Rebuild model card, evidence visuals, and Dot research identity

README.md

@@ -8,59 +8,76 @@ language:
 - en
 datasets:
 - usedot/dot-loom-conductor-v2
+thumbnail: https://huggingface.co/usedot/Dot-Loom-Conductor-14B/resolve/main/assets/dot-loom-conductor-cover.png
 tags:
 - lora
 - peft
 - mistral
+- structured-output
 - model-routing
 - inference-orchestration
 - multi-agent
 - dot-loom
-model-index:
-- name: Dot Loom Conductor 14B
-  results:
-  - task:
-      type: text-generation
-      name: Structured multi-model routing
-    dataset:
-      type: usedot/dot-loom-conductor-v2
-      name: Dot Loom Conductor v2 held-out families
-      split: test
-    metrics:
-    - type: json-validity
-      name: JSON validity
-      value: 100.0
-    - type: policy-accuracy
-      name: Policy accuracy
-      value: 99.7
-    - type: exact-plan-accuracy
-      name: Exact plan accuracy
-      value: 90.2
-    - type: constraint-compliance
-      name: Hard-constraint compliance
-      value: 98.9
-    - type: unsafe-under-escalation
-      name: Unsafe under-escalation
-      value: 0.0
 ---
 
-![Dot Loom Conductor 14B](assets/dot-loom-conductor-hero.svg)
+<p align="center">
+  <img src="assets/dot-mark.png" width="104" alt="Dot">
+</p>
 
-# Dot Loom Conductor 14B
+<h1 align="center">Dot Loom Conductor 14B</h1>
 
-**A Mistral 14B LoRA adapter that converts a task, candidate-model profiles, and hard budgets
-into a structured Lean, Balanced, or Strict execution plan.** It decides which model writes,
-which model independently reviews, whether a finalizer is needed, and what every role may access.
-It does not answer the underlying user task.
+<p align="center"><strong>A learned control policy for budget-constrained multi-model inference.</strong></p>
 
-[Try the public routing lab](https://huggingface.co/spaces/usedot/Dot-Loom-Conductor-Lab) ·
-[Inspect the dataset](https://huggingface.co/datasets/usedot/dot-loom-conductor-v2) ·
-[Read the technical report](TECHNICAL_REPORT.md) ·
-[Reproduce the research](https://github.com/usedotai/dot-loom/tree/main/research/ministral-conductor)
+<p align="center">
+  <a href="https://huggingface.co/datasets/usedot/dot-loom-conductor-v2">Dataset</a>
+  &nbsp;·&nbsp;
+  <a href="https://huggingface.co/spaces/usedot/Dot-Loom-Conductor-Lab">Policy explorer</a>
+  &nbsp;·&nbsp;
+  <a href="TECHNICAL_REPORT.md">Technical report</a>
+  &nbsp;·&nbsp;
+  <a href="https://github.com/usedotai/dot-loom/tree/main/research/ministral-conductor">Source and receipts</a>
+</p>
+
+![Dot Loom Conductor 14B](assets/dot-loom-conductor-cover.png)
+
+Dot Loom Conductor receives a task, three candidate worker profiles, and hard call, credit, and
+latency budgets. It returns a schema-constrained **Lean**, **Balanced**, or **Strict** execution
+plan with role assignments, access rules, estimated receipts, and bounded reason codes.
+
+It does not answer the task. It decides how the task should be executed.
+
+| Release fact | Result |
+|---|---:|
+| Base checkpoint | `mistralai/Ministral-3-14B-Base-2512` |
+| Released artifact | Rank-32 PEFT LoRA adapter |
+| Training corpus | 9,000 synthetic routing traces |
+| Held-out evaluation | 1,200 family-disjoint cases |
+| Exact-plan match | 90.2% raw adapter, 90.8% with runtime guard |
+| Lift over deterministic Loom | +11.5 percentage points with guard |
+| Hard-budget violations after guard | 0 of 1,200 |
+| Training run | 52.6 minutes on 1x NVIDIA H200 SXM5 |
+
+No user prompts, production conversations, wallets, API keys, private code, or customer data were
+used in training.
+
+## Why this model exists
+
+Most agent stacks either use one model for every request or hard-code a fixed writer and reviewer
+pipeline. Both approaches spend the same amount of inference on easy and difficult work.
+
+The conductor learns a different policy:
+
+- **Lean:** one writer call when the task is low-risk and reversible
+- **Balanced:** a writer plus an independent reviewer when verification earns its cost
+- **Strict:** a writer, reviewer, and finalizer for high-consequence or synthesis-heavy work
+
+The roles are model-agnostic. A worker can be OpenAI, Claude, DeepSeek, Qwen, a Dot model, an
+OpenRouter endpoint, or a local model. Brand names never appear in the training prompts. Routing is
+based on capability, reliability, provider independence, cost, latency, risk, and hard budgets.
 
