atom-ai / docs /MERGE_OPERATIONAL.md
thefinalboss's picture
ATOM (atom-ai): code snapshot (100fae0 + batched-train), README with measured numbers, curated checkpoints
9f1076a verified
|
Raw History Blame Contribute Delete
3.84 kB

MERGE operational under hard-v2+ — 2026-09-18

English. Numbers. No fluency / GPT claim.

North star: readable French phrases via persistent toroidal field + Atomizer (CPU continuum only) — no attention / HF / DDP / clusters / Transformers.

Falsified prior (do not reopen)

Linguistic spans 16→32 @ 02891ff (3.325M→3.425M): mean pkt ~19 B, chat still noise. Do not burn another span/printable chunk alone.

Root cause (logged)

Train traj under linguistic +100k: merges=0 always, mph≈0.500 always, n_pairs_above_energy_floor thousands. So energy floor was not the block.

Checkpoint atoms: E≡0.5, φ≈0 → coherence = cos(Δφ)·mean(√(Ei·Ej)) = 0.5. Default phase_coherence_threshold=0.55 used score < thr → every pair rejected.

Mechanism (one) — lower MERGE coherence gate

knob old new default
phase_coherence_threshold 0.55 0.45
merge_energy_floor 0.08 0.08 (unchanged; pef already high)

CLI: --merge-coherence-threshold / --merge-energy-floor (CLI wins on resume).

Keep instrumentation: max_phase_coherence (mph), n_pairs_above_energy_floor.

Also required for live MERGE (latent until merges fired):

  1. Restore MERGE outputs to AtomCompiler shapes (1, d) / rho (1, 1) so the atom collection can stack with live atoms.
  2. Consolidation: expand singleton atom-mean E/stability to n_modes before boolean mask on α (mask shape bug tripped once densified atoms boosted E).

Unit tests: synthetic coherent atoms → merge_count > 0 under defaults; mph=0.5 regime merges at 0.45 and not at 0.55; set_merge_thresholds propagates.

Data / train (smoke)

item value
resume checkpoints/atom_native_obligatory_hard/ @ 3,425,000
data --stream --data-glob data/dialogue_shards/part_*
budget +8k smoke (protocol ~5k–10k)
hyperparams hard ON, L_ign=0, printable_aux=0, --enable-merge, thr=0.45, floor=0.08, linguistic@32, --no-episode-reset, d=64, field_max_rms=3
throughput ~39 tick/s CPU

Trajectory (3.425M → 3.433M = +8k)

stage total step merges (cum) mph atoms (log) logits cos α cos chat
linguistic end 3,425,000 0 ~0.500 grow/flush 0.866 0.713 noise
MERGE smoke +8k 3,433,000 7968 ~0.69 1 (collapse) 0.860 0.705 still noise

Log: merges= column is cumulative merge_count_total. 160/161 traj rows had merge_count>0 this run (~1 merge/tick after first). After each inject, pair merges immediately → list collapses to 1 heavier atom (ρ bumped, E boosted → mph ~0.693).

Chat samples after +8k (honest — not coherent)

  • Bonjour → binary / Latin scraps / loops — no phrase
  • Qui es-tu ? → empty / near-empty — not an answer
  • Il était une fois → short scraps — not a story opener
  • Bonjour (deterministic) → AAa.Mtl… loop — not FR

No fluency claim. MERGE is operational (count > 0); speech still locked.

Probe (after smoke)

metric value
logits off-diag cos 0.860
α off-diag cos 0.705
reconstruction True (partial/diagnostic)
verdict field carrying structure; gen still noise

Artifacts: docs/artifacts/merge_operational/.

Verdict / next

PASS for Step 1 mechanism: MERGE fires under hard-v2+ with thr=0.45. Honest side effect: structural list collapses to ~1 atom/tick — densify works as count, not as a rich multi-atom hierarchy yet.

Next (only after this commit): ONE field→next-packet production loss (small weight), separate train — not stacked in this blob.