Describe the model for users: sources, usage, results, and limitations

README.md

@@ -33,58 +33,47 @@ datasets:
 
 # Dory 1.0 4B Typed Decisions
 
-**Release: `state-32k`.** A Dory fine-tune for fast typed decisions, action selection, and state tracking in software agents.
-This release continues `actions-48k` with 32,000 questions about state tracking, planning, human judgements, and earlier tasks.
-On 512 reserved new questions, native NVFP4 accuracy improves from **46.48% to 75.59%**.
-The main gains come from state tracking and planning. Human-judgement accuracy remains almost unchanged.
+A fine-tune of [NVIDIA Dory 1.0 4B](https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16) for decisions over application-defined answers.
+Supply context, a question, and the allowed answers. The model scores those answers without generating an explanation.
 
-The base model is [NVIDIA Dory 1.0 4B](https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16).
-An application supplies a state, a question, and permitted answers.
-The model scores those answers and returns a binary probability, a named choice, or an ordered score.
+Use it for workflow routing, tool-call validation, evidence checks, rubric scoring, and action selection.
+Questions and answer meanings are supplied at request time, rather than fixed to a single classification task.
 
-This experiment belongs to [Eider](https://github.com/rdaum/eider), a Rust/CUDA inference server for NVIDIA DGX Spark.
-Eider provides the evaluated prompt format and an optimized serving implementation.
-The learned behaviour can also run in another engine that implements Dory and the same decision format.
+The model was trained and evaluated with [Eider](https://github.com/rdaum/eider), a Rust/CUDA inference server for NVIDIA DGX Spark.
+The GGUF also works with llama.cpp's native decision endpoint, using its Dory support branch as described below.
 
-## Available formats
+## Model and formats
 
-All three formats contain the same fine-tuned `state-32k` model. They share this model card and training history.
-
-| Format | Location | Use |
+| Property | Value |
+|---|---|
+| Base model | NVIDIA Dory 1.0 4B |
+| Adaptation | Rank-8 LoRA, merged into the supplied weights |
+| Recurrent computation | Two internal loops per inference |
+| Outputs | Binary probability, named-choice distribution, or ordered-score distribution |
+| Probability calibration | No fitted calibration profile |
+| Licence | [OpenMDW-1.1](LICENSE) |
+
+| Format | Files | Use |
 |---|---|---|
-| BF16 safetensors | Root: eight shards and `model.safetensors.index.json` | Merged weights for a runtime with Dory support. |
-| NVFP4 safetensors | [`nvfp4/`](nvfp4) | Eider's evaluated packed 4-bit checkpoint, with its config and tokenizer. |
-| Q8_0 GGUF | [`dory-decisions-Q8_0.gguf`](dory-decisions-Q8_0.gguf), 4.23 GB | llama.cpp generation and native `/v1/systemone` decisions; instructions below. |
-| Rank-8 LoRA adapter | [`adapter/`](adapter/README.md) | Standalone training changes for the original NVIDIA BF16 model. |
-
-BF16 is the merged full-precision release; NVFP4 and Q8_0 are its quantized exports.
-
-This release has **no fitted probability-calibration profile**.
+| BF16 safetensors | Eight root-level shards and [index](model.safetensors.index.json) | Merged weights for runtimes with Dory support. |
+| NVFP4 safetensors | [`nvfp4/`](nvfp4) | Packed 4-bit weights, configuration, and tokenizer for Eider. |
+| Q8_0 GGUF | [`dory-decisions-Q8_0.gguf`](dory-decisions-Q8_0.gguf), 4.23 GB | 8-bit weights for llama.cpp. |
+| LoRA adapter | [`adapter/`](adapter/README.md) | Standalone adaptation for the original NVIDIA BF16 checkpoint. |
 
-## What this is for
+All formats contain the same fine-tuned model. NVFP4 and Q8_0 are quantized from the merged BF16 weights.
+The adapter uses the supplied Python loader's tensor format, not PEFT format.
+Apply it only to the original NVIDIA checkpoint; the release weights already include it.
 
