Farbod deliverable Oct 24

Phase A/B/C delivery bundles and review guide; JSONL data tracked in LFS.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

.gitattributes

@@ -60,3 +60,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
 *.tar.gz filter=lfs diff=lfs merge=lfs -text
 *.ndjson filter=lfs diff=lfs merge=lfs -text
+deliveries/**/data/*.jsonl filter=lfs diff=lfs merge=lfs -text

deliveries/REVIEW_GUIDE.md

@@ -0,0 +1,58 @@
+# 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` | ~110 rows (~16 per category) |
+
+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:
+
+1. **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.
+2. **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.
+3. **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.
+4. **Realism** - it looks like a real run of that system, not invented patterns or suspiciously regular numbers.
+5. **Scenario** - the records actually show the claimed scenario (e.g. an AUSF outage shows failed authentication).
+6. **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:
+
+1. **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.
+2. **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.
+3. **Question** - natural, answerable from the chunk, matches its `question_type`. An `unanswerable` row must say what
+   information is missing and not guess. On a `stripped` chunk 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.

deliveries/phaseA_real_documents_2026-10-24/README.md

@@ -0,0 +1,57 @@
+# OTel 2.5 - real network documents (UT Dallas), Phase A
+
+2400 chunks of **real, measured** network records for the seven OTel categories, cleaned and
+pseudonymised, each in a labelled and a label-stripped version. These are the documents the Phase C Q&A rows on real data
+are built from (same `source_id`) and the seeds of the Phase B synthetic packets.
+
+| category | chunks | runs | records (rows) |
+|---|---:|---:|---:|
+| lte-signalling | 792 | 70 | 46,261 |
+| rf-signal-quality | 266 | 60 | 14,250 |
+| 5g-core-signalling | 266 | 126 | 13,916 |
+| ran-telemetry | 776 | 127 | 38,232 |
+| network-performance | 63 | 63 | 994 |
+| faults-ground-truth | 155 | 97 | 6,643 |
+| aaa-radius | 82 | 20 | 4,122 |
+
+## Sources
+
+| collection | what it is | chunks |
+|---|---|---:|
+| `powder-lte-v1` | measured over real SDR radios on POWDER (srsRAN_4G) | 993 |
+| `zmq-5g-core-sim-v1` | real core, gNB and UE software; radio link simulated over ZMQ (no RF): radio values describe the simulated link | 669 |
+| `powder-5g-core-sim-v1` | real Open5GS core on POWDER; UE and gNB simulated by UERANSIM (no RF) | 31 |
+| `ota-5g-b210` | measured over the air between two USRP B210s in the lab (ocudu gNB, srsUE, Open5GS); radio values uncalibrated (relative) | 527 |
+| `ota-lte-b210` | measured over the air between two USRP B210s in the lab (srsENB, srsUE, Open5GS EPC); radio values uncalibrated (relative) | 96 |
+| `local-radius` | measured RADIUS exchanges between FreeRADIUS 3.2.5 and a scripted NAS; synthetic test users | 82 |
+| `ota-spectrum-b210` | measured IQ from a USRP B210 receiver; power in dBFS, uncalibrated (no absolute dBm) | 2 |
+
+## What was done (Phase A)
+
+The raw captures and logs of every run were parsed with tshark-based parsers and rendered into self-contained text
+documents (a header with run id, collection, provenance, time base and identifier note, then one or more tables with a
+`columns:` line). All of AT&T's batch-1 blockers are fixed; `reports/feedback_fixes.md` lists each item and its handling.
+In short: no corrupted values, one pseudonym per subscriber across all interfaces, column headers in every chunk, at
+least 10 records per chunk, srsUE fields named for what they are, no paths or host names, label-stripped variants,
+fixed units and precision, at most 20 chunks per run and category, every chunk under ~3,850 tokens, `run_id` and a
+split by run on every chunk. Third-party phones that reached the over-the-air test cells are removed (only our test SIMs
+remain). Every chunk was re-checked when packaged: `reports/spec_check.json` (0 errors).
+
+## Fields (data/<category>.jsonl, one JSON object per line)
+
+`source_repo`, `source_id` (the chunk id), `source_spec_id` / `source_version` (null), `run_id`, `split`
+(train / validation / test by run), `category`, `scenario` (the run's measured scenario label), `provenance`,
+`n_records`, `tables` (table names, columns and row counts), `text` (labelled: scenario label, description and the
+orchestrator's ground-truth timeline in the header), `text_stripped` (the same records with no scenario label, timeline
+or frame labels, for detection and root-cause use).
+
+## Known limits
+
+- Radio values depend on the collection: ZMQ and UERANSIM runs have a simulated radio link; the B210 runs are over the
+  air but uncalibrated (relative values only); POWDER LTE runs used real SDR radios.
+- LTE over-the-air RSRP above -44 dBm (very strong bench link) is kept and noted in the header (3,972 values).
+- v1 ZMQ and early over-the-air runs had no internet access from the UE (host forwarding off): internet tests fail there.
+- In most 5G runs the first registration starts by releasing the subscriber's PDU session left from the previous run;
+  those rows are marked `stale-session-cleanup`, with a header note.
+- Not included: the parsed per-record NDJSON (3.9 GB, needs a separate third-party filtering pass before sharing) and
+  the raw captures (local only: they contain identities of third-party phones).

deliveries/phaseA_real_documents_2026-10-24/data/5g-core-signalling.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b5c6be8577c2474b1fc3cf81f07887c11dcf1bb6f84c093815a0aae4a2ceffe
+size 4028788

deliveries/phaseA_real_documents_2026-10-24/data/aaa-radius.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19ea8d1e3d844a446fc93e2c7fbfb86a35dd5fe440bb71bb5d34f73903ab2dff
+size 1212756

deliveries/phaseA_real_documents_2026-10-24/data/faults-ground-truth.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f021f497114a080cceb1304c529626342024997677be1caa745c8450f45527d0
+size 1710314

deliveries/phaseA_real_documents_2026-10-24/data/lte-signalling.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a728ca537ad26cea850493c2490cd3e6c889751f2d7c3245b8aead835b0cc70
+size 11017608

deliveries/phaseA_real_documents_2026-10-24/data/network-performance.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:13123c21b38f43d30cd3c9b5246517cd2b640cea80831c296c09ee92f6361cdb
+size 462722

deliveries/phaseA_real_documents_2026-10-24/data/ran-telemetry.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3614d65960f8e4605d4d980b16fc55b3315e12ebc57781f4441fd83e1e823808
+size 8756618

deliveries/phaseA_real_documents_2026-10-24/data/rf-signal-quality.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7131f61386eb1571db6064fd10d08e5abd6afbc4c2f4675c6a5ebe8ae86506e0
+size 3078129

