temporal-context-gap-e1

Code and results for E1: temporal-provenance audit and E2: era-classification probe — the first two empirical experiments of the position paper "The Temporal Context Gap: What Meaning Is Lost Between a Corpus and Its Consumer" (draft, September 2026, §6).

E1 result (measured 2026-09-11, job 6aa330505527934177ec3185)

Temporal metadata is fully present and fully crawl-time in every corpus audited:

Stratum	Docs sampled	% with temporal field	Distinct years	Years seen
c4/en	900,000	100.0%	1	2019 only
c4/realnewslike	80,858	100.0%	1	2019 only
c4/en.noclean	445,231	100.0%	1	2019 only
fineweb/CC-MAIN-2013-20	900,000	100.0%	1	2013 only
fineweb/CC-MAIN-2016-18	900,000	100.0%	1	2016 only
fineweb/CC-MAIN-2019-18	900,000	100.0%	1	2019 only
fineweb/CC-MAIN-2022-21	900,000	100.0%	1	2022 only
fineweb/CC-MAIN-2025-26	900,000	100.0%	1	2025 only
dolma3/common_crawl-{adult_content,art_and_design,crime_and_law}	~67k each	0% top-level; per-record metadata.warc_date confirmed	varies	2015–2020

Finding. The headline was not "metadata is missing" but something stronger: every temporal field that survives curation is a fetch/crawl timestamp (C4's timestamp is the April-2019 crawl snapshot for all 1.43M sampled docs; FineWeb's date is the WARC fetch date of its dump — a single year per snapshot config; Dolma3's is metadata.warc_date, the per-record WARC fetch time, nested under metadata). The consumer side of the temporal context gap survives curation; the production side — the authorship time a temporal key needs (paper §5.2) — is absent everywhere. A model trained on these corpora can in principle know when it fetched a document but not when it was written, except through self-anchoring expressions inside the text itself (paper §4, Level 1 — measurable with temporal_keys.selfanchor_density).

Notes and caveats:

• c4/realnewslike and c4/en.noclean sampled fewer docs than c4/en because those configs' shard files contain fewer records per shard (3 shards × ~80–150k), not because of sampling failures.
• FineWeb's date column name suggests document dates; it is the crawl date (verified: within-week spread, one year per snapshot config).
• Dolma3's top-level schema is {id, metadata, text}; created (Dolma v1) is gone. The E1 audit script's created-top-level check returned 0% — the corrected measurement is the nested metadata.warc_date, verified in job 6aa343175527934177ec364f.

E2 result (measured 2026-09-11, job 6aa3644e5527934177ec410b)

Era-classification probe: TF-IDF + logistic regression predicting dump year from raw text (first 4,000 chars, 100k features), FineWeb sample-10BT file 0, 15,000 docs per year cap. The sampled file contained 8 dump years (2013, 2015–2020, 2022); 120,000 docs, 24,000-doc test set.

• Year-from-text accuracy: 22.83% vs. 12.5% majority baseline (1.8×, 8 balanced classes)
• Within ±1 year: 41.35%; within ±2 years: 61.8%; mean abs year error: 2.27
• Per-year accuracy flat (0.151–0.269; lowest 2017/2018)
• Month probe not implementable on FineWeb: each dump's fetch dates concentrate in the crawl window, so the corpus provides no fine-granularity time labels — itself a gap datum (consumer-side metadata cannot even define authorship-month labels)
• Self-anchor density by year (temporal_keys.selfanchor_density): declines monotonically, 0.588 (2013) → 0.415 (2022)
• E1b — C4 URL-year heuristic (100k docs): 19.37% of URLs carry a path-encoded year. Histogram concentrates 2009–2019 (peak 3,092 in 2018), collapsing after 2019 (40 at 2020) — a validation: a URL year cannot postdate the April-2019 C4 crawl. A small id-like tail (years ≤1900 / far-future) is regex noise.

Reading. Coarse era signal survives curation and is measurable in raw text alone (±2-year precision at 62%); fine granularity is unrecoverable — the corpus cannot even supply its labels. The declining self-anchor trend suggests the Level-1 anchor (paper §4) is eroding in web text.

Files

• temporal_keys.py — the TemporalKey dataclass, <KEY ... /> serialization and parsing (round-trip), K2 key-dropout masking, SemEval-2020-Task-1-style LSC drift-data annotation, and self-anchoring density. Paper §5.2/§5.3 implemented.
• test_temporal_keys.py — unit checks (run python test_temporal_keys.py).
• audit_temporal_provenance.py — the E1 audit script (--quick for validation, --files N --rows M for deep sampling).
• e2_era_probe.py — the E2 + E1b probe script (year probe, month probe, self-anchor density, URL-year heuristic).
• RESULTS.md / results.json — E1 deep-audit results (written by the job).
• E2_RESULTS.md / E2_results.json — E2 + E1b results (written by the job).

Reproduce

pip install pyarrow zstandard huggingface_hub
python audit_temporal_provenance.py --files 3 --rows 300000 --out results.json --md RESULTS.md
pip install scikit-learn pyarrow numpy huggingface_hub
python e2_era_probe.py --per-year 15000 --out E2_results.json --md E2_RESULTS.md

Costs

• E1 deep audit: ~$1.13 (cpu-xl, 67.5 min actual)
• Dolma3 follow-up + sandbox validation: ~$0.01
• E2 probe: ~$0.004 (cpu-upgrade, 7 min actual vs 60m ceiling)