-The intended use is a small, local decision component inside a larger application or agent.
-Examples include tool-call validation, workflow routing, document-grounding judgements, and assessments against a rubric.
-The application defines the question and answer meanings at request time, so the model can serve many classification tasks.
+## How typed decisions work
 
-The objective is a fast “System 1” judgement without an autoregressive explanation for every decision.
-The application can use the answer distribution to choose an action or request further analysis.
-These are intended uses, not guarantees of accuracy on every task.
-
-| Answer type | Meaning | Example |
+| Answer type | Result | Example |
 |---|---|---|
-| Binary (`noul` in the API) | Estimated probability of the supplied true condition | Does this proposed call satisfy the request? |
-| Choice | Distribution over named options and the highest-scoring option | Which queue receives this ticket? |
+| Binary (`noul`) | Probability of the supplied true condition | Does this tool call satisfy the request? |
+| Choice | Distribution over named options | Which queue should receive this ticket? |
 | Ordered score | Distribution over rubric levels and their expected index | How well does this response satisfy the rubric? |
 
-## How the decision works
-
-The model keeps the original Dory architecture and vocabulary output head.
-LoRA training changes attention and MLP weights. It adds no separate classification head.
-The model uses **two recurrent loops**, which repeat internal computation through Dory layers without generating a chain of thought.
-
-The prompt combines the state, evaluation instructions, a question, and JSON-described options with single-token labels such as `A` and `B`.
-This excerpt illustrates the question portion:
+The prompt contains context, a question, and answer descriptions mapped to single-token labels such as `A` and `B`.
+For example:
 
 ```text
 Question: Is the response supported by the supplied evidence?
@@ -94,206 +83,104 @@ A: {"option":"true","description":"Every claim is supported by the evidence."}
 B: {"option":"false","description":"At least one claim lacks support."}
 ```
 
-The complete prompt also uses the Dory chat template, disabled explicit reasoning, and an assistant prefix of `Answer:\n`.
-At that position, the runtime reads the next-token logits for the permitted labels.
-It applies softmax across those labels and maps the resulting distribution to the requested answer type.
-These are conditional probabilities over the supplied options. They do not establish that the option list covers every possible answer.
+The runtime reads the next-token logits for those labels and applies softmax over the allowed answers.
+It returns the resulting distribution without generating answer tokens.
+The model uses its ordinary vocabulary head, with no separate classification head.
+Dory's two recurrent loops repeat internal computation without producing a chain of thought.
 
-The application enforces the response type and option membership.
-The model supplies the judgement within that structure.
-This training does not establish support for arbitrary JSON schemas or open-ended structured generation.
+The runtime enforces the answer type and option membership. The model supplies the judgement.
+This method supports the answer types above, rather than arbitrary JSON schemas.
+The exact prompt and label mapping matter; the GGUF includes the decision template.
 
-## How Eider serves decisions
+### Eider serving
 
-Eider exposes a `/v1/decisions` endpoint. A request contains a shared state and one or more typed questions.
-Each question supplies its permitted answers and their meanings.
-
-1. Eider renders the state and each question with the decision prompt described above.
-2. It processes the shared state once and evaluates separate question branches. Questions cannot attend to each other.
-3. It reads only the vocabulary-head rows for the answer labels at the end of each branch.
-4. It applies softmax over those labels and returns the requested binary probability, choice distribution, or ordered score.
-
-The decision path generates no answer tokens or explanation. Dory still performs its two internal recurrent loops.
-Eider batches question branches, reuses cached state prefixes, and uses a compact label readout to reduce repeated work.
-These are serving optimizations. The weights contain the learned decision behaviour and need no Eider-specific classification head.
-Another Dory runtime can reproduce the method with the same prompt, label mapping, recurrent settings, and probability calculation.
-
-A fitted calibration profile can adjust the distributions after inference, but this release does not include one.
-The probabilities describe the supplied options; they are not independently verified confidence estimates.
+Eider exposes `/v1/decisions` for a shared context and multiple typed questions.
+It processes the context once, then evaluates independent question branches in batches.
+Questions cannot attend to each other. Eider reuses cached context and reads only the vocabulary rows for answer labels.
+These optimizations reduce repeated work without requiring an Eider-specific classification head in the weights.
 
