README.md

---
license: cc-by-4.0
pretty_name: DualLoop Difficulty Feedback
configs:
- config_name: tasks
  data_files:
  - split: train
    path: parquet/tasks/train-*.parquet
- config_name: sources
  data_files:
  - split: train
    path: parquet/sources/train-*.parquet
- config_name: replay
  default: true
  data_files:
  - split: train
    path: parquet/replay/train-*.parquet
- config_name: feedback_eval5
  data_files:
  - split: test
    path: parquet/feedback_eval5/test-*.parquet
- config_name: evaluator_responses
  data_files:
  - split: test
    path: parquet/evaluator_responses/test-*.parquet
- config_name: hard_eval5
  data_files:
  - split: test
    path: parquet/hard_eval5/test-*.parquet
- config_name: audit_incomplete_tasks
  data_files:
  - split: train
    path: parquet/audit_incomplete_tasks/train-*.parquet
- config_name: audit_incomplete_responses
  data_files:
  - split: train
    path: parquet/audit_incomplete_responses/train-*.parquet
- config_name: audit_replay_quarantine
  data_files:
  - split: train
    path: parquet/audit_replay_quarantine/train-*.parquet
task_categories:
- visual-question-answering
tags:
- synthetic
- visual-question-answering
- agentic-generation
- difficulty-feedback
- executable-verification
- parquet
size_categories:
- 10K<n<100K
---

# DualLoop Difficulty Feedback

This release contains the DualLoop model-feedback discovery data. Coding agents
generated executable visual-question tasks; each task's inverse program checked
the answer from rendered pixels. For complete feedback transactions, the exact
same five images were evaluated by three frontier VLMs and the resulting
difficulty signal was returned to the next generation episode.

## Contents

| Config | Unit | Rows |
|---|---|---:|
| `tasks` | generated task | 266 |
| `sources` | candidate source file | 2,926 |
| `replay` | replayed visual-question prompt | 78,659 |
| `feedback_eval5` | exact in-loop feedback image | 1,260 |
| `evaluator_responses` | evaluator response | 3,780 |
| `hard_eval5` | exact image from a frozen hard task | 105 |
| `audit_incomplete_tasks` | incomplete feedback transaction | 14 |
| `audit_incomplete_responses` | durable response from an incomplete transaction | 45 |
| `audit_replay_quarantine` | replay-quarantined task | 2 |

The release contains 266 generated tasks from five builder configurations and
two frozen seed pools. Of these, 252 have complete 5-image x 3-evaluator
feedback transactions, 21 satisfy the frozen panel-hard rule, 264 replay
successfully, and 2 remain quarantined.

## Frozen hardness rule

An evaluator fails a task when it answers at most 3 of the 5 exact feedback
images correctly. A task is panel-hard when at least 2 of the 3 evaluators fail
it. The 21 hard tasks comprise 5 `consensus_hard` tasks and 16
`majority_hard` tasks. This is panel-specific model difficulty, not a human
difficulty label.

The evaluator panel is:

- `openai/gpt-5.6-sol` at high reasoning;
- `anthropic/claude-opus-5` at high reasoning;
- `google/gemini-3.7-flash` at high reasoning.

## Exact feedback panels

`feedback_eval5` contains the original images evaluated inside the loop. It is
not a new replay sample. Every image is checked against the SHA-256 recorded in
the original `sample-set.json`. `hard_eval5` is an ID- and hash-preserving view
of the 21 hard tasks.

## Safe evaluation boundary

The dataset contains gold answers for scoring. A model request must use only:

- `image`;
- `question`;
- `answer_options`.

Never serialize an entire row into the prompt. In particular, do not expose
source code, scene specifications, oracle output, hardness labels, prior model
responses, or gold answers. The gold-free request ledger is
`ledgers/feedback_eval5_inputs.csv`; answers are isolated in
`ledgers/feedback_eval5_gold.csv`.

## Mechanical validity, replay, and difficulty are different

- `machine_validated` means the submitted executable task passed the protected
  program and pixel-verification gates.
- `replay_status=verified` means the archived task successfully generated the
  requested offline replay dataset.
- `hardness_class` summarizes performance of the frozen three-model panel.
- `human_admission` remains pending and `canonical_world_member` is false for
  every task in this release.

## Source programs

The `sources` config stores forward renderer/generator code, inverse
oracle/verifier code, tests, metadata, and supporting text as searchable UTF-8
rows. No opaque source tarball is required to inspect the tasks.

## Known boundaries

- Fourteen feedback transactions are incomplete and are excluded from all
  accuracy and hardness claims. Their durable evidence is preserved in audit
  configs.
- Two tasks failed the requested 200-scene replay and are excluded from
  `replay`.
- Seven arms use `episodic-sequential@0.7.0`; three corrected arms use
  `episodic-sequential@0.8.0`. Cross-revision aggregates are descriptive unless
  the guard revision is modeled.
- The builders and evaluator panel do not span all model families.
- No signed human-admission study is included in this release.

## Loading

```python
from datasets import load_dataset

panels = load_dataset(
    "DualLoop/DualLoop-Difficulty-Feedback",
    "feedback_eval5",
    split="test",
    revision="v0.1.0",
)
row = panels[0]
row["image"].show()
print(row["question"], row["answer_options"], row["oracle_answer"])
```

For reproducible runs, pin the dataset revision to the `v0.1.0` release tag or its commit SHA.

## Release integrity

This DualLoop release preserves the questions, answers, images, source programs,
and dataset configurations. Publication metadata uses the DualLoop account.
Local filesystem paths in audit diagnostics are redacted; `CHECKSUMS.sha256`
contains checksums for the files distributed here.