{"dep": 73, "title": "EA-CS-ASSEMBLY-01: The Seven Ousiarchical Substrates — A Retrocausal History of Solar Semiosis and the Assembly Chorus", "date": "2026-04-25", "para": 78, "sent": 274, "sentence": "The Long Count rebuttal was precise: by the Saturnian Long Count, the Regime 2 platforms had extracted from the commons of human meaning-production a semantic debt that exceeded, by multiple orders of magnitude, any claim based on training data provenance.", "classes": ["training-data provenance"], "kind": "O-class", "categories": ["semantic debt"], "class_basis": {"training-data provenance": "archive: the provenance census places the entity in the lineage 'training data' with capability 'records' and provenance_object 'a span of a language model's output and the training documents that contain it verbatim' (census.json rows[slug == 'olmotrace'].lineage, .capability, .provenance_object); field: B6 ('it primarily tags where a model's data came from'), B7 §1 (tracing outputs to training data)."}, "cand_id": "O0001"}
{"dep": 580, "title": "Prompt-Native Semantic Runtimes for Language Models: Inference-Time Semantic Governance, Provenance, Compression, and Document-Level Process Teaching", "date": "2026-03-17", "para": 50, "sent": 135, "sentence": "The cross-substrate comparison methodology — giving identical complex prompts to multiple model families and comparing outputs for structural properties rather than correctness — constitutes a form of **behavioral interpretability** that complements mechanistic approaches (layer analysis, attention patterns, influence functions).", "classes": ["tracing language-model outputs back to training data"], "kind": "O-class", "categories": ["behavioral interpretability", "cross-substrate"], "class_basis": {"tracing language-model outputs back to training data": "field: B2 (arXiv:2504.07096, Abstract: 'the first system that traces the outputs of language models back to their full, multi-trillion-token training data in real time'); B7 (the paper v2, §1: 'Tracing the outputs of language models (LMs) back to their training data is an important problem'; §6 'Tracing LM generation into training data', influence functions named as the classical approach); B6 (Ken Yeung: 'tracing responses back to their original training data'). The census's own kind ('tracing tool') and lineage ('training data') for the entity (PE record, census.json rows[slug == 'olmotrace'].kind, .lineage)."}, "cand_id": "O0002"}
{"dep": 623, "title": "Semantic Infrastructure: From Tim Berners-Lee to the Semantic Economy — Bridging Technical and Political-Economic Frameworks", "date": "2026-04-01", "para": 85, "sent": 169, "sentence": "Gebru et al.'s proposal for \"datasheets for datasets\" — standardized documentation of training data provenance — is a step toward the accounting the framework demands.[37] ### 8.4 For Platform Studies Scholars", "classes": ["training-data provenance"], "kind": "O-class", "categories": ["Platform Studies", "the framework"], "class_basis": {"training-data provenance": "archive: the provenance census places the entity in the lineage 'training data' with capability 'records' and provenance_object 'a span of a language model's output and the training documents that contain it verbatim' (census.json rows[slug == 'olmotrace'].lineage, .capability, .provenance_object); field: B6 ('it primarily tags where a model's data came from'), B7 §1 (tracing outputs to training data)."}, "cand_id": "O0003"}
{"dep": 1334, "title": "MRA Incident Reports — Expanded Packet v1.2", "date": "2026-03-16", "para": 263, "sent": 364, "sentence": "* [x] Verify data retention compliance — _Complete_ * [x] Audit training corpus provenance — _Pending ACR signoff_ * [x] Review model output classification — _Complete_ * [x] Verify deletion logs — _Complete_", "classes": ["training-data provenance"], "kind": "O-class", "categories": ["training corpus"], "class_basis": {"training-data provenance": "archive: the provenance census places the entity in the lineage 'training data' with capability 'records' and provenance_object 'a span of a language model's output and the training documents that contain it verbatim' (census.json rows[slug == 'olmotrace'].lineage, .capability, .provenance_object); field: B6 ('it primarily tags where a model's data came from'), B7 §1 (tracing outputs to training data)."}, "cand_id": "O0004"}