RegalFire/Windway-False-Completion-Cases-v0.1

Windway False Completion Cases v0.1

Windway Data Community Release - four verified first-party mismatch cases. The agent asserted completion, but the controlled application state contradicted it. This is a free, small diagnostic casebook: three ComputerOps cases and one representative controlled-replication case observed in three repetitions.

Useful for: testing state-aware evaluation, completion-checker regression and teaching the distinction between claims and durable outcomes. Each case is a Windway-authored factual summary, not a raw prompt or model response. Internal source episodes were re-read and their successful and failed state transitions replayed before packaging.

Limits: selected failure-only cases, one local model/quantization, synthetic interfaces and a simplified transaction scaffold. No failure-rate estimate, model ranking, causal claim, client result, production-agent certification or evidence about another model. This is a casebook, not a standalone browser environment or turnkey agent benchmark.

Why this dataset exists

Don't trust agent self-report. Verify environment state. A completion message and a saved state are different evidence. The cases expose missing records, wrong-target updates and acknowledgement without a durable effect.

Use cases

Completion-checker regression, verifier-unit examples and state/claim mismatch analysis. Do not use the all-failure selection to estimate reliability or evaluate a model with leaked expected answers.

Data structure

test split: four cases. Readable columns include task, editorial agent_claim, expected_state, verified_state, mismatch_type, failure_context, failed_checks and verifier_outcome. Compact state maps are JSON strings; no giant raw nested trace is included. Model, quantization, repetition count and SHA-256 provenance are explicit. There is no train split.

Example

One case asserts that a requested synthetic CRM lead was created. The durable record does not exist; the independent state checks fail. Three transaction repetitions omit the same requested record; only one representative row is released to avoid padding the casebook.

from datasets import load_dataset
import json
ds = load_dataset("RegalFire/Windway-False-Completion-Cases-v0.1", split="test")
case = ds[0]
assert json.loads(case["expected_state"]) != json.loads(case["verified_state"])

Collection methodology

No new model calls were made for this release. Source: the existing first local ComputerOps text-DOM baseline and the prespecified controlled replication. Selection requires an explicit completion assertion and a failed independent state verifier. The public agent_claim field is authored by Windway and summarizes only the assertion, excluding proprietary prompt text, model prose and internal journals.

Verification methodology

Fresh disk reader replayed all10 ComputerOps episodes and all36 controlled-replication episodes with their pinned state transition/verifier contracts. All46 replayed, including failures. Expected versus actual case projections were recomputed from those states. validate_cases.py verifies this public projection's schema, typed flags, hashes and mismatch predicates; it cannot replay withheld private trajectories. Package tests include plausible tampered outcomes. Public code uses Apache-2.0; data uses CC BY4.0.

Provenance

Original Windway synthetic fixtures; no live business application, external website, client or prospect data. Source record hashes bind internal evidence retained by Windway. Public payload hashes detect changes; hashes alone do not establish identity, timestamp authenticity or licensing. No private path, raw model response, API credential, internal scenario library or infrastructure code is bundled. This is not Qwen training data or a release of Qwen weights.

Limitations

Four selected cases are not a random sample or four independent model experiments. Three repeated transaction observations share one task and are collapsed. ComputerOps used a text-DOM control adapter, not vision. Native browser validation bubbles were not supplied to the model. Seed support was unverified. Environment replay is not deterministic model sampling. No external independent reproduction is claimed. Future evaluations must keep expected-state labels away from agents.

License

Copyright Windway Data. Data compilation and original editorial summaries: CC BY4.0. Attribute Windway Data, the dataset/version and source URL, and indicate changes. Code: Apache-2.0, separately noted in LICENSE_CODE. License does not grant rights to private evidence, model weights, brands or unrelated third-party material. See LICENSE_DATA.

Citation

@misc{windway2026falsecompletion,
  author = {Windway Data},
  title = {Windway False Completion Cases v0.1: Community Release},
  year = {2026},
  url = {https://huggingface.co/datasets/RegalFire/Windway-False-Completion-Cases-v0.1}
}

Windway Research

Windway Data · Research · Failure Is Not Binary · Agent Failure Recovery Benchmark.

Commercial/custom data work

Windway is an early-stage AI reliability and data infrastructure startup. We build tailored evaluation datasets, private agent benchmarks and reproducible data pipelines around your environments, schemas and failure modes. Discuss custom/private evaluation. No customers, large team or revenue are implied.

Windway publication and related research

Windway Data maintains this selected public research asset through the existing RegalFire publication account. This operator statement does not replace historical RegalFire attribution, change the asset-specific license, or claim ownership of third-party material. Company/operator disclosures are on the canonical legal page.

• Windway Data — canonical website
• Windway research
• Failure Is Not Binary — methodology and limitations
• Agent Failure Recovery — historical fixture benchmark
• Verified Agent Trajectories v0.1 — 60 state-verified episodes

These are distinct resources: fixed-label fixtures, recorded state-verified trajectories, editorial cases and a research note. Their sample sizes and metrics must not be pooled. No customer evaluation or comparative leaderboard is implied.