Punk Records Atom Token LLM (v1.3-prep)
Separable ψ memory for agents. Session state lives in a neural ψ field, not in transcript token replay — O(1) input per turn by construction, and experience is a portable, auditable .atom file.
| Paper | Full PDF + figures (Draft v2.0) · Research · Evidence |
| Author | Umberto Canessa Cerchi · umberto.canessa@gmail.com |
| Public cut | 2026-07-16 · code pin 31edd83 · evidence ship_validation_20260716_095230 · paper evidence pin PAPER_V2_EVIDENCE_PIN.md |
| Ship atomizer | atomizer_v9_3_19d_forensic_repair_nl_sidecar.pt |
| Private code | punk-records-atom-token (access on request) |
Demo Space (
wasnaga/atom-token-demo) is deferred — this model repo is the public face for paper, evidence, and ship weights.
What this is
Mainstream chat agents remember by re-tokenizing prior turns (or by paging retrieved text back into the prompt). Cost grows with session length; multi-agent latent state is hard to audit.
Atom Token LLM inverts the stack. The native unit is an AtomToken (σ, ψ, π, φ); memory is a bank of neural cells (κ, ν) with learned addresses. Each inference step takes only the latest user message + ψ — no transcript replay, no runtime weight training. Surface text appears at RENDER egress only. Weights provide ability (atomize, route, compile teaching shape); ψ accumulates experience (facts and procedures planted at inference). Portable .atom files (format v8) carry session state between agents.
The memory law is the claim. What the memory unlocks for procedure execution — and where novel compile remains open — is reported as epilogue evidence, not headline.
Headline results (TEST · ship stack · scaffolds off)
Memory and routing
| Claim | Metric |
|---|---|
| Horizon recall @ N=8/16/32 | 8/8 · 16/16 · 32/32 (pinned capacity sweep) |
| Horizon capacity N→1500 | 1403/1500 (~93.5%) PARTIAL — soft ceiling, not the ship bar (Finding #16: constant --seed) |
| Routing @ banks 8–64 | 100% recall and read-given-routed (window 8) |
| Codec exhaustive / horizon / induction | 3942/3942 · 18733/18733 · 12/12 (train-pool coverage — disclosure in paper §3.6) |
| Token economics @ 32 facts | ~97.6% session-token savings vs naive transcript replay (flat ~86 tok/query vs linear growth; OpenRouter hits the 128k wall @ N=512, COMPARATIVE) |
Society (Act 4 · seeds 4500–4999)
| Claim | Metric |
|---|---|
| Completion (reactive / planner / dreamer) | 500/500 each |
ψ-transfer mentorship (.atom file) |
500/500 · control 0/500 |
| Fact diffusion | 500/500 · mean latency 1.0 rounds |
| Post-lock deploy refused (plan_act) | 500/500 |
| Stale redeploy | planner 0% · reactive 100% |
| Dream quarantine | 0 violations |
| Transcript ablation | 220/500 (visibility-limited, informational — not a gate) |
| Society input token savings | 31% |
World model + v3 closure
| Claim | Metric |
|---|---|
| Plan-act pick / refuse | 500/500 each (post-lock deploy) |
| G-wm-* gates | 100% on TEST trajectories (incl. counterfactual) |
| G-skill-study-mul10 / chain | 50/50 · 50/50 |
| G-behavior-study-plant / transfer | 97/97 · 97/97 |
| G-nl-induction-skill / behavior / transfer | 50/50 · 97/97 · 97/97 |
Template / skill scaffolds are off on the default path.
Epilogue evidence — what the memory unlocked (scoped, not the ship claim)
Reported with the same honesty bar as the paper's Epilogue; these are research-in-motion results, not merged ship claims.
| Evidence | Result |
|---|---|
| Phase E arithmetic (operator textbooks) | Ch.3 15/15 · Ch.4 10/10 held-out APPLY with frozen weights; stress 13/13 on saved .atom. In v1.3 the compile of those procedures was assisted by notation regexes and (for division) a hardcoded template — disclosed and since deleted |
| Physics 4c.2 | 9/9 on the same run — physics-planted procedures were unused by graded cases; method mapping was regex (deleted). Not a "studied physics" claim |
| Sacred chat-teach ÷ (2026-07-17) | Basic math taught in chat: division (as repeated subtraction with a count) 12/12, and flooble — a novel doubling-count operation that exists nowhere in training or vocabulary — 10/10; Arm B 0/12 + 0/10 controls — a 3-turn teacher-register lesson compiles into ψ as an executable loop and answers held-out prompts on certified-clean corpora (gate-0). ÷ schema deliberately starved ≤10%: the binding is lesson-supplied, not memorized |
| Register boundary | Free-prose lessons 0/12 (compile) — the open frontier (Phase 2 language front-end) |
Methodology: findings ledger #1–#19 (incl. scaffold fraud found and deleted in public, gate-0 corpora certification, checkpoint-selection and seed-law fixes).
