
llm(frank-archon): claim the card's settling slack, slab 24000 -> 25600
Frank
2026-10-13 20:18:41 -07:00
066813ea95
llm(frank-archon): 3 x 200,000, pool cut to the contract, slab +1,792 MiB
Frank
2026-10-13 20:00:39 -07:00
7daa886dd6
docs(frank-archon): work-log entry and the 4-way capacity probe
Frank
2026-10-13 19:30:46 -07:00
bae78e7dae
docs(frank-archon): the ttm reset outage, and the 4 x 200,000 fp8 shape
Frank
2026-10-13 19:28:25 -07:00
a6f55f4cc6
fix(frank-archon): raise the ttm knob from a root ExecStartPre — User=dank cannot write it
Frank
2026-10-13 19:20:33 -07:00
a9da6589a6
fix(frank-archon): :" guard needs a space — :"${VAR:?}"` runs :204800
Frank
2026-10-13 19:16:48 -07:00
1a6a78ef73
fix(frank-archon): raise the ttm pages cap in the serve path; reshape to 4 x 200,000 fp8 KV
Frank
2026-10-13 19:09:58 -07:00
c52ba781e3
docs(frank-archon): record the bf16 KV shape, its measured cost, and the probe
Frank
2026-10-11 01:36:48 -07:00
0bf26b0741
llm(frank-archon): unquantized bf16 KV at 2 x 200,000
Frank
2026-10-11 01:01:58 -07:00
394fd8535b
docs(log): record the service-rename cutover outcome and the zealot --metrics gap
Frank
2026-10-10 17:17:59 -07:00
55f87fcdce
llm: retire the trio service names — fleet services renamed to frank-archon/frank-dragoon/frank-zealot
Frank
2026-10-10 17:03:14 -07:00
71c6b5e758
docs(openwebui): frank-dragoon enabled in the picker beside frank-zealot
Frank
2026-10-09 20:43:28 -07:00
5ced474eb9
docs(log): wire-rename wave — three served ids, OpenWebUI + Hermes re-key, bench drop-in retired
Frank
2026-10-09 17:10:25 -07:00
2992a7ddd6
docs: record the wire rename, the OpenWebUI re-key, and the frank-primary bench drop-in trap
Frank
2026-10-09 17:06:44 -07:00
7c77326c33
llm: complete the Protoss rename on the wire — served ids frank-archon/dragoon/zealot
Frank
2026-10-09 16:47:19 -07:00
65dec6a80e
docs(authentik): provision kirby/ry/nic/debo/freya (pk 17-21); OWU pending-role first-login flip law
Frank
2026-10-09 11:49:58 -07:00
f99b698d25
docs(LLM): 3090 worked example — archon w4dn8 methodology ports to GGUF-native quants; 24 GiB size law for 35B-A3B
Frank
2026-10-08 23:44:13 -07:00
757a90057f
docs(LLM): recipe portability audit — qwen35moe plugin carries the calibration tap; 35B-A3B w4 fits one R9700 resident
Frank
2026-10-08 23:42:58 -07:00
009e8a885d
docs(LLM): Quantization Production — 3090 as the unlocked card (runtime menu, Ampere ceilings, Playbook B safetensors ladder)
Frank
2026-10-08 23:38:49 -07:00
6b677139ad
docs(LLM): Quantization Production reference — per-card self-quant playbooks, REAP/Heretic/abliteration/K_P prior art
Frank
2026-10-08 22:39:18 -07:00
184a5c665a
homepage+docs: frank-archon runs our own GPTQ-w4 build, link our HF publication
Frank
2026-10-08 22:15:33 -07:00
958b10545e
homepage+docs: bottom-pinned privacy disclosure chips, live OWU api-keys screenshot, docs steps for 0.6 UI, drop house-DNS/Pi-hole bullet
Frank
2026-10-08 18:07:31 -07:00
c4dca68a03
homepage: link Model Information panel to the docs AI section
Frank
2026-10-08 18:05:50 -07:00
f29dbdf7c7
homepage: rounded corners, aligned title glyphs, greyed closed items, spliced dragoon tagline, inline privacy notes
Frank
2026-10-08 17:29:45 -07:00
98f11a19c1
docs+AGENTS: model ids officially frank-archon/frank-dragoon/frank-zealot (services keep old names)
Frank
2026-10-08 17:16:21 -07:00
a30ac9457a
merge: swift15 runbook LIVE
Frank
2026-10-08 05:02:10 -07:00
75695c312a
swift15: AGENTS.md aligned to deployed shape (fp8 KV, spec 2, seqs 3, PLE fp8 n-gram)
Frank
2026-10-08 05:02:10 -07:00
c507f91438
swift15: runbook LIVE + deployed verification + vision probe law
Frank
2026-10-08 05:02:10 -07:00
b175f4afd1
swift15: vision probe reads reasoning_content fallback, 256-token budget
Frank