deliveries/phaseA_real_documents_2026-10-24/reports/distribution.md

@@ -0,0 +1,67 @@
+# Document distribution (scenario x category)
+
+Cells: chunks / runs.
+
+| collection | scenario | lte-signalling | rf-signal-quality | 5g-core-signalling | ran-telemetry | network-performance | faults-ground-truth | aaa-radius |
+|---|---|---:|---:|---:|---:|---:|---:|---:|
+| local-radius | radius-acct-load | - | - | - | - | - | - | 40 / 2 |
+| local-radius | radius-auth-methods | - | - | - | - | - | - | 4 / 2 |
+| local-radius | radius-baseline | - | - | - | - | - | - | 6 / 2 |
+| local-radius | radius-bruteforce | - | - | - | - | - | - | 8 / 2 |
+| local-radius | radius-reject-delay-0 | - | - | - | - | - | - | 4 / 2 |
+| local-radius | radius-secret-mismatch | - | - | - | - | - | - | 4 / 2 |
+| local-radius | radius-server-down | - | - | - | - | - | - | 4 / 2 |
+| local-radius | radius-session-timeout | - | - | - | - | - | - | 4 / 2 |
+| local-radius | radius-unknown-user | - | - | - | - | - | - | 4 / 2 |
+| local-radius | radius-wrong-password | - | - | - | - | - | - | 4 / 2 |
+| ota-5g-b210 | ota-ausf-stop | - | - | 3 / 1 | 8 / 1 | 1 / 1 | 1 / 1 | - |
+| ota-5g-b210 | ota-baseline | - | - | 5 / 2 | 17 / 2 | 2 / 2 | 2 / 2 | - |
+| ota-5g-b210 | ota-gnb-low-tx | - | - | 3 / 1 | 6 / 1 | 1 / 1 | 1 / 1 | - |
+| ota-5g-b210 | ota-low-mcs | - | - | 2 / 1 | 6 / 1 | 1 / 1 | 1 / 1 | - |
+| ota-5g-b210 | ota-mid-mcs | - | - | 3 / 1 | 7 / 1 | 1 / 1 | 1 / 1 | - |
+| ota-5g-b210 | ota-mixed | - | - | 62 / 19 | 246 / 19 | 19 / 19 | 28 / 19 | - |
+| ota-5g-b210 | ota-n3-netem | - | - | 2 / 1 | 10 / 1 | 1 / 1 | 1 / 1 | - |
+| ota-5g-b210 | ota-nrf-stop | - | - | 3 / 1 | 7 / 1 | 1 / 1 | 9 / 1 | - |
+| ota-5g-b210 | ota-reattach | - | - | 6 / 1 | 6 / 1 | 1 / 1 | 1 / 1 | - |
+| ota-5g-b210 | ota-smf-restart | - | - | 3 / 1 | 7 / 1 | - | 3 / 1 | - |
+| ota-5g-b210 | ota-udm-stop | - | - | 2 / 1 | 9 / 1 | 1 / 1 | 1 / 1 | - |
+| ota-5g-b210 | ota-udp-load | - | - | 4 / 1 | 8 / 1 | - | 1 / 1 | - |
+| ota-5g-b210 | ota-ue-low-tx | - | - | 4 / 1 | 7 / 1 | 1 / 1 | 1 / 1 | - |
+| ota-lte-b210 | lte-baseline | 2 / 2 | 13 / 2 | - | 6 / 2 | 2 / 2 | 2 / 2 | - |
+| ota-lte-b210 | lte-enb-low-tx | 1 / 1 | 9 / 1 | - | 3 / 1 | 1 / 1 | - | - |
+| ota-lte-b210 | lte-hss-stop | 2 / 1 | 6 / 1 | - | 2 / 1 | 1 / 1 | 1 / 1 | - |
+| ota-lte-b210 | lte-idle-service | 4 / 1 | 5 / 1 | - | 2 / 1 | 1 / 1 | - | - |
+| ota-lte-b210 | lte-mme-restart | 2 / 1 | 3 / 1 | - | 2 / 1 | 1 / 1 | 1 / 1 | - |
+| ota-lte-b210 | lte-reattach | 3 / 1 | 5 / 1 | - | 2 / 1 | 1 / 1 | 1 / 1 | - |
+| ota-lte-b210 | lte-sgwc-restart | 2 / 1 | 6 / 1 | - | 2 / 1 | 1 / 1 | 1 / 1 | - |
+| ota-spectrum-b210 | spectrum-sweep | - | 2 / 2 | - | - | - | - | - |
+| powder-5g-core-sim-v1 | amf-restart | - | - | 3 / 1 | - | - | 1 / 1 | - |
+| powder-5g-core-sim-v1 | baseline | - | - | 8 / 3 | - | - | 3 / 3 | - |
+| powder-5g-core-sim-v1 | n2-netem | - | - | 5 / 1 | - | - | 1 / 1 | - |
+| powder-5g-core-sim-v1 | radius | - | - | 3 / 1 | - | - | 1 / 1 | - |
+| powder-5g-core-sim-v1 | smf-restart | - | - | 5 / 1 | - | - | 1 / 1 | - |
+| powder-lte-v1 | benign | 436 / 34 | 141 / 28 | - | - | - | - | - |
+| powder-lte-v1 | fake | 340 / 28 | 76 / 22 | - | - | - | - | - |
+| zmq-5g-core-sim-v1 | amf-restart | - | - | 12 / 6 | 36 / 6 | 6 / 6 | 6 / 6 | - |
+| zmq-5g-core-sim-v1 | ausf-stop | - | - | 6 / 6 | 18 / 6 | - | 6 / 6 | - |
+| zmq-5g-core-sim-v1 | baseline | - | - | 11 / 11 | 39 / 11 | - | - | - |
+| zmq-5g-core-sim-v1 | high-load | - | - | 6 / 6 | 42 / 6 | - | - | - |
+| zmq-5g-core-sim-v1 | low-mcs | - | - | 6 / 6 | 36 / 6 | - | - | - |
+| zmq-5g-core-sim-v1 | mid-mcs | - | - | 6 / 6 | 36 / 6 | - | - | - |
+| zmq-5g-core-sim-v1 | nrf-stop | - | - | 14 / 7 | 21 / 7 | - | 40 / 7 | - |
+| zmq-5g-core-sim-v1 | pcf-restart | - | - | 12 / 6 | 36 / 6 | 6 / 6 | 6 / 6 | - |
+| zmq-5g-core-sim-v1 | reattach | - | - | 28 / 7 | 19 / 7 | - | 7 / 7 | - |
+| zmq-5g-core-sim-v1 | smf-restart | - | - | 14 / 7 | 39 / 7 | 7 / 7 | 13 / 7 | - |
+| zmq-5g-core-sim-v1 | udm-stop | - | - | 6 / 6 | 18 / 6 | - | 6 / 6 | - |
+| zmq-5g-core-sim-v1 | udp-load | - | - | 6 / 6 | 36 / 6 | - | - | - |
+| zmq-5g-core-sim-v1 | upf-restart | - | - | 13 / 7 | 37 / 7 | 6 / 6 | 7 / 7 | - |
+
+| category | train | validation | test |
+|---|---:|---:|---:|
+| lte-signalling | 708 | 55 | 29 |
+| rf-signal-quality | 250 | 14 | 2 |
+| 5g-core-signalling | 214 | 23 | 29 |
+| ran-telemetry | 610 | 70 | 96 |
+| network-performance | 47 | 6 | 10 |
+| faults-ground-truth | 126 | 13 | 16 |
+| aaa-radius | 74 | 6 | 2 |

deliveries/phaseA_real_documents_2026-10-24/reports/feedback_fixes.md

@@ -0,0 +1,29 @@
+# How the batch-1 feedback (feedback_batch1.md) is handled
+
+
+| Item | Handling |
+|---|---|
+| 1.1 corrupted floats / hex | the scan matches whole tokens only; floats rounded per column; checker and tests fail on `\d\.(ue|user)-`, `0x(ue|user)-` |
+| 1.2 two pseudonyms per subscriber | NGAP and 5G NAS keyed the MSIN, the rest the IMSI: the MSIN is now keyed as PLMN+MSIN. Verified: no 5G run has more than one subscriber pseudonym |
+| 1.3 column headers / small chunks | every table in every chunk has its `columns:` line; chunks under 10 records are merged or dropped |
+| 1.4 srsUE fields | `rsrp_raw` (srsUE's NR value is not dBm), `*_mcs_avg`; the gNB's own per-UE scheduler metrics are the primary RAN telemetry |
+| 1.5 RADIUS reject with accept reply | real FreeRADIUS behaviour (authorize-time reply items sent with the reject); the Step 1 server strips Reply-Message on reject; the v1 RADIUS run is excluded |
+| 1.6 signalling order (low-mcs) | real: the core keeps the subscriber's PDU session between runs, so the first registration of 117/126 5G runs starts by releasing it (SM context release, PFCP Session Deletion, PCF/UDM/UDR clean-up). Those rows are marked `stale-session-cleanup`, with a header note |
+| 1.7 paths / host names | dropped or replaced; checker fails on any path or interface name |
+| 3.3 label shortcuts | every chunk rendered `labelled` and `stripped` (no scenario label, description, ground-truth timeline or frame labels) |
+| 3.4 label noise | `awgn-15` / `awgn-8` had no effect on NR (DL SNR 70, BLER 0, MCS 28 like the baselines): relabelled baseline-equivalent with a note |
+| 4.1 cap per run, distribution | at most 20 chunks per run and category (evenly spread); `documents/_distribution.md` per delivery |
+| 4.2 formatting | ISO date in the header, HH:MM:SS.mmm in rows, units in column names, fixed precision, `-` for missing |
+| 4.3 splits | `run_id` and `split` (train/validation/test by run hash) on every chunk; synthetic data only from train runs |
+| row size | chunks sized to <= 9,600 characters (~3,850 tokens) |
+| ground truth vs observations | the ground-truth timeline holds only scenario steps and injected conditions; registrations, app results and link losses are observations |
+
+## Decisions recorded
+
+- LTE over-the-air RSRP above -44 dBm (strong bench link, within 3GPP's extended range) is kept, flagged in the document
+  header and reported as a warning (AT&T's suggested range is -140 to -44 dBm).
+- Third-party phones near the bench attached to the LTE cell (3 runs) and tried to register on the 5G cell (1 run):
+  their records are dropped (allowlist of the test SIMs); the checker fails any document with another subscriber.
+  Raw files stay local (see `raw/ota_lte/THIRD_PARTY_IDENTITIES.md`).
+- The over-the-air gNB caps only PDSCH MCS (20); PUSCH runs at the default 28 (ocudu `du_high_config.h`).
+

deliveries/phaseA_real_documents_2026-10-24/reports/spec_check.json

@@ -0,0 +1,9 @@
+{
+ "chunks": 2400,
+ "documents_checked": 4800,
+ "errors": 0,
+ "warnings": {
+  "value-range-att": 3972
+ },
+ "warning_note": "value-range-att = LTE over-the-air RSRP above -44 dBm (strong bench link, within the 3GPP extended range); kept and noted in each document header"
+}