 ## Training data
 
-The adapter carries five training stages. The tables below list all source families, grouped by purpose.
-A source can appear in several stages through replay.
-Pinned versions, source-file hashes, recorded source licences, and split rules are in [training-data.json](training-data.json).
-This repository contains model weights and metadata; they do not redistribute the source datasets.
-
-### Latest stage: state tracking, planning, and human judgements
-
-| Source | Questions in this stage | What the model learns |
-|---|---:|---|
-| [SCONE](https://nlp.stanford.edu/projects/scone/) — Alchemy and Tangrams | 6,000 | Track a world state through a complete instruction sequence and answer from the recorded final state. |
-| Independent Eider planning generators, motivated by [PlanBench](https://github.com/karthikv792/LLMs-Planning/tree/main/plan-bench) | 8,000 | Choose optimal next moves and assess legality or progress in block stacking and package delivery. |
-| [NVIDIA HelpSteer2](https://huggingface.co/datasets/nvidia/HelpSteer2) | 2,000 | Learn correctness, helpfulness, coherence, and preference judgements from individual human ratings. |
-| Replay of earlier action questions | 8,000 | Retain action-selection behaviour from the four environment sources below. |
-| Replay of earlier general decision questions | 8,000 | Retain earlier decision, tool, document, reasoning, and code behaviour. |
-| **Total** | **32,000** | **16,000 new questions and 16,000 replay questions.** |
-
-SCONE supplies recorded state labels; its Scene domain is excluded pending an alignment audit.
-The planning generators and breadth-first search implementation are independent Eider code.
-No published PlanBench instances or generator code were copied.
-Planning targets share mass among equally optimal actions.
-HelpSteer2 targets use individual rating and preference frequencies, preserving disagreement and ties.
-Inputs exclude annotator ratings and justifications.
+The following sources provide the decision questions and targets used for fine-tuning.
+[training-data.json](training-data.json) records source versions, licences, hashes, sampling, and split rules.
+Source datasets are not redistributed here.
 
