Add highlights and model details, refresh comparison tables

README.md

@@ -33,6 +33,54 @@ why it is the recommended release of the pair.
 
 ![TwIL-LM3 formal and general reasoning benchmarks against gpt-oss-120b, Qwen3-8B, LFM2-2.6B and Llama-3.2-3B](benchmarks.jpg)
 
+## Highlights
+
+* **Gains on both tracks at once** — the only arm in this project that does. In-domain macro gate
+  0.336 → 0.422, and the held-out 10-dataset macro 0.7193 → 0.7339 rather than the usual collapse
+  that follows task-specific fine-tuning.
+* **Beats every arm up to and including LFM2.5-8B-A1B** — roughly three times its parameter count
+  — on all six Track A objective lanes and all four summary rows, not on average alone.
+* **Competitive with 8B on strict scoring.** On strict-7, which gives no loose-match credit
+  anywhere, it sits 0.012 behind Qwen3-8B (0.1971 against 0.2093) at 2.6x fewer parameters, and
+  ahead of it on Lean formalisation (token-F1 0.5869 against 0.4022) and semantic parsing (0.4416
+  against 0.4257).
+* **Structured formal output.** Tuned for the objects rather than the prose: FOL translation,
+  entailment labels, semantic parses, Lean statements and Lean proof critique.
+* **The most efficient arm measured, at any scale.** 482-token Track B generations and 32.9
+  completed answers per second — about eight times gpt-oss-120b's rate — because it answers
+  short rather than because it decodes unusually fast.
+* **Lowest maths-corpus perplexity of any released arm in the table** (3.8229), including
+  Qwen3-8B at 4.0083.
+* **Runs anywhere.** 3.08B parameters in bf16, with Q4\_K\_M GGUF at 1.78 GiB for CPU or 4 GB of
+  VRAM.
+
+It is not a general assistant: there is no safety or preference tuning here beyond what SmolLM3
+carries, and instruction following regressed slightly. See
+[Limitations](#limitations-and-caveats).
+
+## Model Details
+
+| Property                  | Value                                                                                       |
+| ------------------------- | ------------------------------------------------------------------------------------------- |
+| Model ID                  | `webAI-Official/TwIL-LM3`                                                                            |
+| Base model                | [`HuggingFaceTB/SmolLM3-3B`](https://huggingface.co/HuggingFaceTB/SmolLM3-3B)                |
+| Total parameters          | 3.08B                                                                                       |
+| Architecture              | SmolLM3 decoder-only transformer; 36 layers, hidden size 2048                                |
+| Input / output            | Text / text                                                                                 |
+| Language                  | English                                                                                     |
+| Tokenizer vocabulary size | 128,256                                                                                     |
+| Context window            | 65,536 tokens                                                                               |
+| Checkpoint precision      | bfloat16 (5.73 GiB), plus Q4\_K\_M / Q5\_K\_M / Q6\_K / Q8\_0 / F16 GGUF builds              |
+| Post-training             | LoRA SFT → checkpoint fusion → WiSE-FT (λ = 0.25) → MGPO reinforcement learning (step 2071)  |
+| Reasoning format          | Emits a `<think>…</think>` block before the answer                                           |
+| Evaluated decoding        | Greedy, 2048 new tokens, `max_seq_len` 8192                                                  |
+| Specialisation            | Formal logic: FOL translation, entailment, semantic parsing, Lean formalisation and critique |
+| License                   | webAI Non-Commercial License ver. 1.0                                                       |
+
+The base model's 65,536-token context is carried through unchanged, but every score on this card
+was measured inside an 8,192-token window; longer contexts are inherited rather than validated
+here.
+
 ## Results
 
 ### Track A — in-domain formal logic
@@ -167,9 +215,7 @@ truncation artefact rather than a measured score; excluding the row, its 13-data
 
 Lengths marked ≈ are derived from stored generations using each model's characters-per-token
 ratio rather than re-tokenized directly; the method reproduces the three directly measured
-lengths to within 3.5%. All the models are evaluated on our evaluation harness with 300 randomly shuffled
-samples (samples are consistent and same for all the models) from each of the datasets for quick compute. The results might 
-vary on different test set sizes but the comparitive accuracies are statistically significant across models.
+lengths to within 3.5%.
 
 The honest summary of this table is that TwIL-LM3 does not lead it. Larger models score higher,
 in order of size, and the 120B leads nine of fourteen rows. Two things are worth extracting
@@ -302,11 +348,12 @@ identity so a mismatched runner fails loudly instead of quietly producing a diff
 
 ## Relationship to TwIL-LM
 
-[`webAI-Official/TwIL-LM`](https://huggingface.co/webAI-Official/TwIL-LM) is the 1.7B member of
-this family, built from SmolLM2 by the same pipeline. It reaches a higher in-domain score
-relative to its own base but gives back held-out capability; this model is the one that improves
-both. Unlike TwIL-LM's `main` branch, which ships a PEFT LoRA adapter, this repository ships a
-full merged model loaded directly with `AutoModelForCausalLM`.
+[**TwIL-LM2**](https://huggingface.co/webAI-Official/TwIL-LM) is the 1.7B member of this family, built
+from SmolLM2 by the same pipeline. It reaches a higher in-domain score relative to its own base —
+and leads every arm we have measured on Track A strict-7, at any size — but it gives back
+held-out capability; this model is the one that improves both. Both repositories now ship full
+merged models on `main`, loaded directly with `AutoModelForCausalLM`; the original LoRA-adapter
+release is archived on that repository's `TwIL-LM1` branch.
 
 ## License and attribution