 ## Quick start
 
-The released artifact is a PEFT adapter. The pinned base model is downloaded separately.
+The release is a PEFT adapter. The pinned base model is downloaded separately.
 
 ```bash
 git clone https://huggingface.co/usedot/Dot-Loom-Conductor-14B
@@ -69,15 +86,13 @@ python -m pip install -r requirements.txt
 python inference_example.py
 ```
 
-The example loads base revision `5b0ceedbb42dff466ae60b258ba296f32da51384`, attaches this
-adapter, performs greedy decoding, and prints one JSON plan. BF16 inference requires a GPU with
-sufficient memory. The published evaluation did not test a quantized deployment.
+The example loads base revision `5b0ceedbb42dff466ae60b258ba296f32da51384`, attaches the
+adapter, runs greedy decoding, and prints one JSON plan. The published evaluation used BF16. A
+quantized deployment was not evaluated.
 
-### Input and output
+### Output contract
 
-Inputs contain task risk and complexity, three model or worker profiles, and maximum calls,
-credits, and latency. A real H200 inference from the archived demo produced this plan for a
-high-risk payment race:
+A real inference from the archived H200 demo produced this plan for a high-risk payment race:
 
 ```json
 {
@@ -107,106 +122,99 @@ high-risk payment race:
 }
 ```
 
-The roles are model-agnostic. `worker_a`, `worker_b`, and `worker_c` may represent hosted models,
-local models, or models from different providers. Production code must recompute every declared
-receipt field and enforce the hard budgets outside the model.
+Production code must recompute costs, latency, role validity, access, and hard-budget compliance
+outside the model. The adapter proposes a plan. Deterministic runtime code enforces it.
 
-## Held-out evaluation
+## Evaluation
 
-The evaluation contains 1,200 examples from eight task families absent from training and
-validation.
+The frozen test split contains 1,200 examples from eight task families that do not appear in
+training or validation: payment races, tenant isolation, webhook replay, OAuth integrity, SSRF
+egress, stream settlement, health triage, and contract risk.
 
 | Lane | JSON valid | Policy accuracy | Exact plan | Constraints met | Unsafe under-escalation | Mean regret |
 |---|---:|---:|---:|---:|---:|---:|
 | Deterministic Loom | 100.0% | 93.9% | 79.3% | 100.0% | 5.7% | 1.733 |
 | Raw Ministral 14B Base | 1.0% | 0.4% | 0.1% | 0.1% | 98.1% | 1051.175 |
 | Dot-trained raw adapter | 100.0% | 99.7% | 90.2% | 98.9% | 0.0% | 11.823 |
-| Dot-trained plus guard | 100.0% | 100.0% | 90.8% | 100.0% | 0.0% | 0.245 |
-
-The adapter improved exact-plan accuracy over deterministic Loom by 11.5 percentage points.
-Paired bootstrap 95% CI: 9.4 to 13.7 points. Exact-plan McNemar two-sided p-value: `< 1e-12`.
+| **Dot-trained plus guard** | **100.0%** | **100.0%** | **90.8%** | **100.0%** | **0.0%** | **0.245** |
 
