Datasets:

confetti3
/

workledger-structured-edit-interface-study-v0.1

pretty_name: WorkLedger Structured Edit Interface Study v0.1
license: cc-by-4.0
tags:
  - structured-editing
  - structured-edit-interface-study
  - language-model-interfaces
  - verified-code-edits
  - fail-closed-evaluation
  - software-engineering-agents
  - controlled-interface-study
task_categories:
  - text-generation
language:
  - en
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: generations
        path: data/generations.jsonl
      - split: tasks
        path: data/tasks.jsonl
      - split: aggregate_results
        path: data/aggregate_results.jsonl

WorkLedger Structured Edit Interface Study v0.1

Author: Confetti3

This upload root is the offline release-candidate dataset card for workledger-structured-edit-interface-study-v0.1, destined for confetti3/workledger-structured-edit-interface-study-v0.1. The Hugging Face repository is hosted under the authenticated namespace confetti3, while public authorship is credited only to the pseudonym Confetti3. It is built from the sanitized public package under publications/workledger-structured-edit-interface-study-v0.1; it is not a live runner, it constructs no provider transport, and it performs zero network or credential operations.

Study scope

WorkLedger v0.1 evaluates structured editing interfaces for language-model-generated code edits under fail-closed validation. The package contains sanitized final assistant answers, task metadata, aggregate outcomes, claim-survival records, schemas, provenance, leakage checks, and deterministic manifests. It supports public inspection of the controlled interface study records; it does not include hidden evaluator source, raw provider bodies, request identifiers, reasoning traces, credentials, SQLite ledgers, or private artifact paths.

Registered findings

Exact replacement advantage replicated for Qwen3 14B: U1 1/19 and R 13/19 on the registered paired primary comparison.
Exact replacement advantage replicated for Qwen3 235B A22B: U1 3/19 and R 16/19 on the registered paired primary comparison.
Qwen3-32B is an external anchor from existing calibration evidence: U1 0/19 and R 9/19 in the same direction, not jointly randomized and not in the paid multimodel denominator.
Qwen3 8B was directionally favorable to R, with U1 2/19 and R 5/19, but paired coverage was 18/19 and the registered interface-advantage rule was not met.
DeepSeek V4 Pro is infrastructure-unassessable: 38/38 cells were provider-route privacy/data-policy HTTP 404 failures with empty final answers. This is not a coding-performance score and must not be treated as a model-performance denominator.
CROSS_MODEL_GENERALIZATION_SIGNAL is false. The model-strength interaction is descriptive only, not a parameter-scaling law.
Strict native and one-outer-fence-normalized policies are reported on identical response bytes and were identical in the new multimodel run.

Limitations

The evidence uses exactly one attempt per cell. Provider routes and inference parameters are confounded with model identity. The task bank is calibrated and bounded; it is not a production-readiness claim and should not be generalized to broad software engineering workloads. DeepSeek V4 Pro produced no valid model outputs in this run, so it supports neither the interface advantage nor its refutation.

Ethics, privacy, and security

The public package is sanitized by field selection and leakage scans. Private execution evidence is withheld, including authorization material, request envelopes, response identifiers, raw provider bodies, reasoning fields, hidden evaluator source, private filesystem paths, and SQLite ledgers. The release candidate includes text-only upload rehearsal files, but no upload, tag, release, DOI, repository creation, provider request, or credential access has been executed.

Data structure

data/generations.jsonl: sanitized calibration and multimodel final assistant answers plus public outcome metadata.
data/tasks.jsonl: public task metadata and hashes with source/evaluator bytes withheld.
data/aggregate-results.json and data/aggregate_results.jsonl: immutable aggregate outcomes.
data/claim-survival.json and data/claim_survival.jsonl: audited public claims and safest wording.
schema/*.schema.json: public record schemas.
provenance/*.json: build provenance, manifest aliases, leakage checks, sanitization proof, license audit, and withheld-content policy.
MANIFEST.json and MANIFEST.sha256: release-root integrity manifest and digest.

Provenance and task wording

Low-affinity calibration tasks are described as newly authored for calibration when supported by release provenance. High-affinity tasks are described as adapted or reconstructed from upstream issue/task patterns; upstream source bytes and evaluator bytes are withheld. The package preserves frozen hashes, pass counts, and decision labels.

Citation

Please cite as: Confetti3. WorkLedger Structured Edit Interface Study v0.1. Version 0.1.0. Release date 2026-07-20.

License treatment

Public dataset metadata and public release records are licensed under CC-BY-4.0. Space code and tooling are licensed under MIT. This release does not relicense upstream source bytes, evaluator bytes, private execution evidence, or withheld fixtures.

Withheld-content policy

The release discloses sanitized records and non-sensitive hashes only. Hidden evaluators, raw provider evidence, request/response identifiers, reasoning traces, credentials, private paths, SQLite ledgers, and upstream source/evaluator bytes are withheld.