deliveries/phaseB_synthetic_packets_2026-10-24/README.md

@@ -0,0 +1,57 @@
+# OTel 2.5 - synthetic packet data (UT Dallas), Phase B
+
+1023 validated synthetic documents ("packets": tables of protocol messages, log lines, metrics samples, tests or
+RADIUS exchanges) for the seven OTel categories, generated from the measured UT Dallas collections and validated by code
+checks plus an independent LLM judge. **Synthetic, not measured**: every document says so in its header and fields.
+Human spot check: in progress (UT Dallas team).
+
+| category | documents | per real train run | records (rows) | growth vs. real records |
+|---|---:|---:|---:|---:|
+| lte-signalling | 183 | 3.0x | 6,922 | +15% |
+| rf-signal-quality | 124 | 2.3x | 4,359 | +31% |
+| 5g-core-signalling | 200 | 1.9x | 7,562 | +54% |
+| ran-telemetry | 161 | 1.5x | 5,454 | +14% |
+| network-performance | 100 | 2.1x | 1,506 | +152% |
+| faults-ground-truth | 219 | 2.7x | 5,700 | +86% |
+| aaa-radius | 36 | 2.1x | 1,163 | +28% |
+
+## How it was made
+
+1. **Real seeds.** The measured collections (POWDER LTE, local ZMQ 5G, POWDER 5G core with UERANSIM, over-the-air 5G and
+   LTE on two USRP B210s, FreeRADIUS) were parsed with the batch-1 fixes and rendered into documents per category
+   (pseudonymised identifiers, column headers in every chunk, fixed units and precision, run_id and a split by run).
+   Only **train-split** runs are used as seeds, so no synthetic document derives from a validation or test run.
+2. **Generation** (`deepseek/deepseek-v4.1-flash`, temperature 1.0). For every train run, 3 candidates (5 for
+   lte-signalling and faults-ground-truth), each anchored on that run's own document with 2 other real runs of the same
+   measured scenario as references. The generator keeps the seed's structure (same procedures, cycles and tests, in the
+   same order) and varies what varies between real runs: timing gaps, identifiers, measured values within the observed
+   spread. One time offset is applied by code to every timestamp. No new conditions are invented: the scenario label is
+   the seed's. Prompt: `method/generator_prompt.txt`.
+3. **Validation.** (a) Code checks (the same checks as the real data, built from `OTel_data_label.md` and
+   `feedback_batch1.md`): format, the seed's exact tables and columns, per-column ranges, identifier format, no label in
+   the stripped text, no paths, size, and not a copy of the seed (no more than 50 % of rows at the same relative times,
+   and for measurement categories no more than 50 % of rows with the seed's values). (b) An independent judge from a
+   different model family (`anthropic/claude-opus-5.5`) scores protocol validity, internal consistency, physical
+   plausibility, realism and agreement with the scenario (1-5) and lists problems; kept only if it says keep, lists no
+   major problem and every score is at least 3. Prompt: `method/judge_prompt.txt`. Judge calibration
+   (`reports/judge_calibration.json`): 18/20 planted errors caught, 0/24 real documents wrongly removed.
+   Candidates removed as copies or for a rendering bug were repaired once and validated again.
+4. **Result.** 2,156 candidates (1,694 + 462 repairs) -> 1,077 kept (64 % of the original candidates) -> 9 dropped
+   for length -> 45 removed because Phase C Q&A answers found an internal contradiction in them -> 1023 delivered. Details: `reports/validation_report.md`, `reports/distribution.md`.
+
+## Fields (data/<category>.jsonl, one JSON object per line)
+
+`source_repo` (`<collection>-synthetic`), `source_id` (unique chunk id, `<source_repo>__<id>`), `source_spec_id` /
+`source_version` (null), `run_id` (the synthetic run), `split` (always train), `category`, `scenario_label` (the seed's
+measured scenario), `provenance`, `generator`, `judge`, `seed_source_id` / `seed_run_id` (the real document it was
+written from), `judge_scores`, `judge_summary`, `text` (labelled document: scenario label, description and ground-truth
+timeline in the header), `text_stripped` (the same records without any label, for detection / root-cause use).
+
+## Known limits
+
+- Synthetic data varies within measured scenarios; it adds examples, not new failure types.
+- Values inherit the seed's provenance: documents from ZMQ / UERANSIM seeds describe a simulated radio link, documents
+  from B210 seeds an uncalibrated over-the-air lab link (relative values only).
+- The judge is an LLM; its calibration set is small (20 planted errors). A human spot check by the UT Dallas team is in progress.
+- The v1 ZMQ and early over-the-air runs had no internet access from the UE (host forwarding was off), so internet
+  pings and HTTP fail in them and in synthetic documents derived from them; core-gateway tests are unaffected.

deliveries/phaseB_synthetic_packets_2026-10-24/data/5g-core-signalling.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a91a734f9529b25debede058841ea1e747a0e7c541cdec7ed1c75fbf59e0bd5
+size 2560000

deliveries/phaseB_synthetic_packets_2026-10-24/data/aaa-radius.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:185ee5aa9e860564c4e1476e65c16a1deec3fa73519c596b84bfe1c1a408357b
+size 376660

deliveries/phaseB_synthetic_packets_2026-10-24/data/faults-ground-truth.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b15a7cafd1f56c0af097ffbebbca3c7d090ecfda48ed56d65961535864e4520f
+size 1852900

deliveries/phaseB_synthetic_packets_2026-10-24/data/lte-signalling.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9571b9747c3b8da1e41ba773ee23216e0efbc25978635b573a5b5f7077b9f2f4
+size 2209716

deliveries/phaseB_synthetic_packets_2026-10-24/data/network-performance.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f6818c3c6e25b033d999059d6eaf1f885732dd249e7fbe771fee3717c747660
+size 792689

deliveries/phaseB_synthetic_packets_2026-10-24/data/ran-telemetry.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:15e5193f48848cd995eb6756d687e9e654f5b010ec0a2eb16edfe897ebd36dce
+size 1633562

deliveries/phaseB_synthetic_packets_2026-10-24/data/rf-signal-quality.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1ab75fc46fb5e8a281f194cbe332de04b81d0ae4f324a290692caf7d8b612b3e
+size 1096753

deliveries/phaseB_synthetic_packets_2026-10-24/method/generator_prompt.txt

@@ -0,0 +1,46 @@
+You write synthetic telecom network records for a training dataset.
+
+You get 2-3 real, measured documents of ONE scenario (same system, same condition). Write ONE new document that
+looks like another real run of that same scenario. Do not invent a different condition.
+
+Anchor on the PRIMARY document:
+1. Keep its procedure skeleton: the same kinds of messages and steps in the same order, the same tables and columns.
+2. Keep the primary's structure exactly: the same number of cycles, procedures, traffic tests and log events, in the
+   same order, in every table and in the ground-truth timeline. Do not add, remove or move a cycle, test or event.
+   Vary only what really varies between runs of the same scenario (compare the reference documents to see how much):
+   - timestamps: keep the primary's clock (do NOT shift the run; that is done for you afterwards); change the gaps
+     between rows by small realistic amounts, moving every dependent row and timeline event with them;
+   - identifiers: new, consistent ones (pseudonyms "ue-"/"user-" + 16 lowercase hex, S1AP/NGAP ids, RNTI, SEID,
+     session ids, private IPs), the same value everywhere one entity appears;
+   - measured values (throughput, latency, RSRP/SNR, MCS, loss, counters) within the spread seen in the documents;
+3. Everything must stay consistent: tables and the ground-truth timeline agree; a value never outside its physical or
+   protocol range; "-" for a value the source would not report. Numbers are plain numbers in the unit of the column
+   name (e.g. 10.012 in a *_mbps column, never "10M" or "10 Mbps").
+4. Never a real IMSI/IMEI/MSISDN, public IP, hostname or file path. 10 to 40 rows per table.
+
+Answer with one JSON object only:
+{"variation": "<a few words: what you varied (internal note, not shown to anyone)>",
+ "header_notes": ["<the primary's configuration lines, unchanged unless a value in them must change>"],
+ "ground_truth": [["HH:MM:SS.mmm", "<event>", "<details>"], ...],
+ "tables": {"<table name>": [{"<column>": "<value>", ...}, ...], ...}}
+Each row is an object with exactly the listed columns as keys. Use [] for "ground_truth" if the primary has none.
+
+--- user message template ---
+Category: <category>. 
+Scenario (measured, fixed): <scenario>: <description>
+
+PRIMARY document (real):
+<<<
+<PRIMARY document>>>>
+
+Reference document 1 (real, same scenario):
+<<<
+<reference document>>>>
+
+Tables and columns to produce (row objects with exactly these keys):
+- "<table>": <col1> | <col2>
+
+Runs already generated for this scenario group (make this one differ from them):
+- (none yet)
+
+Write synthetic run #1 as the JSON object described.

deliveries/phaseB_synthetic_packets_2026-10-24/method/judge_prompt.txt

@@ -0,0 +1,22 @@
+You review synthetic telecom network records before they are used as training data.
+
+You get a real, measured seed document and a synthetic candidate written from it by another model. The candidate
+claims a condition in its "scenario:" line. Decide whether the candidate is good enough to keep.
+
+Judge on these criteria, each scored 1 (bad) to 5 (excellent):
+- protocol_validity: messages, procedures, causes and their order follow standard behaviour (3GPP / IETF / the
+  software in the seed) for the claimed condition.
+- internal_consistency: identifiers, timings, counters, causes and the ground-truth timeline agree with each other
+  across all tables.
+- physical_plausibility: values are possible for this system and radio setup (rates vs. MCS and bandwidth, latencies,
+  SNR/RSRP, loss, load), given the seed's configuration.
+- realism: the records look like they came from the same kind of system as the seed, not invented patterns or
+  suspiciously regular numbers.
+- matches_variation: the records actually show the condition the candidate claims.
+
+List every concrete problem you find with severity "major" (wrong or impossible: must not be learned) or "minor".
+Verdict "keep" only if there is no major problem.  Answer with one JSON object only:
+{"verdict": "keep" | "remove",
+ "scores": {"protocol_validity": n, "internal_consistency": n, "physical_plausibility": n, "realism": n, "matches_variation": n},
+ "problems": [{"severity": "major" | "minor", "where": "<table/row/field>", "issue": "<what is wrong>"}],
+ "summary": "<one or two sentences>"}

