Human review guide: synthetic packets (Phase B) and Q&A rows (Phase C)
AT&T requires people to check about 2 % of every batch and report the error rate. Their acceptance limit is at most 2 % of items with a factual or domain error. An AI judge has already checked everything; this review measures how often it was wrong.
Method: open the data files, scroll to random places, and review whatever you land on. No list is pre-selected. Spread your picks: different categories, different places in each file, real and synthetic, different question types.
What to review and how many
| Package | Files | Target sample (~2 %) |
|---|---|---|
| Phase B - synthetic packets (AI-written data documents modelled on a real run of the same scenario) | phaseB_synthetic_packets_2026-10-24/data/<category>.jsonl |
~21 packets (3 per category) |
| Phase C - Q&A rows (a data chunk + a question + an answer; real and synthetic chunks) | phaseC_qa_rows_2026-10-24/data/<category>.jsonl |
Split by category, so each person checks what they know best: lte-signalling, rf-signal-quality, 5g-core-signalling, ran-telemetry, network-performance, faults-ground-truth, aaa-radius. One category (3 packets + 16 Q&A rows) takes about 1.5-2 hours.
Part 1 - Phase B synthetic packets
A packet claims to be another run of the measured scenario in its header (scenario_label); seed_source_id names
the real document it was modelled on. Check:
- Protocol - messages, procedures and causes are standard and in the right order (3GPP / IETF / Open5GS, srsRAN, FreeRADIUS behaviour). Example errors: Attach complete before Initial Context Setup; an Authentication Reject carrying a cause code.
- Consistency - the same subscriber, IDs, sessions and timings agree across all tables and the ground-truth timeline. Example errors: a ping succeeds while the timeline says the UE was stopped; a test is "uplink" in the table but "downlink" in the timeline.
- Plausibility - values are possible for this setup: throughput vs. MCS and 10/20 MHz bandwidth, latencies, RSRP/SNR, loss %, counters. Example error: 40 Mbps downlink with the MCS capped at 8.
- Realism - it looks like a real run of that system, not invented patterns or suspiciously regular numbers.
- Scenario - the records actually show the claimed scenario (e.g. an AUSF outage shows failed authentication).
- Label-free version (
text_stripped, spot-check a few): no scenario label, no "ground-truth timeline", no fault events.
(45 packets that Phase C's Q&A answers found internally contradictory were already removed, with the Q&A rows built
on them; their ids are in phaseB_synthetic_packets_2026-10-24/reports/removed_flagged_packets.json.)
Part 2 - Phase C Q&A rows
Every answer has two parts: "What the records show" (facts) and "What it means" (explanation). provenance
says whether the chunk is real or synthetic; document_variant is labelled or stripped (no scenario labels; used
for detection and root-cause questions). Check:
- Facts - every value, time, identifier, message, count and difference in "What the records show" is in the chunk or correctly computed from it. Spot-check at least 3 numbers per row. Example errors: "27 rejects" when the chunk has 25; a time that is not in the chunk.
- Explanation - "What it means" is correct standard behaviour and does not overclaim. Example errors: a wrong NF role ("the UDM authenticates the UE" - it is the AUSF, with vectors from the UDM); claiming a root cause the records cannot show; treating simulated (ZMQ / UERANSIM) or uncalibrated radio values as real field measurements.
- Question - natural, answerable from the chunk, matches its
question_type. Anunanswerablerow must say what information is missing and not guess. On astrippedchunk the question must not reveal the scenario (e.g. "in this AUSF-stop run..." is an error).
Please do not edit the data files, and do not share the packages outside the team before the review is done.