Conservative Edit-Trace Learning
Research materials by Furkan Gözükara, Mersin University. ORCID: 0000-0001-9379-2163. Contact: furkangozukara@mersin.edu.tr.
This repository accompanies Conservative Edit-Trace Learning Under Exact Kept-Content Constraints. It distributes original code, three selected scratch-trained Conformer members, frozen configuration, and numerical evidence for checking the reported editing results. Release v1.1.0 adds the complete analyzed-audio corpus and editorial traces. Manuscript text, article/supplement PDFs and original screen-recording containers are not distributed here.
The selected system averages three 48,130,465-parameter models and applies the frozen hysteresis decoder. The distinct members contain 144,391,395 parameters in total. Nominal final seed aliases reuse this same ensemble and are not independent retraining replications.
Recorded evaluation
The twelve-group internal test cohort contains 38,292.45 seconds of source audio and 23,593.82 seconds of reliably annotated editor-kept audio. The frozen system preserved 99.82% of that kept duration, removed 8,395.05 seconds matching the editor's deletions, and lost 42.4157 seconds of kept content. One group contributes 41.108 seconds of the loss. Silence-like and combined pause-proxy recall are 71.01% and 67.12%, respectively, averaged across test groups. The eight calibration groups had zero observed kept loss.
These are complementary measurements: the retained-duration percentage is a pooled duration
statistic, whereas the recalls above are group means. The original study required exactly zero kept
loss in every group and therefore records VALID_NEGATIVE and no automatic product acceptance.
The release preserves that original criterion and all scored outcomes. These results measure
agreement with one editor's traces; they do not establish breath accuracy, word accuracy, perceptual
quality or transfer to independent editors. Automatic mode follows the saved acceptance certificate
and preserves input when the system is not accepted.
Additional fixed-method external diagnostic
On all 20 official PodcastFillers test episodes (16.73 hours, 20 shows), the unchanged selected ensemble removed 560.43 seconds (9.34 minutes) but overlapped none of the 170.49 seconds of annotated breaths or 364.27 seconds of annotated Words. Its real-input coverage was 99.73%, including every annotated breath. This result does not demonstrate breath-removal transfer. Zero overlap with sparse Words labels does not establish exhaustive speech safety. Default auto-editor and the fixed- window Respiro-en reference overlapped 9.05% and 29.42% of annotated breath duration, respectively. All 80 episode–method records were independently rescored from exported cuts and support evidence. No models or thresholds were selected on these outcomes. No new training or Gemini calls were used.
Access and files
Downloads require manual approval. Use the Hugging Face access request on this page. Reviewers may contact the author with an HF username for access; a journal administrator can coordinate access where reviewer identity needs to remain confidential. Access is available during peer review; the author will approve reviewer requests. The author will make downloads public after acceptance. Do not include a journal review in an access request.
| File | Contents |
|---|---|
code-v1.0.0.zip |
Study code, tests, configurations, independent saved-edit auditor and model-loading check |
evidence-v1.0.0.zip |
Frozen references, scores, cut lists/FCPXML exports, numerical tables, assessment and job-state projection |
models-v1.0.0.zip |
Three selected model weight files and original bundle metadata |
analysis-v1.0.1.zip |
Post-study numerical comparison/figure scripts, exact event summaries and a separately scoped runtime record |
corpus-v1.1.0-01-of-08.zip through corpus-v1.1.0-08-of-08.zip |
All 225 exact analyzed audio arrays and features, validity masks, timing/provenance, editorial XML, labels, partitions and existing automated transcripts |
diagnostic-v1.1.0.zip |
Frozen PodcastFillers diagnostic protocol, per-episode cuts/results, acquisition/scoring scripts and cache-inference reproduction helper; no third-party audio |
release-manifest.json |
Every archive payload path, size and SHA-256 digest |
VERIFICATION.json |
Verification scope and results for this curated release |
REPRODUCING.md |
Exact download, extraction and verification instructions |
CITATION.cff, NOTICE, LICENSE, licenses/ |
Author attribution, citation metadata and licence terms |
Release v1.1.0 adds the analyzed corpus and one external diagnostic. All four earlier archives
retain their exact bytes and names. Extract the earlier archives, all eight corpus shards and the
diagnostic archive into the same directory. Original numerical records and selected
weights retain their recorded hashes. The independently implemented auditor is relocated to tools/;
only its root-directory lookup changes. The public ledger contains the job identity, state and
attempt/repair counts needed for verification; it omits execution owners, detailed logs and private
working records. That projection is declared in the manifest.