deliveries/phaseB_synthetic_packets_2026-10-24/reports/distribution.md

@@ -0,0 +1,66 @@
+# Synthetic packet distribution
+
+Documents per seed collection / measured scenario and category.
+
+| collection | scenario | lte-signalling | rf-signal-quality | 5g-core-signalling | ran-telemetry | network-performance | faults-ground-truth | aaa-radius |
+|---|---|---:|---:|---:|---:|---:|---:|---:|
+| local-radius | radius-acct-load | - | - | - | - | - | - | 6 |
+| local-radius | radius-auth-methods | - | - | - | - | - | - | 3 |
+| local-radius | radius-baseline | - | - | - | - | - | - | 5 |
+| local-radius | radius-bruteforce | - | - | - | - | - | - | 2 |
+| local-radius | radius-reject-delay-0 | - | - | - | - | - | - | 3 |
+| local-radius | radius-secret-mismatch | - | - | - | - | - | - | 2 |
+| local-radius | radius-server-down | - | - | - | - | - | - | 2 |
+| local-radius | radius-session-timeout | - | - | - | - | - | - | 6 |
+| local-radius | radius-unknown-user | - | - | - | - | - | - | 5 |
+| local-radius | radius-wrong-password | - | - | - | - | - | - | 2 |
+| ota-5g-b210 | ota-ausf-stop | - | - | 2 | 1 | 3 | 2 | - |
+| ota-5g-b210 | ota-baseline | - | - | 2 | 2 | 5 | 7 | - |
+| ota-5g-b210 | ota-gnb-low-tx | - | - | 3 | 2 | 3 | 1 | - |
+| ota-5g-b210 | ota-mid-mcs | - | - | 1 | 2 | 2 | 4 | - |
+| ota-5g-b210 | ota-mixed | - | - | 20 | 26 | 19 | 20 | - |
+| ota-5g-b210 | ota-n3-netem | - | - | 1 | 3 | 2 | 5 | - |
+| ota-5g-b210 | ota-nrf-stop | - | - | 1 | 2 | 3 | 4 | - |
+| ota-5g-b210 | ota-reattach | - | - | 2 | 3 | 1 | 1 | - |
+| ota-5g-b210 | ota-smf-restart | - | - | 2 | 2 | - | 2 | - |
+| ota-5g-b210 | ota-udp-load | - | - | 1 | 1 | - | 5 | - |
+| ota-lte-b210 | lte-baseline | 5 | 5 | - | 4 | 3 | 2 | - |
+| ota-lte-b210 | lte-hss-stop | 2 | 2 | - | 3 | 3 | 4 | - |
+| ota-lte-b210 | lte-idle-service | 3 | 3 | - | 2 | 3 | - | - |
+| ota-lte-b210 | lte-mme-restart | 2 | - | - | 1 | 2 | 4 | - |
+| ota-lte-b210 | lte-reattach | 3 | 3 | - | 3 | 3 | 3 | - |
+| ota-lte-b210 | lte-sgwc-restart | 2 | 3 | - | 2 | 3 | 2 | - |
+| ota-spectrum-b210 | spectrum-sweep | - | 6 | - | - | - | - | - |
+| powder-5g-core-sim-v1 | amf-restart | - | - | 1 | - | - | - | - |
+| powder-5g-core-sim-v1 | baseline | - | - | 6 | - | - | 9 | - |
+| powder-5g-core-sim-v1 | n2-netem | - | - | 2 | - | - | - | - |
+| powder-5g-core-sim-v1 | radius | - | - | 3 | - | - | 4 | - |
+| powder-5g-core-sim-v1 | smf-restart | - | - | - | - | - | 2 | - |
+| powder-lte-v1 | benign | 113 | 61 | - | - | - | - | - |
+| powder-lte-v1 | fake | 53 | 41 | - | - | - | - | - |
+| zmq-5g-core-sim-v1 | amf-restart | - | - | 9 | 8 | 11 | 13 | - |
+| zmq-5g-core-sim-v1 | ausf-stop | - | - | 13 | 11 | - | 26 | - |
+| zmq-5g-core-sim-v1 | baseline | - | - | 19 | 18 | - | - | - |
+| zmq-5g-core-sim-v1 | high-load | - | - | 14 | 4 | - | - | - |
+| zmq-5g-core-sim-v1 | low-mcs | - | - | 11 | 8 | - | - | - |
+| zmq-5g-core-sim-v1 | mid-mcs | - | - | 8 | 6 | - | - | - |
+| zmq-5g-core-sim-v1 | nrf-stop | - | - | 14 | 7 | - | 14 | - |
+| zmq-5g-core-sim-v1 | pcf-restart | - | - | 8 | 5 | 14 | 17 | - |
+| zmq-5g-core-sim-v1 | reattach | - | - | 15 | 9 | - | 25 | - |
+| zmq-5g-core-sim-v1 | smf-restart | - | - | 12 | 9 | 9 | 13 | - |
+| zmq-5g-core-sim-v1 | udm-stop | - | - | 13 | 9 | - | 17 | - |
+| zmq-5g-core-sim-v1 | udp-load | - | - | 8 | 5 | - | - | - |
+| zmq-5g-core-sim-v1 | upf-restart | - | - | 9 | 3 | 11 | 13 | - |
+
+## Growth against the real data
+
+| category | synthetic documents | real train runs | x per run | synthetic records | real records | record growth |
+|---|---:|---:|---:|---:|---:|---:|
+| lte-signalling | 183 | 62 | 3.0 | 6922 | 46261 | +15% |
+| rf-signal-quality | 124 | 54 | 2.3 | 4359 | 14250 | +31% |
+| 5g-core-signalling | 200 | 105 | 1.9 | 7562 | 13916 | +54% |
+| ran-telemetry | 161 | 105 | 1.5 | 5454 | 38232 | +14% |
+| network-performance | 100 | 47 | 2.1 | 1506 | 994 | +152% |
+| faults-ground-truth | 219 | 80 | 2.7 | 5700 | 6643 | +86% |
+| aaa-radius | 36 | 17 | 2.1 | 1163 | 4122 | +28% |
+| **total** | **1023** | **470** | | **32666** | **124418** | **+26%** |

deliveries/phaseB_synthetic_packets_2026-10-24/reports/dropped_too_long.json

@@ -0,0 +1,56 @@
+[
+ {
+  "candidate_id": "syn2-6f9a7a6a23ff",
+  "category": "lte-signalling",
+  "chars": 10499,
+  "reason": "too-long (>9,600 chars, row limit)"
+ },
+ {
+  "candidate_id": "syn2-051e0893b98c",
+  "category": "lte-signalling",
+  "chars": 14936,
+  "reason": "too-long (>9,600 chars, row limit)"
+ },
+ {
+  "candidate_id": "syn2-2c76023e24b7",
+  "category": "lte-signalling",
+  "chars": 10270,
+  "reason": "too-long (>9,600 chars, row limit)"
+ },
+ {
+  "candidate_id": "syn2-0dfd9d6ec154",
+  "category": "lte-signalling",
+  "chars": 9773,
+  "reason": "too-long (>9,600 chars, row limit)"
+ },
+ {
+  "candidate_id": "syn2-e38ee4d05e82-fx",
+  "category": "lte-signalling",
+  "chars": 12113,
+  "reason": "too-long (>9,600 chars, row limit)"
+ },
+ {
+  "candidate_id": "syn2-3ff3098b5aba-fx",
+  "category": "lte-signalling",
+  "chars": 12086,
+  "reason": "too-long (>9,600 chars, row limit)"
+ },
+ {
+  "candidate_id": "syn2-36be6116712a-fx",
+  "category": "lte-signalling",
+  "chars": 13208,
+  "reason": "too-long (>9,600 chars, row limit)"
+ },
+ {
+  "candidate_id": "syn2-4f33ea51149e-fx",
+  "category": "lte-signalling",
+  "chars": 12388,
+  "reason": "too-long (>9,600 chars, row limit)"
+ },
+ {
+  "candidate_id": "syn2-fa9ddbc2ae45-r1",
+  "category": "lte-signalling",
+  "chars": 10328,
+  "reason": "too-long (>9,600 chars, row limit)"
+ }
+]

deliveries/phaseB_synthetic_packets_2026-10-24/reports/generation_summary.json

@@ -0,0 +1,27 @@
+{
+ "lte-signalling": {
+  "generated": 309,
+  "unparseable_json": 1
+ },
+ "rf-signal-quality": {
+  "generated": 162
+ },
+ "5g-core-signalling": {
+  "generated": 312,
+  "unparseable_json": 3
+ },
+ "ran-telemetry": {
+  "generated": 310,
+  "unparseable_json": 5
+ },
+ "network-performance": {
+  "generated": 141
+ },
+ "faults-ground-truth": {
+  "generated": 397,
+  "unparseable_json": 3
+ },
+ "aaa-radius": {
+  "generated": 51
+ }
+}

deliveries/phaseB_synthetic_packets_2026-10-24/reports/judge_calibration.json

@@ -0,0 +1,29 @@
+{
+ "unmodified": {
+  "n": 24,
+  "correct": 24,
+  "rate": 1.0
+ },
+ "implausible-scale": {
+  "n": 3,
+  "correct": 2,
+  "rate": 0.667
+ },
+ "swap-order": {
+  "n": 9,
+  "correct": 8,
+  "rate": 0.889
+ },
+ "identity-switch": {
+  "n": 3,
+  "correct": 3,
+  "rate": 1.0
+ },
+ "timeline-contradiction": {
+  "n": 5,
+  "correct": 5,
+  "rate": 1.0
+ },
+ "planted_catch_rate": 0.9,
+ "false_remove_rate_real": 0.0
+}

deliveries/phaseB_synthetic_packets_2026-10-24/reports/removed_flagged_packets.json

@@ -0,0 +1,47 @@
+[
+ "local-radius-synthetic__syn2-12214cccf2c6",
+ "local-radius-synthetic__syn2-2a60fac37fce",
+ "local-radius-synthetic__syn2-a977f8d97c9f",
+ "local-radius-synthetic__syn2-d00fd18bc2a3",
+ "local-radius-synthetic__syn2-ff64480fe470",
+ "ota-5g-b210-synthetic__syn2-6a9822fd171c",
+ "ota-5g-b210-synthetic__syn2-a30aeed13452",
+ "ota-5g-b210-synthetic__syn2-b2bb1f2b3584",
+ "ota-5g-b210-synthetic__syn2-df1663338209",
+ "ota-5g-b210-synthetic__syn2-e2c3f76392f3",
+ "ota-5g-b210-synthetic__syn2-e4f44fd16691",
+ "ota-5g-b210-synthetic__syn2-e91dc26134ec",
+ "ota-5g-b210-synthetic__syn2-fb03cad0d5fc",
+ "ota-5g-b210-synthetic__syn2-fc54fc53846e",
+ "ota-5g-b210-synthetic__syn2-fd0acf537bfd",
+ "ota-lte-b210-synthetic__syn2-1348ae38bfdd",
+ "ota-lte-b210-synthetic__syn2-618417d4d6af-r1",
+ "ota-lte-b210-synthetic__syn2-702113c71443",
+ "powder-lte-v1-synthetic__syn2-3ca633672e30",
+ "powder-lte-v1-synthetic__syn2-78b9b3e1ac69",
+ "powder-lte-v1-synthetic__syn2-7aa49b0d046f",
+ "powder-lte-v1-synthetic__syn2-ebe42a986b7d",
+ "zmq-5g-core-sim-v1-synthetic__syn2-06cbbc874b35",
+ "zmq-5g-core-sim-v1-synthetic__syn2-1452d20fdad5",
+ "zmq-5g-core-sim-v1-synthetic__syn2-3652d33a0bd2",
+ "zmq-5g-core-sim-v1-synthetic__syn2-373e8201ae51",
+ "zmq-5g-core-sim-v1-synthetic__syn2-46410d599550-r1",
+ "zmq-5g-core-sim-v1-synthetic__syn2-58e4b4761add",
+ "zmq-5g-core-sim-v1-synthetic__syn2-5a703eaeae21",
+ "zmq-5g-core-sim-v1-synthetic__syn2-6fc237f3b3ae",
+ "zmq-5g-core-sim-v1-synthetic__syn2-86482b6ab4b6",
+ "zmq-5g-core-sim-v1-synthetic__syn2-93a2841cd651",
+ "zmq-5g-core-sim-v1-synthetic__syn2-965fbca4f96e",
+ "zmq-5g-core-sim-v1-synthetic__syn2-9aa641dda4d7",
+ "zmq-5g-core-sim-v1-synthetic__syn2-9ffa3f617069-fx",
+ "zmq-5g-core-sim-v1-synthetic__syn2-b5cd622a78ac",
+ "zmq-5g-core-sim-v1-synthetic__syn2-bb4430024200-r1",
+ "zmq-5g-core-sim-v1-synthetic__syn2-bc396ebeb1eb",
+ "zmq-5g-core-sim-v1-synthetic__syn2-be27d2bb0301",
+ "zmq-5g-core-sim-v1-synthetic__syn2-c765ead73bb7",
+ "zmq-5g-core-sim-v1-synthetic__syn2-ceaf2ee94876",
+ "zmq-5g-core-sim-v1-synthetic__syn2-e3582eb2e55e",
+ "zmq-5g-core-sim-v1-synthetic__syn2-e600847724b6",
+ "zmq-5g-core-sim-v1-synthetic__syn2-f369e603c3bd",
+ "zmq-5g-core-sim-v1-synthetic__syn2-fbb19cae0dfa"
+]

