RegalFire/windway-verified-agent-trajectories

Windway Verified Agent Trajectories v0.1

60 verified trajectories · 50 task IDs · 10 synthetic workflow domains · 42 successes / 18 failures.

Study how an agent's tool choices affect business state, including failure injection and recovery. Every included trajectory was re-read from immutable evidence, replayed in its source simulator, and checked against an independent exact-state verifier. GOLD means evidence quality, not task success.

Intended uses: offline trajectory inspection, failure taxonomy research, state-based evaluation tooling, and provenance/verification pipeline development. This is one internal Qwen2.5-Coder-14B-Instruct Q5_K_M baseline (temperature 0, context 8192, maximum output 180 tokens), not a comparison of models or vendors. The environment uses a shared three-stage scaffold, not live applications. Raw prompts and model prose are excluded; historical weights hashes are missing. Local release candidate: publication and the proposed CC BY 4.0 grant await owner approval.

Why this exists and research question

Can recovery decisions complete a synthetic workflow while preserving unrelated state? The dataset keeps failed trajectories and exact final-state verdicts so that a completion narrative cannot substitute for environmental evidence. Source does not capture agent completion claims: agent_claim is null, never inferred from actions or outcome. This release cannot measure false-completion rates.

Load and inspect

from datasets import load_dataset
import json
ds = load_dataset("parquet", data_files={"eval": "eval.parquet"}, split="eval")
item = ds[0]
actions = json.loads(item["actions"])
verdict = json.loads(item["verifier_result"])
print(item["episode_id"], item["outcome"], verdict["goal"]["passed"])

JSONL contains native nested structures. Parquet stores heterogeneous nested fields as JSON strings to provide a stable Dataset Viewer schema; manifest.json lists those columns. Scalars retain native types. The HF configuration serves only eval.parquet, avoiding duplicate loading of JSONL and Parquet. This is one eval split, with no claimed held-out train/test independence.

Schema

schema.json defines the native JSONL form. Identifiers: dataset_id, dataset_version, item_id, episode_id. Task/state: task, environment, initial_state, goal, actions, observations, final_state. Failure/recovery: failure_type, failure_injection, recovery. Evidence: verifier, verifier_result, outcome, provenance hashes, source_type, synthetic, generator, generator_version, projection_version, model, seed, seed_support, timestamp, human_review, license, publication_status, limitations. agent_claim_status documents why the completion claim is unavailable.

Each observation retains its before/result/after state. Failure labels describe injected disturbances, not necessarily a failed final outcome. Recovery actions can include unsuccessful recovery attempts.

Verification methodology

The canonical adapter validates schema, record hashes, source-line hashes, source-manifest membership, task identity, original tool choices, deterministic transitions and independent goal verdicts. The release audit checks exact projection equality, duplicate episode/trajectory hashes, state-chain continuity, outcome/verifier agreement, nested JSONL/Parquet equivalence, privacy/secret patterns, and package byte hashes. Negative cases test these gates. VALIDATION.json is the executed receipt. Source simulator and raw traces remain private; public consumers can inspect trajectories and hashes but cannot independently reconstruct the complete original inference run from this package alone.

Provenance and rights

Source: Windway Arena internal local-model interaction captured October 2026, exported by Data Factory 2.0. Tasks assert original Windway synthetic fixtures with no real customer data; that statement is corroborated by the pinned local catalog and state/argument inspection. Model output is limited to structured tool choices against those fixtures. No free-form model responses, external documents, customer data, model weights, or third-party code are included. Source hashes and record linkage are retained; private absolute paths are removed. All 60 rows share one source run.

Proposed dataset/compilation license: CC BY 4.0, within rights held by Windway Data. The generation model and quantizer declare Apache-2.0; see LICENSE_REFERENCES.md. Model terms are not an ownership transfer or endorsement. Canonical per-asset rights review is maintained internally; all included projections are PUBLIC_CANDIDATE pending approval of this exact package.

Limitations

• Small, fully observable simulator with a shared validate/commit/finalize procedure scaffold.
• 60 episodes cover 50 task IDs; task reuse is not semantic episode duplication. Ten repeated trajectories were rejected upstream and are not included. Do not randomly split reused task scaffolds.
• 42/60 success applies only to these included episodes. It is not the source's 32/50 primary-task metric.
• Missing historical model revision/checksum and unverified requested seed support.
• Mutated source request histories prevent faithful prompt reconstruction. No training export is claimed.
• Goal fields and verifier results are oracle-bearing analysis data; do not expose them to evaluated agents.
• Automated content/privacy inspection is not a legal certification or a guarantee of broad real-world validity.
• No per-episode human annotation, customer results, live third-party evaluations or comparative leaderboard.

Citation and research

See CITATION.cff. Windway Data (2026), Windway Verified Agent Trajectories v0.1. Windway research · Community datasets

Custom / Private Evaluation Work

Windway builds private evaluation environments and verified data around a client's mechanics, action schema and failure modes. Public work demonstrates the pipeline; custom environments, integrations and evaluations are adapted privately. Contact Windway.

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
• False Completion Cases v0.1 — four editorial cases

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.

The CEO-approved v0.1.0 package is preserved at release commit 1b539c4. Its manifest SHA-256 is 051f5f4862dcf2f983262011dc8e65ded51bafc8cf193fe01c1420871308cc1e. Preparation-time candidate/pending labels in that immutable snapshot describe its pre-publication state; the owner approved and published those exact bytes on 6 October 2026. This subsequent README-only editorial link update does not replace that manifest or alter its data, license text or verification evidence.