2026-10-08 04:57:31 -07:00
971e23a25c
swift15: fix fleet-probes vision heredoc (quoted 'PY' ate $B); probe header
Frank
2026-10-08 04:57:10 -07:00
3ca64e14fc
merge: swift15 KLD re-PASS at production shape
Frank
2026-10-08 04:50:28 -07:00
f3dfea13ad
swift15: KLD re-PASS at production MNBT 2048 (KL 0.0131, top-1 97.92%)
Frank
2026-10-08 04:50:23 -07:00
02fa1e22ae
merge: swift15-radiance-w4dn8 — frank-primary GPTQ w4 (Swift1.5)
Frank
2026-10-08 04:46:11 -07:00
1aef7418ad
swift15 cutover: AGENTS.md + runbook + log + fleet probes + KLD gate shape fix
Frank
2026-10-08 04:41:07 -07:00
fa5a444b64
homepage: simplified console — flat navy, static down-state, no scanlines/lattice/frames/glows
Frank
2026-10-08 04:31:30 -07:00
90f678582e
homepage: navy Protoss console — maximal octagon cornering, Share Tech Mono + Rajdhani, pictograms for CW/RomM/Copyparty/Marreta, dragoon teal
Frank
2026-10-08 04:17:33 -07:00
cc0fdd15a1
swift15: env -> candidate payload, RAD floor 110000 (measured 116,416 MiB)
Frank
2026-10-08 04:11:02 -07:00
f148ecab4a
homepage: HF source links per model; docs reorder (AI last), arr stack out of family docs
Frank
2026-10-08 04:00:26 -07:00
c2f3411342
swift15: campaign decode probe (43.6k greedy, timings-based) + runbook gate results
Frank
2026-10-08 03:58:17 -07:00
8c5bde7034
homepage: Protoss theme, brand-color icons, dragoon portrait, GameBoy RomM, OWU white art + search warn
Frank
2026-10-08 03:51:50 -07:00
22a7331100
homepage: icon swaps, clickable tiles, down pulse, header nav, compaction callout
Frank
2026-10-08 03:37:13 -07:00
c1e192df00
homepage: app-card art, zealot-first order, common uncensored warning, admin identity gate
Frank
2026-10-08 03:21:26 -07:00
8375567de3
swift15: restore expert_tiered + disk-tier flags dropped in the gate rewrite
Frank
2026-10-08 03:18:04 -07:00
79c2e64c2c
swift15: gate MNBT 3072 per Chief (w4 arena scales hard with MNBT; 32k step is absurd here)
Frank
2026-10-08 03:17:33 -07:00
f20da6a6ff
swift15: gate shape MNBT 16384 (w4 arena at 32768 costs 33 GiB), kv 2048, headroom 512
Frank
2026-10-08 03:16:45 -07:00
058ea6cca8
swift15: gate boots at the production pool (w4 gate_up experts are RAM-resident; disk tier refuses non-serving stored forms)
Frank
2026-10-08 03:14:20 -07:00
3f8531179e
swift15: gate flags per engine (--kld-ref takes no corpus; --kld-out JSON parsed)
Frank
2026-10-08 03:12:18 -07:00
3c5d77d1fa
e2e: count fused Model Information names in the user homepage spec
Frank
2026-10-08 02:51:09 -07:00
203b0be145
homepage: Protoss rename + fused Model Information component
Frank
2026-10-08 02:40:45 -07:00
74da746c51
docs(homepage): record the capability equivalence notes
Frank
2026-10-08 01:32:35 -07:00
7066accc25
homepage: capability equivalence on the AI cards
Frank
2026-10-08 01:30:50 -07:00
76aaa2543b
frank-large: raise ctx to 262144 to close the harness 400-zone
Frank
2026-10-08 00:40:12 -07:00
8220030ba1
swift15: canonical recipe = host copy the convert loaded (all 24 protected experts; committed copy was a partial splice)
Frank
2026-10-07 23:45:45 -07:00
5191c49967
swift15: KLD gate script (candidate vs bf16 ref, KL<=0.03 / top-1>=95%)
Frank
2026-10-07 23:43:06 -07:00
f9330fa201
docs(homepage): record the @data allowlist law for new assets
Frank
2026-10-07 23:41:58 -07:00
4b71c014e9
caddy: serve /img/* from the homepage root
Frank
2026-10-07 23:38:18 -07:00
cee90eea52
homepage: amber CRT theme, VT323, current-fleet docs + API access guide
Frank
2026-10-07 23:27:23 -07:00
73c37f4301
swift15: all experts through GPTQ w4 (plugin one-table-one-encoding law)
Frank
2026-10-07 21:31:10 -07:00
6632d0a2f6
docs(openwebui,authentik): frank-secondary-only picker, API gating, artem account
Frank
2026-10-07 18:49:22 -07:00
caf5e1e0ca
docs(authentik): record the human-user template and alex's recreation
