Ship trained weights in the kit; refocus moat on integration + V2

Weights now Apache-2.0 alongside the rest of the kit. The license-as-moat
strategy was already broken (Apache 2.0, ~$1-2 reproduction cost); a
weights-as-moat strategy was thin for the same reasons. Shipping them
makes the kit actually runnable and makes the HONEST_RESULTS reversal
reproducible end-to-end.

Files added to checkpoints/:
  - atome_944k.bin     (271 KB)  packed C-engine blob (ATOME01 format)
  - atome_1m_v1.pt     (3.7 MB)  PyTorch source checkpoint (for fine-tune)
  - vanilla_1m_v1.pt   (3.7 MB)  FP32 vanilla baseline (for the 944K A/B
                                 reversal in HONEST_RESULTS.md)

.gitignore: keep `checkpoints/*.pt` and `checkpoints/*.bin` as the default,
with explicit `!` allow-list for the three released files. Future random
checkpoints stay ignored.

README: drop the "trained model weights are not in this release" line;
add a Weights section explaining what each file is, plus an honest
caveat that the 944K is a research-demo artifact, narrow and sometimes
incoherent, here to make the architecture runnable, not to set a quality
bar.

PROJECT_CONTENT.md "What's open vs. what's not" table rewritten so the
moat columns reflect reality: weights and architecture are open; the
defensible value is silicon bring-up, the Secure Boot Pack, per-platform
hardening, and the larger internal V2 (3M-param, mixed-domain) model
which stays proprietary.

Author: Atome LM Team

.gitignore

@@ -5,11 +5,17 @@ __pycache__/
 .pytest_cache/
 build/
 
-# Trained checkpoints — too big and binary
+# Trained checkpoints — too big and binary, EXCEPT the three released artifacts
+# (944K demo blob + its PyTorch source + the vanilla FP32 baseline used for
+# the HONEST_RESULTS reversal A/B). These are tracked.
 checkpoints/*.pt
 checkpoints/*.atome
 checkpoints/*.atome2
+checkpoints/*.bin
 checkpoints/ab_sweep/
+!checkpoints/atome_944k.bin
+!checkpoints/atome_1m_v1.pt
+!checkpoints/vanilla_1m_v1.pt
 
 # Generated corpus + on-disk fetch cache (sample.txt is intentionally tracked)
 data/*.txt

PROJECT_CONTENT.md

@@ -91,14 +91,16 @@ That's the only signal that matters before declaring done. If you change anythin
 
 ## What's open vs. what's not
 
-| Open (this repo, Apache 2.0)              | Not open (commercial)                  |
-|-------------------------------------------|----------------------------------------|
-| Architecture, training code, C engine     | Trained model weights (e.g. the 944K)  |
-| Export format + parity tests              | Per-customer MCU bring-up              |
-| Sample data, A/B sweep harness            | Atome Secure Boot Pack (signed blobs)  |
-| All docs (PAPER, HONEST_RESULTS, etc.)    | Per-platform hardening playbooks       |
-
-The architecture is public by design and the training cost is ~$1–2 — a license-as-moat strategy was never going to work. The actual value sits in the weights, the per-deployment integration work, and the security hardening, none of which are in this repo.
+| Open (this repo, Apache 2.0)                        | Not open (commercial)                         |
+|-----------------------------------------------------|-----------------------------------------------|
+| Architecture, training code, C engine               | Silicon bring-up (per-platform integration)   |
+| 944K trained weights (`checkpoints/atome_944k.bin`) | Atome Secure Boot Pack (signed `.atome` blobs)|
+| PyTorch source `atome_1m_v1.pt` + vanilla baseline  | Per-platform hardening + attestation flows    |
+| Export format + parity tests                        | Larger internal V2 model (3M params, mixed-domain) |
+| Sample data, A/B sweep harness                      | Custom fine-tuning + per-customer integration |
+| All docs (PAPER, HONEST_RESULTS, etc.)              | Marketing / live-demo site at atomelm.com     |
+
+The architecture is public by design and the training cost is ~$1–2 — a license-as-moat strategy was never going to work, and weights-as-moat would have been thin. The actual defensible value is the per-deployment integration work, the security hardening, and the larger V2 model kept proprietary — none of which sit in this repo.
 
 ## If you need to dig deeper
 

README.md

@@ -22,8 +22,24 @@ own checkpoint with `scripts/train_demo.py` in ~30 min on a CPU.
 **MCU status:** QEMU ARM (Cortex-M3, MPS2-AN385) parity passes to FP32
 epsilon. **Silicon bring-up is not done in this repository.** For real
 device deployment we sell integration — silicon bring-up, the Atome Secure
-Boot Pack, per-platform hardening — at [atomelm.com](https://atomelm.com).
-Trained model weights are not part of this open-source release.
+Boot Pack (signed `.atome` blobs, dev/prod flags, per-platform secure-boot,
+attestation), per-platform hardening — at [atomelm.com](https://atomelm.com).
+
+**Weights are included** in `checkpoints/`:
+
+- `atome_944k.bin` (271 KB) — the packed C-engine blob (`ATOME01` format),
+  loadable directly by the inference engine.
+- `atome_1m_v1.pt` (3.7 MB) — the PyTorch source checkpoint that produced
+  it; use this to fine-tune or to re-export at different `#define`s.
+- `vanilla_1m_v1.pt` (3.7 MB) — the FP32 vanilla GPT baseline used for
+  the 944K A/B reversal in [`HONEST_RESULTS.md`](HONEST_RESULTS.md);
+  shipped so you can reproduce the comparison end-to-end.
+
+The 944K checkpoint is a research-demo artifact, not a product: it is
+narrow, sometimes incoherent, and trained on a single corpus. It is here
+to make the architecture *runnable*, not to set a quality bar. Reproduction
+is ~$1–2 of CPU/GPU using the included training code; nothing in this kit
+is a reproduction barrier.
 
 ---
 
@@ -163,7 +179,7 @@ Apache License 2.0 — see [`LICENSE`](LICENSE) and [`NOTICE`](NOTICE).
 
 The kit is fully open: use, modify, redistribute, and ship in commercial products without per-seat or per-device fees. The Apache 2.0 patent grant covers the 3-pathway routed-ternary architecture as released here.
 
-Trained model weights are not in this release. Commercial integration, the Atome Secure Boot Pack (signed model blobs, secure-boot guidance for ESP32/STM32/RP2040), and per-customer MCU bring-up are available at [atomelm.com](https://atomelm.com).
+The released checkpoints in `checkpoints/` (atome_944k.bin, atome_1m_v1.pt, vanilla_1m_v1.pt) are likewise Apache-2.0. They are reference / research artifacts, not products. Commercial integration — silicon bring-up, the Atome Secure Boot Pack (signed `.atome` blobs, dev/prod flags, per-platform secure-boot, attestation), per-platform hardening, custom-domain fine-tuning of the larger internal V2 model — is available at [atomelm.com](https://atomelm.com).
 
 ## Citation