-The raw adapter exceeded a hard budget on 13 of 1,200 proposals. Dot Loom's deterministic runtime
-guard caught all 13 and used its fallback. The guard is runtime code and is not contained inside
-the adapter weights.
+![Held-out benchmark](assets/benchmark-overview.png)
 
-![Held-out benchmark](assets/benchmark-overview.svg)
+The guarded conductor improved exact-plan match over deterministic Loom by 11.5 percentage
+points. The paired bootstrap 95% confidence interval was 9.4 to 13.7 points. The exact-plan
+McNemar two-sided p-value was `< 1e-12`.
 
-![Held-out task-family generalization](assets/family-generalization.svg)
+The raw adapter exceeded a hard budget on 13 of 1,200 proposals. The published runtime guard
+caught all 13 and selected its deterministic fallback. The guard is runtime code and is not
+contained in the adapter weights.
 
-### How to interpret the baselines
+![Held-out task-family generalization](assets/family-generalization.png)
 
-The raw Base checkpoint is included only as a pre-training capacity control. It shows that the
-structured behavior was learned during the disclosed fine-tune. It is not presented as a fair
-comparison against an instruction-tuned general assistant. Deterministic Loom is the operationally
-relevant baseline because it receives the same task, worker, and budget fields and already emits a
-valid routing plan. A future release should also evaluate instruction-tuned routing baselines and
-multiple training seeds.
+### Baseline interpretation
 
-## What the conductor decides
+The raw base checkpoint is a pre-training capacity control. It demonstrates that the structured
+routing behavior was acquired during the disclosed fine-tune. It is not presented as a fair
+comparison against an instruction-tuned general assistant.
 
-- Lean, Balanced, or Strict execution depth
-- Writer, reviewer, and finalizer assignments
-- Whether the reviewer comes from an independent provider group
-- Which previous outputs each role may access
-- Maximum calls, credits, and latency
-- Estimated combined quality and pass rate
-- Bounded machine-readable routing reasons
+Deterministic Loom is the operational baseline. It receives the same task, worker, and budget
+fields and already emits a valid plan. Future work should add instruction-tuned routing baselines,
+live downstream task outcomes, and multiple training seeds.
 
-This is a rank-32 LoRA adapter for a Mistral base model, not a 14B foundation model trained from
-scratch.
+## Training
 
-## Artifact and training receipt
+The adapter was trained on the public [Dot Loom Conductor v2
+corpus](https://huggingface.co/datasets/usedot/dot-loom-conductor-v2). Each label was selected by
+enumerating all 15 valid role assignments, rejecting plans that exceed hard budgets, and choosing
+the highest-utility feasible plan under the disclosed simulator.
 
 | Property | Value |
 |---|---:|
 | Adapter size | 560,078,656 bytes |
-| SHA-256 | `e0e6655a16a10f28cbce898e564640bf6a4a64ca84bb2014a1da3a60c5f11eda` |
+| Adapter SHA-256 | `e0e6655a16a10f28cbce898e564640bf6a4a64ca84bb2014a1da3a60c5f11eda` |
 | LoRA rank / alpha / dropout | 32 / 64 / 0.05 |
 | Trainable parameters | 139,984,896 |
 | Total base parameters | 14,085,016,576 |
 | Base revision | `5b0ceedbb42dff466ae60b258ba296f32da51384` |
-| Training hardware | 1× NVIDIA H200 SXM5, 143,771 MiB reported VRAM |
-| Training data | 9,000 synthetic examples |
-| Validation data | 900 family-disjoint synthetic examples |
+| Hardware | 1x NVIDIA H200 SXM5, 143,771 MiB reported VRAM |
+| Train / validation examples | 9,000 / 900, family-disjoint |
 | Optimizer steps / epochs | 564 / 2 |
 | Effective batch size | 32 |
 | Learning rate / schedule | `1.5e-4` / cosine, 28 warmup steps |
+| Precision | BF16 |
 | Seed | `20260716` |
 | Duration | 52.6 minutes |
 | Mean GPU utilization | 96.57% |
 | Integrated GPU energy | 0.588 kWh |
 | Selected checkpoint | Step 564, validation loss `0.070168` |
 
-BF16 LoRA targeted `q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, and
-`down_proj`. No user prompts, private code, wallets, API keys, or production conversations were
-used.
+LoRA targeted `q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, and `down_proj`.
 