Frank and omp
2026-10-07 18:20:11 -07:00
5278f5be26
frank-secondary: record the deployed boot, keep the unit description honest
Frank and omp
2026-10-07 17:33:21 -07:00
ec6184de9a
frank-secondary: restore the shebang my rewrite dropped (unit failed 203/EXEC)
Frank and omp
2026-10-07 14:50:49 -07:00
1d062f06a3
frank-secondary: swap Gemma4-26B-A4B Q4_K_M -> Qwen3.5-9B-Uncensored Q6_K 2x200,704
Frank and omp
2026-10-07 14:47:32 -07:00
b51e078e36
frank-secondary: record the deployed f16-K boot ledger, correct the k_eq_v billing law
Frank
2026-10-07 13:58:30 -07:00
68a525740d
frank-secondary: swap Swift-1.5 RCO-IQ3_XXS -> Gemma4-26B-A4B QAT Q4_K_M
Frank
2026-10-07 13:48:54 -07:00
5b7b3ccb90
frank-secondary: swap to Gemma4-26B-A4B Q4_K_M, 1x200704, q8_0 KV, MTP via separate assistant file
Frank
2026-10-07 12:41:07 -07:00
2137080d1c
docs(index): top-level link for the staged Swift15 runbook (seam 135)
Frank
2026-10-07 04:42:36 -07:00
f747cb0dab
docs(runbook): staged Swift15 w4dn8 runbook skeleton + index link
Frank
2026-10-07 04:40:31 -07:00
db8faa68ff
docs(worklog): KLD reference complete (rows 50235), script-swap law third occurrence
Frank
2026-10-07 02:49:14 -07:00
cf6c62d44a
fix(kld-then-calib): restore S= path clobbered by the SKIP_KLD edit
Frank
2026-10-07 02:45:28 -07:00
0aafad5f2c
fix(kld-then-calib): SKIP_KLD resume + rows-50235 receipt gate
Frank
2026-10-07 02:44:48 -07:00
a3f45ce028
docs(campaign): interim work log for the Swift1.5 w4dn8 campaign
Frank
2026-10-06 23:42:32 -07:00
45de2fd47f
perf(run-swift15-bf16): calib at the vendor big-step shape
Frank
2026-10-06 23:39:10 -07:00
fa858f4c10
feat(campaign): protected-cast applier, recipe markers, KLD gate runner
Frank
2026-10-06 22:41:18 -07:00
07af1a9366
feat(kld-then-calib): chain KLD record -> 1.2.0 calib serve -> 10M push
Frank
2026-10-06 22:39:14 -07:00
f0d0614ce6
fix(run-swift15-bf16): define KLD mounts (were unbound-empty)
Frank
2026-10-06 22:37:16 -07:00
9aed9987b6
fix(run-swift15-bf16): kv pool 3072 both offline arms
Frank
2026-10-06 22:33:19 -07:00
4048dc8630
fix(run-swift15-bf16): per-mode engine pins + explicit KV pools
Frank
2026-10-06 22:31:17 -07:00
9bc40fd899
fix(run-swift15-bf16): drop duplicated radiance arg (image entrypoint already runs radiance)
Frank
2026-10-06 22:25:01 -07:00
bb60e8524f
run-swift15-bf16: boot the bf16 container for KLD record / calibration (vendor disk-tier shape)
Frank
2026-10-06 21:46:05 -07:00
319dd0f7da
derive: restore the header read
Frank
2026-10-06 21:42:16 -07:00
650dfff5a2
derive: RADGRAM2 header is magic,k,rank,rows,domain (28 B)
Frank
2026-10-06 21:42:00 -07:00
de70320cde
derive-protected-experts: RADGRAM2 scan + 3%/3x law + parity fill from calib recording
Frank
2026-10-06 21:41:16 -07:00
910c54b6ff
frank-primary: keep 1.0.13 engine pin — 1.2.0 bench rejected (accept 0.7123 vs 0.7549, decode -2.9%; vram pair exonerated, ledger-exact)
Frank
2026-10-06 21:35:39 -07:00
8c887e6fc9
calib-push: drive calib.jsonl through scratch prefill until the tap writes Hessians
Frank
2026-10-06 21:32:56 -07:00
49dd2be4cd
corpus: handle codeparrot content column + per-class caps
Frank
2026-10-06 21:16:16 -07:00
a11f703c49
swift15: corpus generator v2 — HF code/prose ingest, richer tool JSON, 8k KLD prompts
Frank
2026-10-06 20:58:32 -07:00
6cdf88f611
swift15: restore the numpy import the census edit clobbered
Frank
2026-10-06 20:41:33 -07:00
259fc5505d
swift15: PLE audit adds over-cap census + next-bf16 scale suggestion
Frank
2026-10-06 20:41:07 -07:00
3e46d51df8
swift15: BF16 absmax bits belong in the HIGH half of the float32
Frank
2026-10-06 20:39:21 -07:00
d062a376b2
swift15: corpus generator (10M calib + 1.8M KLD, Froggeric-rendered, disjoint pools)
Frank
2026-10-06 20:38:26 -07:00
Frank
