Fractus — evolution log (timestamped)

Canonical chronology of the system, not the mythology.
Sister repos: thefinalboss/fractus-cte, fractus-p0, fractus-vorax, fractus-datasets.
Last updated: 2026-08-28 03:50 UTC.

Dates are UTC calendar days as recorded in docs, manifests, and HF lastModified. Token counts are manifest snapshots, not live GPU clocks.

2026-04 — seed (outside this repo)

Preprint work on K3 elliptic CM / automorphy (separate mathematical line).
Fractus as an idea: thought is continuous, accumulating, oscillatory — not one forward pass.

2026-07 — architectures in parallel

Fractus CTE specified as multi-block continuous engine (linear attention + Kuramoto + PhaseRoutedMoE).
Palimpseste-max: hypervector append-only memory, CPU, released on HF.
Fractus-Vorax (late July / mid-August design): sealed CTE brain + organs that write knowledge (.kn) without training the 1B again.
eQchip-Q / SOLARIS: other tracks, not this codebase.

2026-08-12 — training doctrine

docs/2026-08-12-fractus-chinchilla.md: Chinchilla-style epoch stacking is rejected for this MoE + HV setup. One real pass + Vorax ingest.

2026-08-13 → 2026-08-15 — first long 4× GPU burn

4× RTX 5090, ~1k tok/s/GPU class.
Loss from chaotic high values down through the 30s.
Discoveries (logged, not marketing):
- generation collapse to single-token attractors
- train graph ≠ length-1 generate graph
- mid-train surgery is possible without throwing the .pt
- mean-merge of same-shape shards produces a usable brain
Hugging Face Space attempt: torch pin / audioop / Gradio stack failures on Python 3.13.

2026-08-16 — outage

Pod / filesystem incident. Resume from last HF .pt, not from token 0.
Migration toward 8× GPU recipe. Docs: BOOST_AND_OUTAGE.md, BOOST_B4_RECOVERY.md.

2026-08-17 → 2026-08-18 — phase 2 + freeze + Vorax

Phase-2 shards from thefinalboss/fractus-datasets (tokenized npy memmap).
Anti re-ingest via manifest / PHASE_SWITCH.
Freeze snapshot: ~~15–17M tokens/GPU on ~430M-token phase-2 shards (~~4% of that pass).
Vorax repo stands up: compiler → traces / Hebbian / spawn organs, :core a sealed .pt.
Decision: finish the pass later on next GPU budget; do not pretend 4% = trained.

2026-08-19 — Kuramoto bottleneck named

Agent probe: degenerate circular mean, post-RK4 routing arc ~25°/360°, (r \approx 0.01) even on a fresh path; ~70% dead experts on probe windows.
Not “wait for more tokens.” Prep: GATE_TEMP, OMEGA_SCALE, real LB in the loss, Kuramoto in autograd.
Docs: KURAMOTO_BOTTLENECK_AND_FIX.md, CPU_MINI_MERGE_AND_DIMENSIONS.md (CPU minis ≠ 1B progress; mini→1B mean-merge is invalid).

2026-08-21 — wrong cluster

QuickPod heterogeneous RTX (3090 / A4000 / 5060 Ti / A2000 12GB). 1B B=2 does not fit 12GB.
Cluster deleted. Marketplace: no 8× offers at probe time. Credit stays on QuickPod; wait for identical 24GB+.

2026-08-22 → 2026-08-26 — open-heart kernels + x8

Mission: faster phase-2 without breaking living checkpoints.
Linear attention: masked einsum (O(C^2)) → cumsum / chunked.
chunked_cross_entropy, BLOCK_CKPT (1B in ~16GB class), atomic ckpt writes, int32 memmap.
Tests: attention/CE equivalence suite.
White paper v2 + arxiv/main.tex landed on the repo.
X8_MANIFEST (pushed): status: x8_running, start tokens ~34.5–35.8M/GPU.
README of 26 Aug claimed ~1550–1600 tok/s/GPU aggregate ~12.6k on 8×5090 — treat as a run log, re-measure on the next pod.

2026-08-27 — P0 / v4 / DiffusionBlocks / lineage made explicit

Package fractus-p0:
- P0 routing: atan2 encode from (h) (not LayerNorm-mean≈0), phase carry into RK4, per-token MoE phases, Switch LB. Not open-heart-equivalent (dynamics change, shapes do not).
- v4 trainer: SS ramp, anti-repeat λ=0.1, PREFIX unique@40 as language gate. CARRY length-1 is a separate hole.
- CPU mini: PREFIX counts 0 1 2 → 4 5 6…; CARRY locks 4 4 4….
- Fractus-native DiffusionBlocks (Sakana/UTokyo ICLR 2026 mapped, not copied): one CTEBlock trained per step, noise band on residual, CE + MSE denoise, 0 new params. Mini 6/6 tests; /σ² denoise detonated last block — removed.
Explicit statement: Fractus is a linear-attention RNN, not a transformer. Open items listed in the root README (decay on (S), DDP vs merge, MI probe on (\bar\theta), matched PPL, eval split).

What “evolution” is not

Not “AGI shipped.”
Not “4.2B tokens digested” — a few percent of phase-2 plus earlier phase-1 shards.
Not “mean-merge = 8× data efficiency.”
Not “Kuramoto is solved” — P0 changes the drive; the MI test is still due.

Next (operational)

Identical 24GB+ GPUs, x4 acceptable.
Resume from X8 / freeze manifest — never token 0 by accident.
Body = P0 + v4. Gate = PREFIX unique@40.
Optional: (\gamma) decay on (S); MI((\bar\theta), token) vs tick; DDP if the fabric allows; matched transformer PPL on a held-out source split.
Vorax for knowledge writes. Do not re-pretrain the brain for facts.

2026-08-27 evening → 2026-08-28 03:50 UTC — SS on 8×5090 (this night)

Full log: docs/2026-08-28-NIGHT-SS-X8.md.

Named the gap in production: CE teacher-force ≠ generation. unique@40 PREFIX NO-GO ~3.3 while tf on GPU0 was already ~5–8.
Scheduled sampling enabled and surviving after checkpoint-metadata crash on the second backward.
Carry snapshot/restore; SS dense CE; no CE-checkpoint when N≤ce_chunk.
Volume 80→200 GB after disk-full trainer death.
All 8 live checkpoints + hourly HF push to thefinalboss/fractus-cte/checkpoints/x8run/.
B=4 + SS_RATE=1.0 live: ~27 GB, ~990 tok/s/GPU, ~40M tokens/GPU on 430M phase-2 shards.
Speech not claimed. SS is the training-side fix; PREFIX unique@40 is the gate.

2026-08-28 16:35 PDT — Decode surgery I

Space output Retail×32 proved the length-1 carry decode is an attractor, not a sentence.

fractus/generate_aligned.py now decodes with a 64-token causal window and masks the previous token. Trainers not stopped.

unique@32 after: 16 / 14 / 27 / 12. Short cycles left. Gate still PREFIX unique@40.

docs/2026-08-28-DECODE-WINDOW.md