README: adapt pitch to honest framing

Replace "smallest LM that talks / no GPU no cloud no RAM / runs on a $2 MCU"
hero with "reference implementation of a routed-ternary tiny LM with
bit-exact Python <-> C99 inference, sized for MCU-class RAM budgets."

Acknowledge prior art explicitly (BitNet b1.58 for ternary, Hymba/MossNet
for hybrid SSM + sparse-attention routing, Super-Tiny-LM for byte
tokenizer) and state directly that the contribution is integration, not
architecture.

Add explicit MCU status line: QEMU ARM parity passes; silicon bring-up
is NOT done in this repository. Commercial integration (silicon bring-up,
Secure Boot Pack, hardening) is sold separately.

Rename "Frontier finding" -> "Reproducible result, narrow regime" and
keep the 944K reversal up front instead of buried.

Author: Atome LM Team

README.md

@@ -1,31 +1,44 @@
 # Atome LM
 
-> **The smallest language model that talks.**
-> **No GPU. No cloud. No RAM.** Runs on a $2 microcontroller.
-
-A tiny ternary language model designed for the Atome C99 inference engine. Three-pathway router architecture, byte-level tokenizer, integer-arithmetic forward pass, ships in under 100 KB of flash at default config — small enough for a Cortex-M0, an RP2040, an ESP32-C3.
-
-**Try it now:** [live browser demo at atomelm.com](https://atomelm.com/demo.html) (no install, no signup).
-
-**Get the kit:** training code, C engine, MCU firmware, benchmarks, paper — all in this repository, released under the [Apache 2.0 License](LICENSE). Train your own checkpoint with `scripts/train_demo.py` in ~30 min on a CPU.
-
-Trained model weights are not part of this open-source release. Commercial integration support — MCU bring-up, the Atome Secure Boot Pack, per-platform hardening — is available separately at [atomelm.com](https://atomelm.com).
+> A reference implementation of a routed-ternary tiny language model with a
+> bit-exact Python ↔ C99 inference engine, sized for microcontroller-class RAM
+> budgets.
+
+60K-default-parameter LM combining three known ideas into one open kit:
+ternary weights (after [BitNet b1.58](https://arxiv.org/abs/2402.17764)),
+a hybrid SSM + sparse-attention + local-conv block routed per token
+(after [Hymba](https://arxiv.org/abs/2411.13676) and
+[MossNet](https://arxiv.org/abs/2510.26182)),
+and a byte tokenizer at super-tiny scale
+(after [Guertler 2024](https://arxiv.org/abs/2405.14159)).
+**The contribution is integration, not architecture**: a complete train →
+ternary export → base-3 packing → C99 inference path, with bit-exact
+Python ↔ C parity enforced by tests.
+
+**Get the kit:** training code, C engine, benchmarks, paper — all in this
+repository, released under the [Apache 2.0 License](LICENSE). Train your
+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.
 
 ---
 
-## Frontier finding
+## Reproducible result, narrow regime
 
-At fixed parameter count, Atome's 3-pathway ternary architecture beats a
-vanilla GPT-FP32 baseline by **22 % on perplexity** (6.31 vs 8.12 ppl)
-while using **16× less disk** (15.1 KB vs 237.5 KB). At fixed flash
-budget, Atome wins by **52 %** (6.31 vs 13.10 ppl). 3,000 steps on
-TinyStories, single seed; full reproduction in [FRONTIER.md](FRONTIER.md).
+On TinyStories, 3000 steps, single seed: at fixed parameter count Atome's
+routed-ternary block reaches **6.31 ppl vs 8.12** for a vanilla GPT-FP32
+baseline (−22 %); at fixed flash budget **6.31 vs 13.10** (−52 %). Disk
+footprint is 16× smaller at param-match (15.1 KB vs 237.5 KB).
 
-**This win holds at the ~60 K-parameter MCU regime and only there.** In a
-944 K-parameter scale-up A/B the vanilla FP32 baseline reverses it (wins by
-~11 %) — Atome's bet is deliberately the sub-1M, MCU-class regime. The full
-honest reading, reversal included, is in [`FRONTIER.md`](FRONTIER.md) and
-[`HONEST_RESULTS.md`](HONEST_RESULTS.md).
+**The result reverses at 944 K parameters**, where the vanilla FP32 baseline
+wins by ~11 %. Atome's bet is deliberately the sub-1M, MCU-class regime;
+above it ternary's capacity ceiling closes the gap and overruns it. Full
+reproduction in [`FRONTIER.md`](FRONTIER.md), full honest reading including
+the reversal in [`HONEST_RESULTS.md`](HONEST_RESULTS.md).
 
 ## Why