Refresh org card: DD paper published, list public PoC models post-migration

README.md

@@ -1,41 +1,63 @@
----
-title: README
-emoji: 🔩
-colorFrom: gray
-colorTo: blue
-sdk: static
-pinned: false
----
-
-<div align="center">
-
-# Tinman Lab
-
-### Autonomous Machines. Second-Order Systems.
-
-<sub>AGENT MEMORY · ADVERSARIAL SAFETY · AGENTIC ECONOMY · PERCEPTION SYSTEMS · APPLIED RESEARCH</sub>
-
----
-
-</div>
-
-## Disposition Distillation
-
-Tinman Lab develops **Disposition Distillation (DD)** — a multi-teacher distillation methodology that trains *how a model behaves* into weights, not system prompts. DD models plan before acting, acknowledge uncertainty, verify their own reasoning, and know what they don't know.
-
-- **4-stage all-MIT pipeline** — Kimi K2.5 → GLM-5 → MiniMax M2.7 → GLM-5
-- **7 behavioral dispositions** — Eager, Deliberate, Adversarial, Curious, Self-Improving, Humble, Persistent
-- **On-device focus** — 0.6B to 2B parameters, quantized for mobile and edge deployment
-- **100% open training data** — MIT-licensed teachers only, zero proprietary model outputs
-
-## Models
-
-| Model | Size | Description |
-|-------|------|-------------|
-| [tinman-code-0.6B](https://huggingface.co/Tinman-Lab/tinman-code-0.6B) | 418 MB | Coding assistant with meta-cognitive awareness — plans, verifies, flags uncertainty |
-
-## Links
-
-- [Website](https://tinmanlab.com)
-- [GitHub](https://github.com/tinmanlabsl/)
-- Research Paper — coming soon
+---
+title: README
+emoji: 🔩
+colorFrom: gray
+colorTo: blue
+sdk: static
+pinned: false
+---
+
+<div align="center">
+
+# Tinman Lab
+
+### Autonomous Machines. Second-Order Systems.
+
+<sub>AGENT MEMORY · ADVERSARIAL SAFETY · AGENTIC ECONOMY · PERCEPTION SYSTEMS · APPLIED RESEARCH</sub>
+
+---
+
+</div>
+
+## Research
+
+### Disposition Distillation — a three-arc negative result
+
+We set out to train *behavioral dispositions* — self-verification, uncertainty
+acknowledgment, feedback integration — into sub-billion-parameter language
+models. Across three independent operator classes (SFT/DPO LoRA imitation,
+attention-head tempering on `o_proj`, and frozen-base hidden-state confidence
+sidecars), no operator moved judge-measured disposition without simultaneously
+damaging content quality or collapsing into stylistic mimicry. The failure is
+consistent across Qwen3-0.6B, Qwen3-1.7B, Qwen3.5-0.8B, Gemma 4 E2B, and
+SmolLM2-1.7B-Instruct.
+
+The contribution is the falsification, a two-failure-mode taxonomy for linear
+hidden-state probes, and a methodological pipeline that converts
+CV-on-same-distribution false positives into honest negatives.
+
+📄 **Paper** — [arXiv:2604.11867](https://arxiv.org/abs/2604.11867)
+🔧 **Artifacts** — [github.com/tinmanlabsl/disposition-distillation](https://github.com/tinmanlabsl/disposition-distillation)
+
+## Open-source proofs of concept
+
+Apache 2.0. Internal product models are not listed here.
+
+### Tinman SmolOmni (MLA)
+
+- [Tinman-SmolOmni-MLA-256M](https://huggingface.co/Tinman-Lab/Tinman-SmolOmni-MLA-256M)
+- [Tinman-SmolOmni-MLA-500M](https://huggingface.co/Tinman-Lab/Tinman-SmolOmni-MLA-500M)
+- [Tinman-SmolOmni-MLA-Toolkit](https://huggingface.co/Tinman-Lab/Tinman-SmolOmni-MLA-Toolkit)
+
+### Tinman Companion (Gemma 4)
+
+- [Tinman-gemma4-companion-merged](https://huggingface.co/Tinman-Lab/Tinman-gemma4-companion-merged) — full precision
+- [Tinman-gemma4-companion-gguf](https://huggingface.co/Tinman-Lab/Tinman-gemma4-companion-gguf) — GGUF quantized for llama.cpp
+- [Tinman-gemma4-companion-litert-lm](https://huggingface.co/Tinman-Lab/Tinman-gemma4-companion-litert-lm) — LiteRT-LM for on-device deployment
+- [Tinman-gemma4-companion-sft](https://huggingface.co/Tinman-Lab/Tinman-gemma4-companion-sft) — SFT checkpoint
+- [Tinman-gemma4-companion-dpo](https://huggingface.co/Tinman-Lab/Tinman-gemma4-companion-dpo) — DPO checkpoint
+
+## Links
+
+- [Website](https://tinmanlab.com)
+- [GitHub](https://github.com/tinmanlabsl/)