-![Training dynamics](assets/training-dynamics.svg)
+![Training dynamics](assets/training-dynamics.png)
 
 ## Reproducibility
 
-- `requirements.txt` pins the exact tested runtime versions from the archived H200 environment.
 - `adapter_config.json` pins the upstream base revision.
-- `SHA256SUMS` covers the release files and adapter weights.
-- `evidence/` contains the training receipt, telemetry summary, paired statistics, checkpoint
-  selection, and scored benchmark summaries.
-- The [full research package](https://github.com/usedotai/dot-loom/tree/main/research/ministral-conductor)
-  contains the generator, raw predictions, scoring code, audit responses, tests, and charts.
+- `requirements.txt` pins the tested runtime versions from the archived H200 environment.
+- `SHA256SUMS` covers the public release files and adapter weights.
+- `evidence/` includes the training receipt, telemetry summary, checkpoint selection, paired
+  statistics, and scored benchmark summaries.
+- The [research package](https://github.com/usedotai/dot-loom/tree/main/research/ministral-conductor)
+  contains the corpus generator, raw predictions, scoring code, audit responses, tests, charts,
+  and command-line demonstrations.
 
 ## Intended use
 
-- Research on model-agnostic inference orchestration
+- Research on learned, model-agnostic inference orchestration
 - Offline comparison with deterministic routers
 - Experimental routing for coding agents and AI application backends
-- Local demonstrations of cost, latency, and verification tradeoffs
+- Local demonstrations of cost, latency, independence, and verification tradeoffs
 
-## Out of scope
+## Not intended for
 
-- Treating predicted quality as a production guarantee
+- Treating predicted quality or pass rate as a production guarantee
 - Executing financial, health, legal, security, or infrastructure changes without review
 - Granting permissions or selecting tools without a separate authorization layer
-- Using model-generated budgets without deterministic runtime recomputation
+- Trusting model-declared costs or budgets without deterministic recomputation
 
 ## Limitations
 
@@ -215,7 +223,8 @@ used.
 - The benchmark measures routing-plan generation, not final code correctness.
 - Worker capability values can drift from live provider performance.
 - The study uses one seed, one base checkpoint, and one H200 training run.
-- A production deployment must enforce all hard constraints outside the model.
+- The raw adapter violated a hard budget in 13 of 1,200 test proposals.
+- A production deployment must enforce every hard constraint outside the model.
 
 ## Citation
 
@@ -232,5 +241,5 @@ used.
 
 ## License
 
-This adapter is released under Apache 2.0. The upstream base model is also currently published
-under Apache 2.0. Users must comply with the upstream license and any applicable provider terms.
+The adapter is released under Apache 2.0. The pinned upstream base model is also published under
+Apache 2.0. Users must comply with the upstream license and any applicable provider terms.