-### Earlier action and environment training
-
-These sources supplied the 24,000 new questions in the preceding 48,000-question stage; its other 24,000 questions were replay.
-
-| Source | New questions in that stage | Preparation |
-|---|---:|---|
-| [JAT BabyAI](https://huggingface.co/datasets/jat-project/jat-dataset) | 8,000 | Successful expert navigation, doors, keys, and object placement. Maps use past observations only. |
-| [ALFWorld through Early Experience](https://huggingface.co/datasets/osunlp/early-experience) | 6,000 | Recorded expert commands and admissible actions, with the supplied observation history. |
-| [Competitive Game Dataset](https://huggingface.co/datasets/yyyyyyjjjjzzz/competitive-game-dataset) — Connect Four | 4,000 | Solver-labelled legal moves, with equal target mass for tied best moves. |
-| [ScienceWorld gold paths](https://github.com/allenai/ScienceWorld/tree/e8216d6044e8e39be9fcb185e3b2dfb602584b52/goldpaths) | 6,000 | Select among 17 operation types using the goal, room, inventory, and past results. |
-
-ScienceWorld training concerns operation types, not complete object-specific commands.
-Expert imitation labels do not establish that every alternative action would fail.
-
-### General typed decisions and controlled scenarios
-
-| Source | Selected data and use |
-|---|---|
-| [Open-Jev](https://huggingface.co/datasets/ZefanCai/Open-Jev) | `release-v2-redistributable`: control, workflow, reasoning, and game questions. Original target distributions are preserved. |
-| [SargeDev/jev-distill-corpus-v3](https://huggingface.co/datasets/SargeDev/jev-distill-corpus-v3) | Only `yuri_v3` rows: diverse decision questions with supplied probability targets. |
-| [Procedural Typed Decisions](https://huggingface.co/datasets/tasksource/procedural-typed-decisions) | Rule-based and exact-probability tasks covering retrieval, state tracking, arithmetic, routing, and planning. |
-| [Open-Jev-v1.1](https://huggingface.co/datasets/ZefanCai/Open-Jev-v1.1) | `community-hard-mix-v2-redistributable`: WANLI-derived inference and controlled tool, document, support, and rubric scenarios. |
-
-### Tools, documents, reading, mathematics, and code
-
-| Source | Preparation |
-|---|---|
-| [ToolACE](https://huggingface.co/datasets/Team-ACE/ToolACE) | Judge reference tool calls and deliberately invalid variants that violate the supplied argument schema. |
-| [RAGTruth](https://github.com/ParticleMedia/RAGTruth) | Evidence-grounding questions from human span annotations, responses, and their supplied documents. |
-| [RACE](https://huggingface.co/datasets/ehovy/race) | Reading-comprehension questions with original options and labels. |
-| [GSM8K](https://huggingface.co/datasets/openai/gsm8k) | Numeric answer selection with generated distractors. Reference reasoning is excluded from the input. |
-| [Devign through CodeXGLUE](https://huggingface.co/datasets/google/code_x_glue_cc_defect_detection) | Binary C-function defect questions using the source labels. |
-
-### Data separation and target meaning
-
-Complete source groups stay together: episodes, games, goals, conversations, documents, or prompts.
-Planning groups include equivalent problems under block, package, and location renaming.
-Public test examples never enter training, and previous evaluation reservations remain excluded.
-New source groups contribute at most four questions each.
-The latest stage rejects inputs above 1,152 tokens instead of truncating them.
-
-**No JevMan or Pac-Man states, layouts, trajectories, rewards, or labels are used.**
-Other game and environment training is listed explicitly above.
-These results do not measure full-episode game performance.
-
-Training targets have different meanings across sources: recorded answers, expert actions, optimal actions, human votes, or supplied distributions.
-Optimal-action mass is not a win probability, and human vote frequencies do not establish outcome probabilities.
-
-## Training recipe and history
-
-Training starts from NVIDIA's BF16 revision `b366b87f762ea334ba3eb55912031baba1515810`.
-Rank-8 LoRA with alpha 16 updates 16,220,160 adapter parameters across 288 attention and MLP projections.
-The original weights remain frozen. Training and evaluation use two recurrent loops.
-The loss is soft-target conditional cross-entropy plus 0.5 times Brier error over the permitted answers.
-
-| Stage | Training questions | Mixture | Learning rate |
-|---|---:|---|---:|
-| Initial decisions | 49,152 | 24,576 Open-Jev + 24,576 Yuri v3 | 0.0001 |
-| Broader decisions | 65,536 | 24,576 procedural + 16,384 WANLI-derived + 8,192 controlled + 16,384 replay | 0.00002 |
-| Tools and documents | 32,000 | 4,000 tools + 4,000 documents + 8,000 general + 16,000 replay | 0.00001 |
-| Action selection | 48,000 | 24,000 environment questions + 24,000 replay | 0.00001 |
-| **State tracking, planning, and judging** | **32,000** | **16,000 new + 16,000 replay** | **0.00001** |
-
-Replay repeats earlier training questions to retain previous behaviour.
-These are questions processed per stage, not a total of unique examples.
-The latest stage continues the previous adapter with a fresh optimizer.
-It completed 8,000 updates of four questions each on one GB300, with about 45 minutes of training compute.
-
-## Accuracy evaluation
-
-The following results use **native Eider NVFP4 inference without probability calibration**.
