ChrisRPL/enron-mailbox-substrate

Enron Mailbox Context Substrate

This dataset reconstructs temporally ordered Enron mailboxes by linking EnronQA emails to a structured Enron corpus. It is intended as a research substrate for studying historical-context retrieval: given a target email and only messages that existed before it, retrieve the prior evidence needed to resolve references or support downstream reasoning.

This is not currently a gold minimal-context benchmark. The record join, timestamps, and per-user chronology are validated; candidate thread structure and all automatic context annotations are derived or experimental, as described under Validation.

Motivation

Work email routinely refers to prior material without restating it: a value agreed earlier, an earlier decision, a person mentioned by first name, a prior thread. A system that reads such an email in isolation must either guess or consult the mailbox. This dataset supports research on that consultation problem: which prior messages, and how few, are needed to resolve the references in a target email and support correct downstream reasoning or action. Public resources do not currently pair real question-bearing email with the ordered mailbox history needed to study this, so the two existing sources are joined here and the temporal structure is reconstructed.

Dataset construction

The two sources are joined on the EnronQA path field against the corpus file_name field (both are maildir paths inherited from the same CMU release). The join was exact: all 73,772 EnronQA emails matched exactly one corpus record; there were no unmatched records and no one-to-many mappings. A fallback join by basename was implemented but never required.

Temporal histories are constructed with a single rule used throughout the derived experiments: for a target email, the usable history is the set of messages in the same per-user mailbox with a strictly earlier timestamp. This causal constraint is enforced programmatically in all tooling, and every message identifier used in an annotation or retrieval result can be checked against it with one timestamp comparison.

Known data caveats that survive the join: the underlying Enron corpus contains some corrupt dates at the extremes (the observed range is 1980 to 2004; 0.19 percent of mailbox rows predate 1998), and EnronQA email bodies carry a synthetic summary header added by the EnronQA authors while quoted reply text remains inline in the body.

Data sources

Source	Role
MichaelR207/enron_qa_0922 (EnronQA; arXiv:2505.00263)	73,772 emails with 528,304 question-answer pairs over 150 inboxes; defines the target-email population. Its QA answers are generated by an LLM pipeline and are silver, not human gold
corbt/enron-emails	Structured mirror of the raw Enron corpus: message id, subject, from, to, cc, bcc, date, body, file_name for 517,401 messages
CMU Enron corpus (enron_mail_20150507.tar.gz)	Upstream source of both; FERC-derivative public research release

Dataset structure

File	Contents
canonical_mailbox.parquet (73,772 rows)	One row per EnronQA email: message_id, user, timestamp (UTC), subject, from, to, cc, body (raw, including inline quoted text), enronqa_path, QA annotations (questions, gold answers, rationales), join status
user_mailboxes.parquet (517,401 rows)	Full mailboxes of the same 150 users: message_id, user, timestamp, subject, from, to, cc, body, file_name. This is the candidate pool for historical-context experiments
thread_edges.parquet (134,797 rows)	Candidate reply edges between mailbox rows; each edge carries confidence (high/medium) and an evidence string naming the rule that produced it
thread_stats.json, thread_examples.jsonl	Reconstruction statistics and 20 sampled edges for manual inspection
join_report.json	Measured join statistics
dense_cache_<user>.npz (24 users)	bge-small-en-v1.5 embeddings of each user's 2,000 most recent messages (800-character truncation), with row indices; derived, rebuildable
p0_records.jsonl, p0_stats.json, p0_examples.jsonl	LLM reference-prevalence scan over 500 sampled emails; sampler output, not annotation
pilot_targets.json, pilot_pools.jsonl, pilot_annotations.jsonl, pilot_gold_configA.jsonl, pilot_gold_configB.jsonl, pilot_agreement.json	The automatic context-annotation pilot: 40 emails, two independent LLM annotator configurations, per-config provisional labels, and the measured cross-config agreement
p1_p2_results.json, p1_p2_per_requirement.jsonl, error_analysis.json	Retrieval baseline results scored against the provisional pilot labels; indicative only
calibration_targets.json, calibration_prep.json	A 20-email stratified calibration subset selected for a subsequent annotation pipeline
stage_a..f_p1.jsonl, silver_labels_p1.jsonl, pipeline_summary_p1.json	Audit trail and outputs of a multi-LLM annotation pipeline over the calibration subset; may be partial while runs are in progress
unmatched_examples.jsonl, multi_match_examples.jsonl	Empty by construction; the join had no unmatched or ambiguous mappings
join_substrate.py, threads_reconstruct.py, stats_enronqa.py, p0_scan.py, pilot_annotate.py, p1_p2_cached.py, p1_p2_gpu.py, prep_calibration.py, silver_pipe2.py, silver_analysis.py	Scripts that produce every artifact above