Architecture (production ship)
| Component | File | Role |
|---|---|---|
| Atomizer (NL sidecar) | atomizer_v9_3_19d_forensic_repair_nl_sidecar.pt |
Compile + π routing |
| κ encoder | kappa_encoder.pt |
Hierarchical addresses |
| Tension v4 | tension_field_v4.pt |
Write/read routing |
| Value codec | value_codec.pt + value_codec_induction.pt |
ν encode/decode (re-encode verified search) |
| σ encoder | sigma_encoder.pt |
Query↔plant binding |
| Mixer | sigma_mixer_v2.pt |
σ mixing |
| ACT | act_tool_bind.pt, act_arg_extract.pt |
Tool surfaces |
| ψ-predictor | psi_predictor.pt |
plan_act imagination (no ψ write, grounding veto) |
| Transcript baseline | transcript_baseline.pt |
Society ablation arm |
| Chitchat | unified_chitchat_v16.pt |
Catalog + CRS egress |
~69M parameters loaded (~9M memory/routing substrate + ~59M chitchat/render). Manifest: stack.json.
This repo publishes ship files only. Legacy atomizers (v3/v4/v6/v7) and skill-split κ/tension are not included.
Download
pip install -U "huggingface_hub[cli]" torch pyyaml numpy
huggingface-cli download wasnaga/atom-token-llm-v1 --local-dir ./checkpoints/lane_d
export LANE_D_CHECKPOINT_DIR="$(pwd)/checkpoints/lane_d"
export LANE_D_SHIP_STACK=ship
export LANE_D_ALLOW_HF_DOWNLOAD=0
For gates and Phase E, prefer the private GitHub checkpoint release when available — Hub is the public mirror of the same ship filenames.
Reproduce
Private code repo with pinned checkpoints:
export LANE_D_CHECKPOINT_DIR=/path/to/checkpoints/lane_d
export LANE_D_SHIP_STACK=ship
export LANE_D_ALLOW_HF_DOWNLOAD=0
export LANE_D_TEMPLATE_SCAFFOLD=0
export LANE_D_SKILL_SCAFFOLD=0
export PYTHONPATH=.
# One-command ship chain (Windows):
# scripts/run_full_ship_validation.ps1
# Requires operator PDFs via LANE_D_MATH320_PDF / LANE_D_OPENSTAX_PHYSICS_PDF
python -u -m lane_d.product.eval.run_release_gates
python -u -m lane_d.product.eval.run_fp_closure_gates
python -u -m lane_d.society.eval.export_evidence
Evidence JSON in this repo: evidence/.
Limitations (honest)
- Deterministic ops-deploy society simulator; primarily 2-agent episodes.
- Not a frontier general LLM — gated architectural primitive at modest scale (~9M memory substrate).
- Horizon N→1500 is PARTIAL (~93.5% soft ceiling); ship bar is gap audit 8/16/32.
- Codec and routing numbers are partly in-distribution (train-pool coverage; tension cascade pool includes DEV+TEST realm IDs — disclosed in paper §3.6/§4.4).
- Sacred chat-teach succeeds in teacher register (12/12 ÷ · 10/10 flooble · controls hold) — scoped to that register; free-prose lessons still fail (0/12), so the language front-end remains open work.
- Interactive Space demo deferred.
Full discussion: paper §9–§10 and Epilogue.
Citation
Canessa Cerchi, U. (2026). Separable ψ Memory for Agents: Ability/Experience Split with O(1) Input and Portable .atom State (Draft v2.0 · 17 July 2026).
https://huggingface.co/wasnaga/atom-token-llm-v1
@misc{canessa2026atomtoken,
author = {Canessa Cerchi, Umberto},
title = {Separable ψ Memory for Agents: Ability/Experience Split with O(1) Input and Portable .atom State},
year = {2026},
howpublished = {Hugging Face},
url = {https://huggingface.co/wasnaga/atom-token-llm-v1},
note = {Draft v2.0; v1.3-prep ship cut 2026-07-16; PDF in /paper/}
}