-The previous and new adapters use the same frozen Eider executable, prompt format, and two-loop configuration.
-The parent NVFP4 export is byte-identical to the published `actions-48k` weights.
-
-Checkpoint selection uses 1,024 validation questions: 512 new questions and 512 earlier questions.
-The criterion is lowest Brier error, then cross-entropy, then earliest step.
-The unchanged parent and steps 2,000, 4,000, and 8,000 were eligible; step 8,000 won.
-Validation accuracy improved from 60.06% to 76.17%, and Brier error fell from 0.5027 to 0.2779.
-The 512 new test questions were scored after selection.
-
-| Question set | Model | Questions | Accuracy | Brier ↓ | Cross-entropy ↓ |
-|---|---|---:|---:|---:|---:|
-| New state, planning, and judging questions | Previous `actions-48k` NVFP4 | 512 | 46.48% | 0.7261 | 2.2838 |
-| New state, planning, and judging questions | This `state-32k` NVFP4 | 512 | 75.59% | 0.2706 | 1.0355 |
-| Earlier questions, reused | Previous `actions-48k` NVFP4 | 2,048 | 79.54% | 0.2521 | 0.7409 |
-| Earlier questions, reused | This `state-32k` NVFP4 | 2,048 | 79.35% | 0.2642 | 0.7846 |
-| General questions, reused | Previous `actions-48k` NVFP4 | 89 | 64.04% | 0.4997 | 1.0697 |
-| General questions, reused | This `state-32k` NVFP4 | 89 | 65.17% | 0.4791 | 1.0044 |
-
-The new test breaks down as follows:
-
-| Source | Questions | Previous NVFP4 | This NVFP4 | Brier, previous → current |
-|---|---:|---:|---:|---:|
-| SCONE state tracking | 192 | 42.19% | 75.52% | 0.8904 → 0.3779 |
-| Block-stacking and delivery planning | 192 | 48.44% | 92.19% | 0.6956 → 0.1909 |
-| HelpSteer2 human judgements | 128 | 50.00% | 50.78% | 0.5253 → 0.2290 |
-
-State tracking and planning improve substantially.
-HelpSteer2 accuracy changes by one answer; its predicted distributions move closer to the supplied human rating distributions.
-Earlier-question accuracy changes by four answers out of 2,048, while Brier and cross-entropy become slightly worse.
-The 89-question general check improves by one answer and is too small to establish broad generalization gains.
-
-Accuracy accepts any option with maximum target probability, including ties.
-Brier and cross-entropy compare the full predicted and target distributions.
-The earlier and general sets are reused development checks, not independent final benchmarks.
-Converted questions do not follow the original datasets' complete benchmark protocols.
-
-**These scores apply to NVFP4.** Neither the matching BF16 merge nor the Q8_0 GGUF has received a separate full accuracy evaluation.
-The original unfine-tuned NVIDIA model was not rescored in this comparison.
-[Full results](evaluation/nvfp4-comparison.json) include every candidate, answer type, and source.
-[Runtime metadata](evaluation/runtime.json) records the executable and script hashes.
-
-## Probability calibration
-
-No calibration profile has been fitted for `state-32k`.
-The previous profile was removed from the current release because it belongs to different weights.
-The 1,024-question calibration split remains separate from training, checkpoint selection, and the new test.
-Uncalibrated softmax values are conditional scores over the permitted answers; they are not proven real-world outcome probabilities.
-
-## Portability and files
-
-Another engine can implement the same method using Dory, its tokenizer and prompt, two loops, and selected-label scoring.
-Eider-specific CUDA kernels are not part of the learned weights.
-Eider accelerates serving with shared state processing, independent question branches, batching, CUDA graphs, and selected output logits.
-A different prompt, label mapping, loop count, or quantization can change the results.
-
-The root contains eight indexed BF16 safetensors shards, merged from the original NVIDIA weights and the selected adapter.
-Every saved tensor was checked against its CPU merge result. These merged BF16 files have not received a separate accuracy pass.
-
-The matching standalone LoRA adapter is in `adapter/`. The complete NVFP4 package also includes a copy.
-The adapter uses Eider tensor names rather than PEFT format; the supplied Python module loads that format.
-The adapter applies to the original NVIDIA BF16 checkpoint. Do not apply it again to the merged release weights.
-Its SHA-256 is `60723ac470780ac7e8bfe7785c23665be2d5c5d6653cbd2fad653523214aea7d`.
-
-The NVFP4 release quantizes our fine-tuned BF16 weights; it is not NVIDIA's original NVFP4 model.
-Its export uses packed E2M1 weights, E4M3 block scales, an F32 tensor scale, and a ModelOpt-compatible layout.
-The quantizer searches all 126 positive finite E4M3 scales for minimum block reconstruction error after BF16 merging.
-[Export metadata](evaluation/export.json) records the recipe, source hashes, and per-projection statistics.
+| Area | Source | Training task |
+|---|---|---|
+| General decisions | [Open-Jev](https://huggingface.co/datasets/ZefanCai/Open-Jev), `release-v2-redistributable` | Control, workflow, reasoning, and game questions with supplied target distributions. |
+| General decisions | [Jev Distill Corpus v3](https://huggingface.co/datasets/SargeDev/jev-distill-corpus-v3), `yuri_v3` subset | Diverse typed questions with supplied probability targets. |
+| General decisions | [Procedural Typed Decisions](https://huggingface.co/datasets/tasksource/procedural-typed-decisions) | Rule-based retrieval, arithmetic, routing, state tracking, and planning. |