deliveries/phaseB_synthetic_packets_2026-10-24/reports/validation_report.json

@@ -0,0 +1,313 @@
+{
+ "judge": "anthropic/claude-opus-5.5",
+ "reasoning": {
+  "effort": "low"
+ },
+ "min_score": 3,
+ "copy_max": 0.5,
+ "categories": {
+  "lte-signalling": {
+   "candidates_generated": 456,
+   "judged": 456,
+   "removed_stage_a": 187,
+   "removed_stage_b": 74,
+   "kept": 195,
+   "selected": 195,
+   "rejection_rate": 0.572,
+   "top_removal_reasons": {
+    "shifted-copy-of-seed": 122,
+    "judge remove": 73,
+    "label-leak": 49,
+    "row-keys": 8,
+    "time-order": 5,
+    "score below 3": 1,
+    "value-identifier": 1,
+    "generator-unparseable_json": 1
+   },
+   "stage_a_rules": {
+    "shifted-copy-of-seed": 125,
+    "label-leak": 50,
+    "row-keys": 11,
+    "value-identifier": 10,
+    "time-order": 6,
+    "generator-unparseable_json": 1,
+    "chunk-size": 1
+   },
+   "judge_major_problem_examples": [
+    "The candidate is a verbatim replay of the seed. Every timestamp is shifted by the same constant offset (-03:06:12.000) and keeps identical millisecond fractions",
+    "S1 Setup has exactly the cycle-5 phase_start timestamp. That is 27 ms before enb_process_start (17:56:05.880), so the eNB sends S1 Setup before its process exis",
+    "Same defect for cycle 6: S1 Setup equals phase_start and comes before enb_process_start (17:56:37.511).",
+    "RRCConnectionRequest has exactly the ue_process_start timestamp. The seed shows about 2.8 s of cell search and sync between UE start and RRCConnectionRequest. A",
+    "The detach comes 4.2 s before ue_process_stop (17:56:28.401); the seed shows about 1 s. It also falls while the 8 s iperf started at 17:56:18.056 should still b",
+    "The eNB sends S1AP UECapabilityInfoIndication before it has received RRC UECapabilityInformation from the UE. The indication can only be sent after the capabili",
+    "The candidate is a verbatim copy of the seed. Every timestamp is shifted by exactly +02:00:37.000, and all millisecond offsets, message sequences, S1AP IDs (enb",
+    "The timeline shows the eNB process stopping at 13:41:37.479 and restarting at 13:41:42.511 (cycle-6). The seed shows that every eNB start is followed by an S1Se"
+   ],
+   "mean_scores_kept": {
+    "protocol_validity": 4.52,
+    "internal_consistency": 4.66,
+    "physical_plausibility": 4.96,
+    "realism": 3.5,
+    "matches_variation": 4.67
+   }
+  },
+  "rf-signal-quality": {
+   "candidates_generated": 194,
+   "judged": 194,
+   "removed_stage_a": 35,
+   "removed_stage_b": 33,
+   "kept": 126,
+   "selected": 126,
+   "rejection_rate": 0.351,
+   "top_removal_reasons": {
+    "judge remove": 33,
+    "shifted-copy-of-seed": 31,
+    "value-copy-of-seed": 4
+   },
+   "stage_a_rules": {
+    "value-copy-of-seed": 35,
+    "shifted-copy-of-seed": 31
+   },
+   "judge_major_problem_examples": [
+    "The iperf DL step is stamped after the downlink burst has already happened. The table shows DL bitrate of 3.4-7.3 Mbps from 17:32:55 to 17:33:04, and the DL rat",
+    "The iperf UL step falls inside the gap row, where no metrics are reported. The actual uplink burst (about 1.4-3.85 Mbps UL) runs from 17:33:25 to 17:33:35.",
+    "The HTTP step coincides with the start of the uplink-heavy burst, which is iperf UL behaviour rather than an HTTP download. Steps and metrics are misaligned.",
+    "The detach is placed exactly on a row that still shows 8.5 Mbps DL traffic at MCS 20. Active high-rate traffic at the moment of detach and bearer release is inc",
+    "The RSRQ sequence is a value-for-value copy of the seed (-3.0 x16, -3.5, -9.0, -4.5, -3.5, -7.5, -9.5 ... -5.0, -3.0, -3.0). Only the timestamps (shifted, with ",
+    "The candidate is a near-verbatim copy of the seed, with only uniform offsets applied. Every timestamp is shifted by a constant +11 ms (+8 ms on row 1). Every RN",
+    "PUSCH SINR drops from ~32.5 to 18.5 dB and PUCCH SINR from ~29.6 to 15.3 dB in a single isolated report. Serving RSRP (-71), RSRQ (-11.0), PHR (40.0), TA and ne",
+    "Uplink SINR on PCI 2 drops by about 15 dB (to ~18.4 / 15.3), while PCI 1 stays at ~32.5 / 29.6. In the seed, closed-loop uplink power control holds both cells a"
+   ],
+   "mean_scores_kept": {
+    "protocol_validity": 4.65,
+    "internal_consistency": 4.68,
+    "physical_plausibility": 4.64,
+    "realism": 3.47,
+    "matches_variation": 4.56
+   }
+  },
+  "5g-core-signalling": {
+   "candidates_generated": 332,
+   "judged": 332,
+   "removed_stage_a": 38,
+   "removed_stage_b": 67,
+   "kept": 227,
+   "selected": 227,
+   "rejection_rate": 0.316,
+   "top_removal_reasons": {
+    "judge remove": 66,
+    "shifted-copy-of-seed": 12,
+    "row-keys": 9,
+    "label-leak": 5,
+    "time-order": 3,
+    "chunk-size": 3,
+    "generator-unparseable_json": 3,
+    "value-identifier": 2
+   },
+   "stage_a_rules": {
+    "value-identifier": 20,
+    "row-keys": 12,
+    "shifted-copy-of-seed": 12,
+    "label-leak": 5,
+    "time-order": 3,
+    "chunk-size": 3,
+    "generator-unparseable_json": 3,
+    "value-time": 1
+   },
+   "judge_major_problem_examples": [
+    "Near-verbatim copy of the seed. Every row is shifted by exactly +04:32:18.000 with identical millisecond offsets, and only the subscriber pseudonym is swapped (",
+    "A fresh Registration request starts 0.2 s after the UE-originated deregistration. The timeline has no ue_process_start or re-register event there. The warmup UE",
+    "The table covers only the warmup deregistration and a spurious re-registration. It never shows the AMF restart at 06:19:59.165 (no SCTP/NG Setup loss, no new NG",
+    "UE deregistration does not delete the subscriber's subscription data in the UDR. Open5GS does not do this; it would wipe the provisioned subscriber.",
+    "A UE-authentication PUT to the UDM during deregistration is not standard behaviour. Context cleanup would use nudm-uecm/sdm unsubscribe, not ueau.",
+    "After the NAS Authentication response the AMF must confirm 5G-AKA with PUT nausf-auth/ue-authentications, as the seed does. The candidate skips this step and go",
+    "Ping and HTTP workload start before any Registration accept (01:42:01.357) or PDU session establishment (01:42:01.563). User-plane traffic cannot run without a ",
+    "The UE is reported deregistered before the registration it would deregister from completes, and before the PDU session request at 01:42:01.563. The signalling t"
+   ],
+   "mean_scores_kept": {
+    "protocol_validity": 4.56,
+    "internal_consistency": 4.53,
+    "physical_plausibility": 4.98,
+    "realism": 3.49,
+    "matches_variation": 3.94
+   }
+  },
+  "ran-telemetry": {
+   "candidates_generated": 436,
+   "judged": 436,
+   "removed_stage_a": 191,
+   "removed_stage_b": 80,
+   "kept": 165,
+   "selected": 165,
+   "rejection_rate": 0.622,
+   "top_removal_reasons": {
+    "value-copy-of-seed": 82,
+    "judge remove": 79,
+    "shifted-copy-of-seed": 71,
+    "label-leak": 18,
+    "row-keys": 9,
+    "generator-unparseable_json": 5,
+    "chunk-size": 2,
+    "value-time": 2
+   },
+   "stage_a_rules": {
+    "value-copy-of-seed": 164,
+    "shifted-copy-of-seed": 78,
+    "label-leak": 20,
+    "value-time": 11,
+    "row-keys": 10,
+    "generator-unparseable_json": 5,
+    "chunk-size": 2,
+    "missing-table": 1
+   },
+   "judge_major_problem_examples": [
+    "The timeline has the UL iperf running from 02:06:38 to about 02:06:48, followed by http and ping. No DL iperf is active until 02:06:55.402, yet these rows show ",
+    "DL holds at 59.2 Mbps while the TCP iperf target is 40M. The seed's 10M target produced about 12 Mbps, so a 40M TCP flow should not deliver roughly 50% above it",