SHA256SUMS

@@ -1,15 +1,16 @@
 d5fadc36773f0a6e03ae542aac79a3747abd47f77028ab304cb878dc2beb3ec1  .gitattributes
+d980b376af380689d7e60e8c86d2ca9b96a09215337a59670430a90df3eadaf1  CITATION.cff
 cfc7749b96f63bd31c3c42b5c471bf756814053e847c10f3eb003417bc523d30  LICENSE
 a7afc45c75c5ba304cf2bfcd9061e0832f28cb3a42cc29fea2c1284d6def2964  NOTICE
-e345c30dcfee5694236231b9d1c1b2950e0d5ba4143fc27810f5c695065d39e8  README.md
-f1d4dcc5d2b118ee603e3ca389d52251f14c29c82e70867b1d1ec36d5e47b6eb  TECHNICAL_REPORT.md
-d980b376af380689d7e60e8c86d2ca9b96a09215337a59670430a90df3eadaf1  CITATION.cff
+5d7a0c27387950ac75b835844c2092f59e0ab13071a5b7009f151267827233ed  README.md
+e5465da8a0115cb226fc36e640ef534e761f1adec88c1b3345c33d57c483a443  TECHNICAL_REPORT.md
 39d3e713e98dc32ad4b663d58e0ffff719d91550d08545076d5094d804ce5e05  adapter_config.json
 e0e6655a16a10f28cbce898e564640bf6a4a64ca84bb2014a1da3a60c5f11eda  adapter_model.safetensors
-82ec02400973e21d6fae7f7b2fe0c8e5f7b7602f57a315b26dd27f4100cfa165  assets/benchmark-overview.svg
-7030b9293dcbdaf586b115c041bb1e46722cae6ac9d3fdb213dd5ee77efa59a4  assets/dot-loom-conductor-hero.svg
-271d5d2f9d5ad04cec6d7fde6b64ab40b88ad54ba3e39c56769e719176c6dead  assets/family-generalization.svg
-6853cf0e379f5751c4b3a2ee2e267d8645428643c29e28b75080936adab699c4  assets/training-dynamics.svg
+1449b4e04afd787b8594d36bc522c28dbbafd5d9bf40ad0973bb20f3954e84c7  assets/benchmark-overview.png
+50c7c38b901aababb16ccaf599bcb95a36a2ad9bfd41bfcf8210bff46b0f8631  assets/dot-loom-conductor-cover.png
+d8b632da85953ed75b11f9e77d998ad01e89e042a14abd8f3990fc5fc2442b68  assets/dot-mark.png
+4d5dfe31eea6fcd2951805c4fc2999962659521dd89b90e3c9cc6f90ce27eeb4  assets/family-generalization.png
+e72de3a667dbc2ed166224d6297b863e452d02340766c1c03852dcbd4ac32b10  assets/training-dynamics.png
 433ffcf387183f580ce03978f4d2dff0d518416f67083c481c5073dcf3d7d31d  evidence/checkpoint-selection.json
 ced5263c61c6732af7da8fb7c03a227e71ff3f129b54b3a1965ede6c1ba0d33a  evidence/deterministic_scored.summary.json
 dbe6682d157932edeafddf20303c0f0e4380bf467af7934e1070a4862519ec3f  evidence/gpu-telemetry.summary.json

TECHNICAL_REPORT.md

@@ -1,8 +1,12 @@
-# Dot Loom Conductor 14B
+<p align="center">
+  <img src="assets/dot-mark.png" width="84" alt="Dot">
+</p>
 
-## Learned budget-constrained multi-model routing
+<h1 align="center">Dot Loom Conductor 14B</h1>
 
-Dot R&D, July 2026
+<p align="center"><strong>Technical report: learned budget-constrained multi-model routing</strong></p>
+
+<p align="center">Dot R&amp;D · July 2026</p>
 
 ### Abstract
 
@@ -97,7 +101,7 @@ guard is not encoded in the adapter weights.
 
 - [Model and adapter](https://huggingface.co/usedot/Dot-Loom-Conductor-14B)
 - [Dataset](https://huggingface.co/datasets/usedot/dot-loom-conductor-v2)
-- [Interactive routing lab](https://huggingface.co/spaces/usedot/Dot-Loom-Conductor-Lab)
+- [Reference policy explorer](https://huggingface.co/spaces/usedot/Dot-Loom-Conductor-Lab)
 - [Source, tests, raw predictions, and receipts](https://github.com/usedotai/dot-loom/tree/main/research/ministral-conductor)
 
 ### Citation