+| Inference and controls | [Open-Jev-v1.1](https://huggingface.co/datasets/ZefanCai/Open-Jev-v1.1), `community-hard-mix-v2-redistributable` | WANLI-derived inference and controlled tool, document, support, and rubric scenarios. |
+| Tools | [ToolACE](https://huggingface.co/datasets/Team-ACE/ToolACE) | Assess reference tool calls and variants with invalid arguments. |
+| Evidence grounding | [RAGTruth](https://github.com/ParticleMedia/RAGTruth) | Judge response support using documents and human span annotations. |
+| Reading | [RACE](https://huggingface.co/datasets/ehovy/race) | Select answers to reading-comprehension questions. |
+| Mathematics | [GSM8K](https://huggingface.co/datasets/openai/gsm8k) | Select numeric answers from generated alternatives, without reference reasoning in the input. |
+| Code | [Devign through CodeXGLUE](https://huggingface.co/datasets/google/code_x_glue_cc_defect_detection) | Classify C-function defects. |
+| Navigation | [JAT BabyAI](https://huggingface.co/datasets/jat-project/jat-dataset) | Select expert actions involving navigation, doors, keys, and object placement. |
+| Household tasks | [ALFWorld through Early Experience](https://huggingface.co/datasets/osunlp/early-experience) | Select expert commands from admissible actions and observation history. |
+| Board games | [Competitive Game Dataset](https://huggingface.co/datasets/yyyyyyjjjjzzz/competitive-game-dataset), Connect Four | Select solver-labelled moves, preserving ties between optimal actions. |
+| Interactive tasks | [ScienceWorld gold paths](https://github.com/allenai/ScienceWorld/tree/e8216d6044e8e39be9fcb185e3b2dfb602584b52/goldpaths) | Select among 17 operation types from the goal, room, inventory, and past results. |
+| State tracking | [SCONE](https://nlp.stanford.edu/projects/scone/), Alchemy and Tangrams | Infer final world states from instruction sequences. |
+| Planning | Independent Eider generators motivated by [PlanBench](https://github.com/karthikv792/LLMs-Planning/tree/main/plan-bench) | Assess legality and choose optimal moves in block stacking and package delivery. |
+| Human judgements | [NVIDIA HelpSteer2](https://huggingface.co/datasets/nvidia/HelpSteer2) | Learn correctness, helpfulness, coherence, and preference distributions from human ratings. |
+
+Planning examples use independently generated problems and search-derived targets, not published PlanBench instances or generator code.
+Training and evaluation keep related source groups together, such as episodes, games, documents, conversations, and equivalent planning problems.
+Public test examples are excluded from training.
+
+### Training method
+
+Rank-8 LoRA with alpha 16 updates 16,220,160 parameters across 288 attention and MLP projections; the base weights remain frozen during training.
+Training uses two recurrent loops and soft-target conditional cross-entropy plus 0.5 times Brier error over the allowed answers.
+The adapter is merged into BF16 before quantization.
+NVFP4 uses packed E2M1 weights with E4M3 block scales and an F32 tensor scale in a ModelOpt-compatible layout.
+[Export metadata](evaluation/export.json) records the quantization recipe.
+
+## Evaluation
+
+These results use **Eider NVFP4 inference with two recurrent loops and no probability calibration**.
+The 512-question test set was scored after checkpoint selection on a separate 1,024-question validation set.
+
+| Test task | Questions | Accuracy | Brier ↓ |
+|---|---:|---:|---:|
+| SCONE state tracking | 192 | 75.52% | 0.3779 |
+| Block-stacking and delivery planning | 192 | 92.19% | 0.1909 |
+| HelpSteer2 human judgements | 128 | 50.78% | 0.2290 |
+| **Combined** | **512** | **75.59%** | **0.2706** |
+
+Accuracy counts any answer tied for the highest target probability as correct.
+Brier error compares the complete predicted and target distributions.
+These are adapted decision tasks, not scores under the source datasets' original benchmark protocols.
+
+Additional reused development sets scored 79.35% on 2,048 questions and 65.17% on 89 general questions.
+They are not independent final tests. [Full evaluation records](evaluation/nvfp4-comparison.json) contain the per-source results and checkpoint comparisons.
+Neither BF16 nor Q8_0 has received a separate full accuracy evaluation.
+
+The Q8_0 GGUF passed prompt and answer-label parity checks and a 24-request GPU compatibility check across all three answer types.
+On GB10, short requests with changing context had median latency of 62 ms for one question and 191 ms for three questions.
+Each median used five timed requests. These small checks establish API compatibility, not general accuracy or production throughput.
+[GGUF test evidence](gguf/provenance.json) records the setup and results.
+
+## Limitations
+
+Probabilities are normalized over the supplied answers and have no fitted calibration profile.
+They are not verified confidence estimates or real-world outcome probabilities.
+Targets vary by source: factual labels, expert actions, optimal actions, human votes, or supplied distributions.
+A distribution over optimal actions, for example, is not a probability of winning.
+
+The evaluation does not establish general chat quality, reliable decisions in every domain, full-context quality, or complete agent-task success.
+ScienceWorld training predicts operation types, not complete object-specific commands.
+Changing the prompt, label mapping, loop count, or quantization can change the results.
 