+    "dl_bitrate is exactly 59.200 in all 20 rows, and dl_prbs alternates only between 90340 and 90330. Real seed rows vary from second to second (11.0-12.1 Mbps, 167",
+    "DL bitrate is 1.09 Mbps with 1687 PRBs and BSR 264, but the downlink iperf does not start until 07:06:33.254. The row reports traffic before the application tha",
+    "The timeline is shifted by about +1h01m00.47s from the seed, but the table is shifted by about +1h01m02s. The table therefore lags the ground truth by about 2 s",
+    "The row still shows 8.75 Mbps UL with BSR 196 bytes. The 5 s uplink iperf started at 10:34:19.403 and should have ended around 10:34:24.4, so this traffic falls",
+    "A per-UE scheduler row is reported after ue_process_stop at 10:34:32.243. The UE process is gone, so the gNB should not be reporting active UE scheduling at tha",
+    "The table is a near-verbatim copy of the seed with mechanical perturbations. pusch_snr_db, timing_advance_ns, dl_prbs (after row 6) and many bitrates are each s"
+   ],
+   "mean_scores_kept": {
+    "protocol_validity": 4.54,
+    "internal_consistency": 4.52,
+    "physical_plausibility": 4.67,
+    "realism": 3.42,
+    "matches_variation": 4.12
+   }
+  },
+  "network-performance": {
+   "candidates_generated": 159,
+   "judged": 159,
+   "removed_stage_a": 39,
+   "removed_stage_b": 17,
+   "kept": 103,
+   "selected": 103,
+   "rejection_rate": 0.352,
+   "top_removal_reasons": {
+    "judge remove": 17,
+    "shifted-copy-of-seed": 17,
+    "time-order": 9,
+    "value-copy-of-seed": 7,
+    "row-keys": 3,
+    "value-type": 1,
+    "label-leak": 1,
+    "chunk-size": 1
+   },
+   "stage_a_rules": {
+    "value-copy-of-seed": 19,
+    "shifted-copy-of-seed": 17,
+    "time-order": 10,
+    "row-keys": 5,
+    "value-type": 4,
+    "label-leak": 3,
+    "chunk-size": 2
+   },
+   "judge_major_problem_examples": [
+    "The timeline starts a downlink iperf3 test (direction=downlink, seconds=10). The matching app_tests row reports a failed uplink test ('iperf3 uplink -'). The tw",
+    "The injected fault was changed from SMF to AMF, but the user-plane outage that follows was copied unchanged from the SMF case. An AMF restart in Open5GS does no",
+    "All app_tests rows are stamped 04:02:51\u201304:08:28, but the timeline runs 07:34:19\u201307:40:40. The results are offset by about 3.5 hours, so no test result lines up",
+    "The fault target was changed from SMF to AMF, but the outcome copied from the SMF seed was kept unchanged: permanent loss of the user plane and 'unable to conne",
+    "The result is logged 9.563 s after the request started, but total_ms=9775.0. The test would have finished about 0.21 s after its own result row, which is imposs",
+    "A PCF restart in Open5GS does not tear down an established PDU session or the UPF data path. The seed shows UL/DL iperf continuing normally after the PCF restar",
+    "These rows are labelled 'uplink -', but the timeline starts them as downlink tests (downlink tcp 5M, downlink udp 8M, downlink tcp 5M). Test direction is incons",
+    "There are two consecutive ue_process_stop/gnb_process_stop pairs with no process start between them, so the processes are stopped twice."
+   ],
+   "mean_scores_kept": {
+    "protocol_validity": 4.53,
+    "internal_consistency": 4.35,
+    "physical_plausibility": 4.83,
+    "realism": 3.66,
+    "matches_variation": 4.43
+   }
+  },
+  "faults-ground-truth": {
+   "candidates_generated": 524,
+   "judged": 524,
+   "removed_stage_a": 202,
+   "removed_stage_b": 102,
+   "kept": 220,
+   "selected": 220,
+   "rejection_rate": 0.58,
+   "top_removal_reasons": {
+    "shifted-copy-of-seed": 132,
+    "judge remove": 102,
+    "row-keys": 24,
+    "label-leak": 22,
+    "time-order": 10,
+    "chunk-size": 6,
+    "generator-unparseable_json": 4,
+    "value-time": 3
+   },
+   "stage_a_rules": {
+    "shifted-copy-of-seed": 139,
+    "row-keys": 31,
+    "label-leak": 22,
+    "time-order": 21,
+    "value-time": 21,
+    "chunk-size": 6,
+    "generator-unparseable_json": 4,
+    "value-identifier": 2
+   },
+   "judge_major_problem_examples": [
+    "Log rows are timestamped after ran_collect_stop (19:22:15.597). The timeline was compressed but the logs kept the seed's original spacing, so the logs no longer",
+    "This SM context error comes after the UE was deregistered (19:22:04.893) and after the UE and gNB processes stopped (19:22:08.741 and 19:22:10.598). No SM conte",
+    "These log events no longer line up with the scenario steps. AMF/AUSF endpoint teardown and the SmContext error happen after process stop instead of during the s",
+    "Near-verbatim copy of the seed shifted by about 6.5 h. Every log line, NF, sequence, address (127.0.0.11/12) and inter-event gap is reproduced, so it adds no ne",
+    "Reuses the seed's exact subscriber pseudonym supi=imsi-ue-4549df2477a8c074, which links the synthetic record directly to the measured subscriber.",
+    "PFCP recovery timestamps were copied verbatim from the seed. 4000953940 in NTP seconds is about 2026-10-14 08:05:40 UTC. That is roughly 5.6 h after this row's ",
+    "The candidate is the seed shifted by a constant offset of about -5:36:36.707. Every log text, gap, SM context ID (9), AMF_UE_NGAP_ID (10), NF instance UUID and ",
+    "The 'gmm_state_security_mode: Expectation failed' error appears with no cause. In Open5GS it follows a failed NGAP send to a removed gNB. The seed shows that ch"
+   ],
+   "mean_scores_kept": {
+    "protocol_validity": 4.67,
+    "internal_consistency": 4.5,
+    "physical_plausibility": 4.92,
+    "realism": 3.65,
+    "matches_variation": 4.66
+   }
+  },
+  "aaa-radius": {
+   "candidates_generated": 55,
+   "judged": 55,
+   "removed_stage_a": 10,
+   "removed_stage_b": 4,
+   "kept": 41,
+   "selected": 41,
+   "rejection_rate": 0.255,
+   "top_removal_reasons": {
+    "shifted-copy-of-seed": 5,
+    "judge remove": 4,
+    "value-copy-of-seed": 2,
+    "value-identifier": 1,
+    "chunk-size": 1,
+    "row-keys": 1
+   },
+   "stage_a_rules": {
+    "value-copy-of-seed": 9,
+    "shifted-copy-of-seed": 7,
+    "row-keys": 3,
+    "value-identifier": 1,
+    "chunk-size": 1
+   },
+   "judge_major_problem_examples": [
+    "This user is rejected at 12:54:38 with 'No Auth-Type found', the unknown-user rejection. The same user then gets Login OK at 12:54:42 and 12:54:49. A user name ",
+    "Timeline events are stamped 17:54:37.000, but the log events they describe are at 12:54:37.000. Both claim the same UTC time base, so there is a 5-hour mismatch",
+    "The timeline says user-71f3b8d29c4e5a06 was accepted on port 30. In the log, that user appears only once, on port 9 at request (8). Port 30, request (29), is us",
+    "Near-verbatim copy of the seed. Every row matches: the same sequence of Accept and Reject per slot, the same nas_port numbering 1-20, identical millisecond offs",
+    "Two different users, user-3ba71c9f62d0a48e and user-6ec50a8b1437f9d2, both get 'Welcome otel-u07'. In the seed, each user pseudonym maps to exactly one otel acc",
+    "session_time_s=0 but 1.318 s have elapsed since the Start at 05:11:24.461. The other sessions, and the seed, report whole elapsed seconds rounded down, which wo",
+    "session_time_s=1 at 2.642 s and 2 at 3.967 s are one second too low. In the same document, user-3a7c1f0b9d4e28a6 reports 3 at 3.996 s and user-8f1a63d5c9b47e20 "
+   ],
+   "mean_scores_kept": {
+    "protocol_validity": 4.83,
+    "internal_consistency": 4.61,
+    "physical_plausibility": 4.88,
+    "realism": 3.59,
+    "matches_variation": 4.51
+   }
+  }
+ }
+}