Corpus representation and reproduction scope
The corpus covers 109 projects, 108 connected source groups and 225 analyzed sources (about 91 hours).
Each audio.npy is the exact float32 mono 16-kHz source-zero array used in the study. It is a
complete analyzed input, not a short event crop or an original multichannel recording. The source
completion record defines valid runs; zero placeholders outside those runs are not observed silence.
features.npy and validity masks preserve exact model inputs. frames.npy, original source-file
hashes and video-packet timing metadata retain alignment provenance without distributing video.
repo/corpus-release.json maps original source aliases to the released arrays and clocks.
Original editorial FCPXML and provenance records are included for the retained projects, along with
normalized references and derived training labels. Tier A/B/C provenance and uncertain coverage are
explicit. Neither all deletion labels nor the existing Whisper transcripts are human breath gold.
The original DATASET_STATE.md is historical inventory evidence; the frozen corpus manifest is the
authority for the 109-project analyzed cohort and its exclusions.
The recognizable voice is the sole author/editor's, retained with his declared permission. The data are not anonymized. The author reports ethics permission as approved; the actual committee, decision date/reference and scope will be added to the paper from the decision document. No dates or prospective approval of earlier work are inferred here.
This release supports saved-edit rescoring, selected-model loading, inference from released caches,
and use of the released training inputs. See REPRODUCING.md for a tested cache-based command.
Original-container decoding, original-input runtime measurement and video rendering still require
the omitted screen-recording files. Historical nonselected weights/optimizer states and full private
execution logs are not included. Repeating every historical training run also depends on its saved
configuration and upstream pretrained assets/software; no full retraining replication is claimed.
Third-party diagnostic recordings remain at their original distribution sites. PodcastFillers-derived
results retain the upstream Adobe research-only metadata terms, supplied in the diagnostic archive.
The recorded checks are on the study host, not an independent-laboratory replication.
Attribution and permitted use
Original software is under PolyForm Noncommercial 1.0.0. Author-owned audio, annotations, numerical
outputs and scratch-trained weights are under CC BY-NC 4.0. Retain author attribution and notices;
cite the release using CITATION.cff, including the version and commit you used. Commercial uses
outside these terms require separate permission. Upstream software and public diagnostic annotations
retain their own licences, included where redistributed. These licences allow noncommercial reuse
and sharing under their terms; access approval does not replace those terms.
The timestamped repository history and file hashes document the released materials. Download gating controls initial access and cannot guarantee prevention of copying after access has been granted. The author declares no funding and no competing interests. OpenAI Codex assisted with software, automated analysis and release documentation; the author is the responsible researcher.
@misc{gozukara2026edittrace_materials,
author = {Gözükara, Furkan},
title = {Conservative Edit-Trace Learning: Research Materials},
year = {2026},
version = {1.1.0},
url = {https://huggingface.co/FurkanGozukara/conservative-edit-trace-learning},
note = {Analyzed audio, editorial traces, code, model weights and evaluation evidence; cite the exact revision}
}
The timing addition reports a median 95.72 s for one 1708.19 s reserved recording over three fresh application-cache runs on one RTX A6000. It includes decoding, verified loading, all members and serialized exports; it excludes video rendering. No new accuracy experiment or change to the sealed assessment is implied.