No canonical train/dev/test split is defined. The upstream EnronQA release's three splits were verified to contain the same 73,772 emails (only the questions differ), so they are not leakage-safe; consumers who need a split should hold out whole users.

Validation

Component	Status	Notes
EnronQA ↔ Enron join	Validated	73,772/73,772 exact joins
Timestamp recovery	Validated	100% valid timestamps
Sender agreement	Validated	99.4%
Per-user mailbox chronology	Validated	517,401 messages
Candidate thread edges	Derived / provisional	Not gold thread labels
Automatic context annotations	Experimental / rejected as gold	Phase-1 stability was insufficient
Minimal sufficient context labels	Not available yet	Future research target

Supporting measurements: the embedded sender field agrees with the corpus sender in 99.44% of 73,772 joined pairs; 64.22% of address-like header recipients appear in the corpus recipient list (the remainder are bare display names); the corpus date is consistent with the maximum quoted Sent: date in 99.29% of 20,761 checkable messages.

The automatic context-annotation protocol was evaluated before any scale-up and was not sufficiently stable: two independent LLM annotator configurations over the same 40 emails agreed on 41% of requirement mentions, 34% of reference-mechanism labels, and 10% of evidence-set membership, and one configuration produced references to messages dated after their targets. Those annotations are therefore not presented as gold labels. Detailed measurements are in pilot_agreement.json and in the research history below.

Candidate thread reconstruction

Thread structure is not present in the source data and is reconstructed heuristically. Two rules produce candidate reply edges: matching a message's first quoted-message header (from, subject, sent date) against the user's mailbox under a normalized-subject and seven-day window, and matching normalized subjects with participant overlap and strictly earlier timestamps within thirty days. Each edge records which rule produced it and its confidence.

The reconstruction is incomplete and unverified against gold: quoted-message headers resolve a parent for only 8.92% of quoted emails under the strict windows, no public gold thread labels exist for Enron, and the structural relations should be treated as a lower bound on thread membership rather than as fact.

Intended uses

• Research on historical-context retrieval and sufficient-evidence selection over real mailbox histories, under the causal constraint described above.
• Research on thread reconstruction and on annotation-pipeline design, including reproduction of the negative stability result.
• Evaluation of retrieval systems that report their treatment of the dense-embedding window (see Limitations) and of the provisional status of all annotation files.

Limitations

• All automatic annotations in this repository are machine-generated. None have been verified by human annotators, and the initial pilot was explicitly rejected as gold.
• Candidate thread edges are heuristic, incomplete, and unvalidated.
• The dense-embedding caches cover only each user's 2,000 most recent messages; lexical search covers full histories. Experiments using dense retrieval must state this window.
• The corpus is Enron-only, from one energy company circa 2001; findings may not transfer to modern mail.
• Some timestamps are corrupt (see Dataset construction); temporal analyses should filter the extremes and report the number of rows dropped.
• QA answers inherited from EnronQA are LLM-generated silver labels.

