Datasetpublic

spectralbranding/tba-cascade-gap-largen

Cascade-Gap Large-N Study — evidence corpus and coding logs Companion dataset for the working paper A Large-N Confirmatory Necessary-Condition Test of Closing-Time Structural Gaps in Acquisition Integration Failure (Zharnikov, 2026). Concept DOI 10.5281/zenodo.21755969. Dataset DOI 10.57967/hf/9805. Code, paper, and reproduction pipeline: https://github.com/spectralbranding/orgschema-papers/tree/main/cascade-gap-largen Dataset viewer The viewer shows… See the full description on the dataset page: https://huggingface.co/datasets/spectralbranding/tba-cascade-gap-largen.

sourceHugging Facecc-by-4.0updated 2mo agoView on Hugging Face

0likes409downloads

Dataset Card

Cascade-Gap Large-N Study — evidence corpus and coding logs

Companion dataset for the working paper A Large-N Confirmatory Necessary-Condition Test of Closing-Time Structural Gaps in Acquisition Integration Failure (Zharnikov, 2026). Concept DOI 10.5281/zenodo.21755969. Dataset DOI 10.57967/hf/9805.

Code, paper, and reproduction pipeline: https://github.com/spectralbranding/orgschema-papers/tree/main/cascade-gap-largen

Dataset viewer

The viewer shows `full_draw_dataset.csv` — the analysis dataset of record: 350 completed transactions, one row per case, with the per-construct codes (structural gap, outcome pathway), the outcome status, the deal-structure stratum, and the rotated coder assignment. This is the single table the paper's results are computed from.

The bulky raw evidence and logs are archival files under the Files and versions tab (they are not a train/validation split — the auto-inferred split view does not apply to them):

Path	Description
`full_draw_dataset.csv`	The 350-case coded dataset (the viewer table).
`full_draw_dossiers/`	700 blinded per-case evidence dossiers (350 structural + 350 outcome), built from public SEC EDGAR filings — the coder-facing evidence base.
`full_draw_code_out/`	700 per-case coded records (structural + outcome JSON per case) plus reliability outputs.
`logs_fulldraw/`	700 JSONL model-call logs — every rater call with full prompt, parameters, and response (secrets-clean).

Provenance

Four large language models served as independent blinded raters under a pre-registered 3-of-4 per-construct rotation (seed 20260729): Claude (Opus 4.8), Gemini (3.1 Pro), Grok (4.5), GPT (5.6). All source documents are public SEC EDGAR filings. License: CC BY 4.0.