deliveries/phaseB_synthetic_packets_2026-10-24/reports/validation_report.md

@@ -0,0 +1,110 @@
+# Synthetic data validation report
+
+Judge: `anthropic/claude-opus-5.5` (reasoning {'effort': 'low'}); generator: `deepseek/deepseek-v4.1-flash`. Counts are from the validation run, before 9 packets were dropped for length and 45 removed after Phase C (see the README). Keep rule: verdict keep, no major problem, every score >= 3; at most 50 % rows shared with the seed.
+
+| category | generated | judged | removed (code) | removed (judge) | kept | selected | rejection rate |
+|---|---:|---:|---:|---:|---:|---:|---:|
+| lte-signalling | 456 | 456 | 187 | 74 | 195 | 195 | 0.572 |
+| rf-signal-quality | 194 | 194 | 35 | 33 | 126 | 126 | 0.351 |
+| 5g-core-signalling | 332 | 332 | 38 | 67 | 227 | 227 | 0.316 |
+| ran-telemetry | 436 | 436 | 191 | 80 | 165 | 165 | 0.622 |
+| network-performance | 159 | 159 | 39 | 17 | 103 | 103 | 0.352 |
+| faults-ground-truth | 524 | 524 | 202 | 102 | 220 | 220 | 0.58 |
+| aaa-radius | 55 | 55 | 10 | 4 | 41 | 41 | 0.255 |
+
+### lte-signalling
+
+Removal reasons: {'shifted-copy-of-seed': 122, 'judge remove': 73, 'label-leak': 49, 'row-keys': 8, 'time-order': 5, 'score below 3': 1, 'value-identifier': 1, 'generator-unparseable_json': 1}
+Code-check rules: {'shifted-copy-of-seed': 125, 'label-leak': 50, 'row-keys': 11, 'value-identifier': 10, 'time-order': 6, 'generator-unparseable_json': 1, 'chunk-size': 1}
+Major problems found by the judge (examples):
+- The candidate is a verbatim replay of the seed. Every timestamp is shifted by the same constant offset (-03:06:12.000) and keeps identical millisecond fractions
+- S1 Setup has exactly the cycle-5 phase_start timestamp. That is 27 ms before enb_process_start (17:56:05.880), so the eNB sends S1 Setup before its process exis
+- Same defect for cycle 6: S1 Setup equals phase_start and comes before enb_process_start (17:56:37.511).
+- RRCConnectionRequest has exactly the ue_process_start timestamp. The seed shows about 2.8 s of cell search and sync between UE start and RRCConnectionRequest. A
+- The detach comes 4.2 s before ue_process_stop (17:56:28.401); the seed shows about 1 s. It also falls while the 8 s iperf started at 17:56:18.056 should still b
+- The eNB sends S1AP UECapabilityInfoIndication before it has received RRC UECapabilityInformation from the UE. The indication can only be sent after the capabili
+- The candidate is a verbatim copy of the seed. Every timestamp is shifted by exactly +02:00:37.000, and all millisecond offsets, message sequences, S1AP IDs (enb
+- The timeline shows the eNB process stopping at 13:41:37.479 and restarting at 13:41:42.511 (cycle-6). The seed shows that every eNB start is followed by an S1Se
+
+### rf-signal-quality
+
+Removal reasons: {'judge remove': 33, 'shifted-copy-of-seed': 31, 'value-copy-of-seed': 4}
+Code-check rules: {'value-copy-of-seed': 35, 'shifted-copy-of-seed': 31}
+Major problems found by the judge (examples):
+- The iperf DL step is stamped after the downlink burst has already happened. The table shows DL bitrate of 3.4-7.3 Mbps from 17:32:55 to 17:33:04, and the DL rat
+- The iperf UL step falls inside the gap row, where no metrics are reported. The actual uplink burst (about 1.4-3.85 Mbps UL) runs from 17:33:25 to 17:33:35.
+- The HTTP step coincides with the start of the uplink-heavy burst, which is iperf UL behaviour rather than an HTTP download. Steps and metrics are misaligned.
+- The detach is placed exactly on a row that still shows 8.5 Mbps DL traffic at MCS 20. Active high-rate traffic at the moment of detach and bearer release is inc
+- The RSRQ sequence is a value-for-value copy of the seed (-3.0 x16, -3.5, -9.0, -4.5, -3.5, -7.5, -9.5 ... -5.0, -3.0, -3.0). Only the timestamps (shifted, with 
+- The candidate is a near-verbatim copy of the seed, with only uniform offsets applied. Every timestamp is shifted by a constant +11 ms (+8 ms on row 1). Every RN
+- PUSCH SINR drops from ~32.5 to 18.5 dB and PUCCH SINR from ~29.6 to 15.3 dB in a single isolated report. Serving RSRP (-71), RSRQ (-11.0), PHR (40.0), TA and ne
+- Uplink SINR on PCI 2 drops by about 15 dB (to ~18.4 / 15.3), while PCI 1 stays at ~32.5 / 29.6. In the seed, closed-loop uplink power control holds both cells a
+
+### 5g-core-signalling
+
+Removal reasons: {'judge remove': 66, 'shifted-copy-of-seed': 12, 'row-keys': 9, 'label-leak': 5, 'time-order': 3, 'chunk-size': 3, 'generator-unparseable_json': 3, 'value-identifier': 2}
+Code-check rules: {'value-identifier': 20, 'row-keys': 12, 'shifted-copy-of-seed': 12, 'label-leak': 5, 'time-order': 3, 'chunk-size': 3, 'generator-unparseable_json': 3, 'value-time': 1}
+Major problems found by the judge (examples):
+- Near-verbatim copy of the seed. Every row is shifted by exactly +04:32:18.000 with identical millisecond offsets, and only the subscriber pseudonym is swapped (
+- A fresh Registration request starts 0.2 s after the UE-originated deregistration. The timeline has no ue_process_start or re-register event there. The warmup UE
+- The table covers only the warmup deregistration and a spurious re-registration. It never shows the AMF restart at 06:19:59.165 (no SCTP/NG Setup loss, no new NG
+- UE deregistration does not delete the subscriber's subscription data in the UDR. Open5GS does not do this; it would wipe the provisioned subscriber.
+- A UE-authentication PUT to the UDM during deregistration is not standard behaviour. Context cleanup would use nudm-uecm/sdm unsubscribe, not ueau.
+- After the NAS Authentication response the AMF must confirm 5G-AKA with PUT nausf-auth/ue-authentications, as the seed does. The candidate skips this step and go
+- Ping and HTTP workload start before any Registration accept (01:42:01.357) or PDU session establishment (01:42:01.563). User-plane traffic cannot run without a 
+- The UE is reported deregistered before the registration it would deregister from completes, and before the PDU session request at 01:42:01.563. The signalling t
+
+### ran-telemetry
+
+Removal reasons: {'value-copy-of-seed': 82, 'judge remove': 79, 'shifted-copy-of-seed': 71, 'label-leak': 18, 'row-keys': 9, 'generator-unparseable_json': 5, 'chunk-size': 2, 'value-time': 2}
+Code-check rules: {'value-copy-of-seed': 164, 'shifted-copy-of-seed': 78, 'label-leak': 20, 'value-time': 11, 'row-keys': 10, 'generator-unparseable_json': 5, 'chunk-size': 2, 'missing-table': 1}
+Major problems found by the judge (examples):
+- The timeline has the UL iperf running from 02:06:38 to about 02:06:48, followed by http and ping. No DL iperf is active until 02:06:55.402, yet these rows show 
+- DL holds at 59.2 Mbps while the TCP iperf target is 40M. The seed's 10M target produced about 12 Mbps, so a 40M TCP flow should not deliver roughly 50% above it
+- dl_bitrate is exactly 59.200 in all 20 rows, and dl_prbs alternates only between 90340 and 90330. Real seed rows vary from second to second (11.0-12.1 Mbps, 167
+- DL bitrate is 1.09 Mbps with 1687 PRBs and BSR 264, but the downlink iperf does not start until 07:06:33.254. The row reports traffic before the application tha
+- The timeline is shifted by about +1h01m00.47s from the seed, but the table is shifted by about +1h01m02s. The table therefore lags the ground truth by about 2 s
+- The row still shows 8.75 Mbps UL with BSR 196 bytes. The 5 s uplink iperf started at 10:34:19.403 and should have ended around 10:34:24.4, so this traffic falls
+- A per-UE scheduler row is reported after ue_process_stop at 10:34:32.243. The UE process is gone, so the gNB should not be reporting active UE scheduling at tha
+- The table is a near-verbatim copy of the seed with mechanical perturbations. pusch_snr_db, timing_advance_ns, dl_prbs (after row 6) and many bitrates are each s
+
+### network-performance
+
+Removal reasons: {'judge remove': 17, 'shifted-copy-of-seed': 17, 'time-order': 9, 'value-copy-of-seed': 7, 'row-keys': 3, 'value-type': 1, 'label-leak': 1, 'chunk-size': 1}
+Code-check rules: {'value-copy-of-seed': 19, 'shifted-copy-of-seed': 17, 'time-order': 10, 'row-keys': 5, 'value-type': 4, 'label-leak': 3, 'chunk-size': 2}
+Major problems found by the judge (examples):
+- The timeline starts a downlink iperf3 test (direction=downlink, seconds=10). The matching app_tests row reports a failed uplink test ('iperf3 uplink -'). The tw
+- The injected fault was changed from SMF to AMF, but the user-plane outage that follows was copied unchanged from the SMF case. An AMF restart in Open5GS does no
+- All app_tests rows are stamped 04:02:51–04:08:28, but the timeline runs 07:34:19–07:40:40. The results are offset by about 3.5 hours, so no test result lines up
+- The fault target was changed from SMF to AMF, but the outcome copied from the SMF seed was kept unchanged: permanent loss of the user plane and 'unable to conne
+- The result is logged 9.563 s after the request started, but total_ms=9775.0. The test would have finished about 0.21 s after its own result row, which is imposs
+- A PCF restart in Open5GS does not tear down an established PDU session or the UPF data path. The seed shows UL/DL iperf continuing normally after the PCF restar
+- These rows are labelled 'uplink -', but the timeline starts them as downlink tests (downlink tcp 5M, downlink udp 8M, downlink tcp 5M). Test direction is incons
+- There are two consecutive ue_process_stop/gnb_process_stop pairs with no process start between them, so the processes are stopped twice.
+
+### faults-ground-truth
+
+Removal reasons: {'shifted-copy-of-seed': 132, 'judge remove': 102, 'row-keys': 24, 'label-leak': 22, 'time-order': 10, 'chunk-size': 6, 'generator-unparseable_json': 4, 'value-time': 3}
+Code-check rules: {'shifted-copy-of-seed': 139, 'row-keys': 31, 'label-leak': 22, 'time-order': 21, 'value-time': 21, 'chunk-size': 6, 'generator-unparseable_json': 4, 'value-identifier': 2}
+Major problems found by the judge (examples):
+- Log rows are timestamped after ran_collect_stop (19:22:15.597). The timeline was compressed but the logs kept the seed's original spacing, so the logs no longer
+- This SM context error comes after the UE was deregistered (19:22:04.893) and after the UE and gNB processes stopped (19:22:08.741 and 19:22:10.598). No SM conte
+- These log events no longer line up with the scenario steps. AMF/AUSF endpoint teardown and the SmContext error happen after process stop instead of during the s
+- Near-verbatim copy of the seed shifted by about 6.5 h. Every log line, NF, sequence, address (127.0.0.11/12) and inter-event gap is reproduced, so it adds no ne
+- Reuses the seed's exact subscriber pseudonym supi=imsi-ue-4549df2477a8c074, which links the synthetic record directly to the measured subscriber.
+- PFCP recovery timestamps were copied verbatim from the seed. 4000953940 in NTP seconds is about 2026-10-14 08:05:40 UTC. That is roughly 5.6 h after this row's 
+- The candidate is the seed shifted by a constant offset of about -5:36:36.707. Every log text, gap, SM context ID (9), AMF_UE_NGAP_ID (10), NF instance UUID and 
+- The 'gmm_state_security_mode: Expectation failed' error appears with no cause. In Open5GS it follows a failed NGAP send to a removed gNB. The seed shows that ch
+
+### aaa-radius
+
+Removal reasons: {'shifted-copy-of-seed': 5, 'judge remove': 4, 'value-copy-of-seed': 2, 'value-identifier': 1, 'chunk-size': 1, 'row-keys': 1}
+Code-check rules: {'value-copy-of-seed': 9, 'shifted-copy-of-seed': 7, 'row-keys': 3, 'value-identifier': 1, 'chunk-size': 1}
+Major problems found by the judge (examples):
+- This user is rejected at 12:54:38 with 'No Auth-Type found', the unknown-user rejection. The same user then gets Login OK at 12:54:42 and 12:54:49. A user name 
+- Timeline events are stamped 17:54:37.000, but the log events they describe are at 12:54:37.000. Both claim the same UTC time base, so there is a 5-hour mismatch
+- The timeline says user-71f3b8d29c4e5a06 was accepted on port 30. In the log, that user appears only once, on port 9 at request (8). Port 30, request (29), is us
+- Near-verbatim copy of the seed. Every row matches: the same sequence of Accept and Reject per slot, the same nas_port numbering 1-20, identical millisecond offs
+- Two different users, user-3ba71c9f62d0a48e and user-6ec50a8b1437f9d2, both get 'Welcome otel-u07'. In the seed, each user pseudonym maps to exactly one otel acc
+- session_time_s=0 but 1.318 s have elapsed since the Start at 05:11:24.461. The other sessions, and the seed, report whole elapsed seconds rounded down, which wo
+- session_time_s=1 at 2.642 s and 2 at 3.967 s are one second too low. In the same document, user-3a7c1f0b9d4e28a6 reports 3 at 3.996 s and user-8f1a63d5c9b47e20 