Loading the dataset

from huggingface_hub import hf_hub_download
import pandas as pd

REPO = "ChrisRPL/enron-mailbox-substrate"
mailbox = pd.read_parquet(hf_hub_download(REPO, "user_mailboxes.parquet", repo_type="dataset"))
edges = pd.read_parquet(hf_hub_download(REPO, "thread_edges.parquet", repo_type="dataset"))

# Historical state for the message at row t: strictly earlier messages, same user.
t = 125436
history = mailbox[(mailbox["user"] == mailbox["user"][t]) & (mailbox["timestamp"] < mailbox["timestamp"][t])]

# Candidate thread edges for the target (grade by `confidence` and `evidence` before use).
parent_edges = edges[edges["child_idx"] == t]

Reproduction

All artifacts are produced by the scripts in this repository. Fixed seeds are recorded in each script; LLM calls run at temperature 0, but inference-provider nondeterminism exists, so the released records are the canonical run and each record names the model that produced it.

uv run join_substrate.py            # join and canonical table (~2 min, CPU)
uv run threads_reconstruct.py       # mailboxes and candidate thread edges (~5 min, CPU)
uv run stats_enronqa.py             # upstream row-level audit
uv run p0_scan.py                   # reference-prevalence scan (~10 min; requires HF_TOKEN)
uv run pilot_annotate.py            # annotation pilot (~33 min; requires HF_TOKEN)
uv run p1_p2_cached.py              # retrieval baselines, CPU path
uv run p1_p2_gpu.py                 # retrieval baselines, GPU path (~4 min on t4-small)
uv run prep_calibration.py          # calibration subset and missing caches (GPU)
uv run silver_pipe2.py              # multi-LLM annotation pipeline (PASS=1 / PASS=2)
uv run silver_analysis.py           # agreement, stability, tier analysis

Provenance and licensing

The Enron corpus is a public research release of real employee email compiled after FERC investigations. It contains real names, addresses, and opinions of real people. Neither upstream repository carries a formal license; this dataset inherits the Enron release terms (made available as a resource for researchers) and no single dataset-level license identifier is asserted here. Users should not attempt de-anonymization, should not use the data for profiling, and should follow established research norms for the Enron corpus. EnronQA answers are LLM-generated silver labels, and all annotations produced in this repository are machine-generated and marked as such in their files.

Citation

If you use this dataset, please cite the Enron corpus (CMU release), EnronQA (arXiv:2505.00263), the corbt/enron-emails mirror, and this repository, and state explicitly which annotation files, if any, you treated as reference labels.

Research status / changelog

Commit	Content
18c953e	Join and canonical table (exact join verified)
ab81df4	Mailboxes and candidate thread edges
1211203	Reference-prevalence scan
02d4299	Annotation pilot and agreement analysis
eb78538	Per-user dense-embedding caches
a19d23a	Retrieval baselines against provisional labels
748f6e5	Calibration subset and missing caches
ongoing	Multi-LLM annotation pipeline over the calibration subset (pass files update in place)

Research history, summarized. An initial automatic context-annotation pilot (two independent LLM annotator configurations, 40 stratified emails, temperature 0) was measured before scale-up and found unstable: requirement-trigger agreement 0.413, mechanism agreement 0.342, evidence-set overlap 0.103, roughly half of resolvability judgments discordant, and 49 references to non-prior messages in one configuration's output. The pilot was consequently rejected as a source of reference labels, and a revised staged pipeline (independent requirement extraction, normalization, tool-based mailbox search, dual-judge verification with span checks, leave-one-out minimality tests, and an independent adjudicator, with rerun-stability and counterfactual checks) was built to test whether stable silver labels are attainable without human annotation. Retrieval baselines reported against the provisional pilot labels (best configuration: 0.32 recall at 10 for a recency baseline, 0.20 for hybrid BM25+dense) should be read as order-of-magnitude indicators only, since the reference labels they were scored against are the ones this validation rejected.