Card: reviewed release text. HealthBench Hard 39.7 (not 39.8); MedCalc 86.6% is a separate calculator-enabled configuration, not MIA 1; charts from the release package

.gitattributes

@@ -34,3 +34,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 mia1-hero-16x9.png filter=lfs diff=lfs merge=lfs -text
+charts/mia1-healthbench-professional.png filter=lfs diff=lfs merge=lfs -text
+charts/mia1-healthbench-hard.png filter=lfs diff=lfs merge=lfs -text
+charts/mia1-medcalc-bench.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,54 +1,117 @@
 ---
-license: other
-license_name: mia-1-research-release
-pipeline_tag: text-generation
 language:
 - en
 tags:
 - medical
 - clinical-reasoning
-- healthbench
-- medcalc-bench
-- qwen3.8
+- evaluation
 ---
 
 # MIA 1, Medical Intelligent Agent
 
 **Dr. Max Health** · Research release · October 2026
 
-![MIA 1](mia1-hero-16x9.png)
+MIA 1 is a research model, not a medical device; not for diagnosis or treatment decisions.
 
-MIA 1 is a 27B clinical reasoning model built on Qwen3.8-27B with Dr. Max Health's clinical reasoning harness and medical tools. It is our first step toward AI that reasons through real clinical cases at the level of an experienced doctor.
+## What is released
+
+This page is a model card only. **No model weights, adapters or other model files are published here.**
+
+MIA 1 is not a fine-tuned checkpoint. It runs the unmodified open Qwen3.8-27B model with Dr. Max Health's clinical
+reasoning harness, applied at inference time:
+
+- instructions written for clinicians, and
+- an answer step that drafts, checks the draft against what a complete answer needs, and tightens it.
+
+The model's weights are not changed. The harness is not part of this release.
 
 ## Highlights
 
-- **Clinical calculations:** 86.6% on MedCalc-Bench with its medical tools, within 2 points of Claude Opus 5.5.
-- **Hard cases:** within 4 points of Claude Opus 5.5 on HealthBench Hard.
-- **Built for clinicians:** 54.1 vs 41.3 for its base model on HealthBench Professional, with answers half as long.
-- **Efficient:** 27B parameters, running on a single GPU node.
+- **Built for clinicians.** 54.1 on HealthBench Professional, up from 41.3 for Qwen3.8-27B on its own, with answers
+  less than half as long on average. Claude Opus 5.5, run by us on the same questions, scores 65.1.
+- **Hard cases.** 39.7 on a 100-question sample of HealthBench Hard, against 37.6 for Qwen3.8-27B and 43.8 for
+  Claude Opus 5.5.
+- **Compact.** 27 billion parameters.
+- **Clinical calculations.** Separately, our Qwen-based calculator-enabled configuration scored 86.6% on MedCalc-Bench
+  Verified (75.7% without tools; Claude Opus 5.5 88.2% without tools). It is a different configuration from the MIA 1
+  setup evaluated on HealthBench.
+
+## Intended use
 
-## Results
+MIA 1 is released for research and evaluation. It is not yet a certified medical device; we are working toward
+clinical decision support that meets medical-device requirements.
 
-![MIA 1 results](mia1-results-16x9.png)
+Intended users are researchers and partner institutions evaluating clinical reasoning systems together with
+Dr. Max Health.
 
-| Benchmark | MIA 1 | Qwen3.8-27B | Claude Opus 5.5 |
-|---|---|---|---|
-| HealthBench Professional | **54.1** | 41.3 | 65.1 |
-| HealthBench Hard (100-question sample) | **39.8** | 37.6 | 43.8 |
-| MedCalc-Bench (accuracy %) | **86.6** (with calculator tools) | 75.7 | 88.2 |
+## Out-of-scope use
 
-### How we evaluated
+- Diagnosis, treatment, triage or any other decision about a real patient.
+- Medical advice to patients or the public.
+- Use as, or inside, a medical device or a clinical decision support system in patient care.
+- Emergencies, or any setting where an unchecked answer could harm a person.
 