 ## Use the GGUF with llama.cpp
 
-The GGUF works with llama.cpp's native `/v1/systemone` decision endpoint. No Eider server, custom llama.cpp patch, or further training is required.
+The GGUF works with llama.cpp's native `/v1/systemone` decision endpoint. It can serve decisions directly, without an Eider server.
 **Dory architecture support is still in NVIDIA's unmerged [llama.cpp PR #30306](https://github.com/ggml-org/llama.cpp/pull/30306).**
 Use that branch until architecture support reaches upstream. This requirement also applies to ordinary generation.
 
 ### Download
 
-Install the Hugging Face CLI and authenticate with `hf auth login` if this repository is private.
+Install the Hugging Face CLI to download the weights.
 Download only the format you need:
 
 ```sh
@@ -350,7 +237,7 @@ curl http://127.0.0.1:18091/v1/systemone \
 
 The endpoint returns answer probabilities with zero generated tokens.
 The GGUF provides a named `systemone` prompt template and selects llama.cpp's existing `granite-switch` decision convention.
-That name selects letter labels, true-before-false ordering, and selected next-token logits. **The architecture and weights are Dory, not IBM Granite.**
+This metadata selects the answer-label convention; the model architecture remains Dory.
 The normal chat template is also retained.
 
 | Question type | Supported input |
@@ -364,24 +251,10 @@ The tested server rejects excess score levels with HTTP 400. Template validation
 The template accepts text, JSON states, and plain text chat. Serialize tool-call history as state text.
 Image and audio inputs are unsupported.
 
-### What was tested
-
-On 2026-10-21, the Q8_0 GGUF ran on GB10 with all 73 layers offloaded and no local llama.cpp source changes.
-Eight native prompt checks matched Eider's token IDs and answer labels exactly.
-They covered JSON, Unicode, chat, 64 choices, default yes/no descriptions, and 10 score levels.
-The 24-request GPU smoke check covered all three answer types and returned the expected answers with zero generated tokens.
-
-For short requests with changing state, median latency was 62 ms for one question and 191 ms for three questions.
-Each median used five timed requests. Server logs showed full prompt processing without prefix reuse.
-These small compatibility checks do not establish Q8_0 accuracy on the full evaluation sets.
-[GGUF provenance and test evidence](gguf/provenance.json) record the tested source revision, file hash, and checks.
-
-## Scope and provenance
+## Attribution and provenance
 
-This experimental fine-tune targets typed decisions, action selection, state tracking, and planning.
-The results do not establish better general chat, reliable decisions in every domain, or full-context quality.
-[BF16 provenance](provenance.json), [NVFP4 provenance](nvfp4/provenance.json), and [format hashes](formats.json) bind the files to the selected adapter and source weights.
+NVIDIA supplies the base model; the linked dataset authors supply the training sources.
+[Eider](https://github.com/rdaum/eider) supplies the training tooling and evaluated inference implementation.
+This is an independent derivative under [OpenMDW-1.1](LICENSE); see [NOTICE](NOTICE).
 
-NVIDIA supplies the base model, and the linked authors supply the training sources.
-[Eider](https://github.com/rdaum/eider) supplies the inference implementation and experiment tooling.
-The base model licence is [OpenMDW-1.1](LICENSE). This is an independent derivative.
+[BF16 provenance](provenance.json), [NVFP4 provenance](nvfp4/provenance.json), and [file hashes](formats.json) identify the source weights and adapter.