-HealthBench Professional (325 held-out examples) and HealthBench Hard (100-question sample): scores are length-adjusted following the HealthBench Professional paper and are the mean of two blind model graders (Qwen 27B and Claude Opus 5.5) using the official rubrics; per-grader and raw scores are in the technical report. MedCalc-Bench (1,100 test items): scored by exact answer with the same rule for every model; MIA 1 uses its calculator tools and its score is the average over 8 answers per question, Claude Opus 5.5 and Qwen3.8-27B answer without tools (Opus 1 answer per question). All models were run by Dr. Max Health on the same questions. These internal results are not comparable to official HealthBench leaderboard numbers. Technical report to follow.
+## Evaluation
 
-## Intended use
+All models were run by Dr. Max Health on the same questions.
+
+**HealthBench** (length-adjusted score, 0-100; mean of two model graders)
+
+| Model | HealthBench Professional (325 held out) | HealthBench Hard (100-question sample) |
+|---|---|---|
+| MIA 1 | 54.1 | 39.7 |
+| Claude Opus 5.5 | 65.1 | 43.8 |
+| Qwen3.8-27B | 41.3 | 37.6 |
 
-MIA 1 is released for research and evaluation. It is not yet a certified medical device; we are working toward clinical decision support that meets medical-device requirements.
+**MedCalc-Bench Verified** (accuracy on 1,100 test questions, one scoring rule for every model). The first row is a
+separate calculator-enabled configuration, not the MIA 1 setup evaluated on HealthBench.
+
+| Configuration | MedCalc-Bench Verified |
+|---|---|
+| Qwen3.8-27B + calculator tools | 86.6% |
+| Claude Opus 5.5, no tools | 88.2% |
+| Qwen3.8-27B, no tools | 75.7% |
+
+![HealthBench Professional, length-adjusted score on 325 questions held out from development, scale 0-100: MIA 1 54.1, Claude Opus 5.5 65.1, Qwen3.8-27B 41.3.](charts/mia1-healthbench-professional.png)
+
+![HealthBench Hard, length-adjusted score on a 100-question sample, scale 0-100: MIA 1 39.7, Claude Opus 5.5 43.8, Qwen3.8-27B 37.6.](charts/mia1-healthbench-hard.png)
+
+![MedCalc-Bench Verified, accuracy on 1,100 test questions under one pre-registered scoring rule, scale 0-100 %: Qwen3.8-27B + calculator tools 86.6 %, Claude Opus 5.5 without tools 88.2 %, Qwen3.8-27B without tools 75.7 %.](charts/mia1-medcalc-bench.png)
+
+### How we evaluated
+
+HealthBench Professional (325 examples held out from our development selection) and HealthBench Hard (100-question
+sample): scores are length-adjusted following the HealthBench Professional paper and are the mean of two blind model
+graders (Qwen 27B and Claude Opus 5.5) using the official rubrics; per-grader and raw scores will be in the technical
+report. MedCalc-Bench Verified (1,100 test items): one pre-registered scoring rule for every model; the
+calculator-enabled Qwen3.8-27B configuration uses calculator tools, and its score and the score of Qwen3.8-27B without
+tools are averages over 8 answers per question; Claude Opus 5.5 answers once per question, without tools. All models
+were run by Dr. Max Health on the same questions. These internal results are not comparable to official HealthBench
+or MedCalc-Bench numbers.
+
+## Limitations
+
+- These are benchmark results, not clinical validation. MIA 1 has not been clinically validated.
+- Claude Opus 5.5 scores higher than MIA 1 on both HealthBench benchmarks (mean of the two graders).
+- HealthBench Hard is a 100-question sample, so small differences there are within noise.
+- The two graders are models, not clinicians. One grader shares MIA 1's base model; the other is Claude Opus 5.5,
+  which is also one of the compared models.
+- MIA 1 inherits the limits of Qwen3.8-27B: it can give confident wrong answers, miss recent guidance and reflect
+  biases in its training data. Shorter answers can leave out details a clinician would want.
+- We have not evaluated MIA 1 outside the benchmarks listed here.
 
 ## Access
 
-Weights are not public. For research or clinical partnerships, contact us via [drmaxhealth.de](https://drmaxhealth.de).
+No weights are published. Hospitals, clinics and research partners: contact us via drmaxhealth.de.
+
+## Licence and attribution
+
+Built on Qwen3.8-27B.
+
+Qwen3.8-27B is released by the Qwen team under the Apache License 2.0
+([Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B)). This page redistributes no Qwen3.8-27B weights or
+files. The Qwen name identifies the base model only and does not imply endorsement by the Qwen team.
 
-## Attribution
+## Technical report
 
-Built on Qwen3.8-27B by the Qwen team (Alibaba Cloud), used under its license.
+A technical report will follow.