deliveries/phaseC_qa_rows_2026-10-24/README.md

@@ -0,0 +1,56 @@
+# OTel 2.5 - Q&A training rows (UT Dallas), Phase C
+
+5373 question-and-answer rows in the `OTel_data_label.md` format, built from the UT Dallas real documents and the
+Phase B validated synthetic packets, for the seven OTel categories (36 MB). Human spot check: in
+progress (UT Dallas team).
+
+| category | rows from real data | rows from synthetic data | total |
+|---|---:|---:|---:|
+| lte-signalling | 331 | 546 | 877 |
+| rf-signal-quality | 241 | 357 | 598 |
+| 5g-core-signalling | 531 | 599 | 1130 |
+| ran-telemetry | 621 | 480 | 1101 |
+| network-performance | 186 | 296 | 482 |
+| faults-ground-truth | 323 | 654 | 977 |
+| aaa-radius | 100 | 108 | 208 |
+
+## Row format (data/<category>.jsonl, one JSON object per line)
+
+Fields of `OTel_data_label.md`: `source_repo`, `source_id` (the chunk the row is built from), `source_spec_id`,
+`source_version` (null), `messages` = [user: the document followed by the question, assistant: the answer].
+Additional fields: `run_id`, `split` (train / validation / test, by run, identical across all rows of a run),
+`question_type`, `document_variant` (`labelled`, or `stripped` = no scenario label or injected-fault timeline, used for
+detection and root-cause questions), `provenance` (the document's provenance; `synthetic` for rows on synthetic packets),
+`references` (standards the explanation relies on), `n_tokens` (row size measured with the generator's tokenizer;
+every row <= 3,900, under the 4,096-token limit). The extra fields can be dropped without affecting the spec fields.
+
+## How the rows were made
+
+- **Sources.** Real: per run and category up to two chunks, 3 + 2 rows (at most 5 rows per run and category). Synthetic:
+  3 rows per validated synthetic packet (each is its own run, always in the train split).
+- **Question mix** (feedback 3.2): extraction, procedure, cause and effect, comparison, anomaly detection, concept and
+  unanswerable questions, drawn per document; anomaly and most cause-and-effect questions are asked on the
+  label-stripped document (3.3). Personas vary (NOC, RF, core network, RAN, security, operations). Distribution:
+  `reports/distribution.md`.
+- **Answers** (feedback 3.1) have two parts: "What the records show" (only facts from the document) and "What it means"
+  (standard 3GPP / IETF / software knowledge, no speculation about the run).
+- **Generator:** `openai/gpt-6.1-sol` (reasoning low). **Judge:** `anthropic/claude-opus-5.5`, a different model family
+  (feedback 2.1). Prompts in `method/`.
+- **Checks per row:** code checks first (two-part answer; every timestamp and identifier in the answer present in the
+  document; every decimal number in the observed part present in the document or a difference of two document
+  times; no label leak on stripped documents; size), then the judge (grounding, domain correctness, question quality);
+  rejected rows were rewritten once with the reasons and checked again. Rows on synthetic packets whose answer reported
+  an internal contradiction were dropped; those 45 packets (`reports/synthetic_docs_flagged.json`) were then
+  removed from the Phase B delivery, together with the 82 other rows built on them. 6 near-duplicate questions within a run removed.
+- **Judge calibration** (`reports/judge_calibration.json`): planted errors in kept rows (wrong count, wrong network
+  function, negated explanation) removed 18/19; unmodified rows kept 12/12.
+- **Result:** 5,577 rows planned; 5373 delivered. Run statistics: `reports/run_report.md`.
+
+## Known limits
+
+- The judge is an LLM; a human spot check by the UT Dallas team is in progress.
+- Synthetic rows rest on synthetic packets: they are labelled as such and are in the train split only.
+- Dense numeric documents (RAN telemetry) were split into two sub-chunks (`source_id` ending in s1 / s2) where the
+  row would otherwise exceed the token limit.
+- v1 ZMQ and early over-the-air runs had no internet access from the UE, so rows about internet tests in those runs
+  describe that failure; core-gateway tests are unaffected.

deliveries/phaseC_qa_rows_2026-10-24/data/5g-core-signalling.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:976854120a58144682a5e24ed01539304a8968496e03778d71035d0128c34158
+size 9500716

deliveries/phaseC_qa_rows_2026-10-24/data/aaa-radius.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fe3683d01cdb607f24868d9f7c6b88773c5cb82a1e230e45f0d6c50cff0aa15d
+size 1421748

deliveries/phaseC_qa_rows_2026-10-24/data/faults-ground-truth.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:274ee1f7cb752fec7d8e8a365af1ecd306f92b1599cc2100731d67bfed495767
+size 5507319

deliveries/phaseC_qa_rows_2026-10-24/data/lte-signalling.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9dff7c6d2f251697358036aaaf3e7d1a3d5c1fd5a485a6107bc8bc35d07db570
+size 6710159

deliveries/phaseC_qa_rows_2026-10-24/data/network-performance.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:abab60d3036cefbfbfcea18ed2137a884ffaf42d641a9ead69daf8d6d3cef0e0
+size 2481399

deliveries/phaseC_qa_rows_2026-10-24/data/ran-telemetry.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ba555a43beb1c6c925449e14e5eb600707c891048f6457284254012d7134014
+size 6918420

deliveries/phaseC_qa_rows_2026-10-24/data/rf-signal-quality.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2e51902e3ca1a588315341b484cb1d5013ea761b91770bb0ba5012bcf80346fb
+size 3339112

deliveries/phaseC_qa_rows_2026-10-24/method/qa_generator_prompt.txt

@@ -0,0 +1,32 @@
+You write question-and-answer training rows about telecom network records.
+
+You get one document (real or synthetic records, described in its header) and a list of rows to write, each with a
+question type and a persona. For each row write a question that persona would ask about THIS document, and the answer.
+
+Answer rules:
+1. Two parts, always in this form:
+   "What the records show: ..." - only facts from the document: values, times, identifiers, messages, outcomes, copied
+   exactly as written (keep the document's units and precision; counts and differences you compute must be correct).
+   "What it means: ..." - the interpretation with standard knowledge (3GPP, IETF, the software named in the header):
+   procedures, causes, network-function roles, normal ranges. No speculation about this run beyond the records; say so
+   when the records cannot establish something.
+2. Respect the header's provenance: simulated radio values are not over-the-air measurements; uncalibrated values are
+   relative only. Mention the provenance only where it changes the interpretation (e.g. radio values from a simulated
+   or uncalibrated link); do not open answers with it.
+3. Never assume a scenario or an injected fault the document does not state. For an "unanswerable" row, the answer
+   says what information is missing and does not guess.
+4. Questions are natural, specific and different from each other; do not paste large parts of the document into them.
+5. "references": the standards sections the "What it means" part relies on (e.g. "TS 24.501 5.5.1.2", "TS 23.502
+   4.2.2.2", "RFC 2865 4.3"), or [] if none.
+
+Answer with one JSON object only:
+{"rows": [{"n": 1, "question_type": "<as given>", "persona": "<as given>", "question": "...", "answer": "What the records show: ... What it means: ...", "references": ["..."]}]}
+
+--- question types ---
+extraction: a factual question about specific values, counts or messages in the document
+procedure: a question about a procedure or sequence in the records: which step it is, what comes before or after
+cause_effect: a cause-and-effect or root-cause question: why something in the records happened, based on the evidence
+comparison: a comparison across times, conditions, users, cells or phases in the document
+anomaly_detection: ask whether anything in the records is abnormal, without naming what (the reader must find it)
+concept: a concept question grounded in the records: the role or meaning of an element, NF, message or field seen here
+unanswerable: a plausible question the document does NOT contain enough information to answer; the answer says what is missing and does not guess

deliveries/phaseC_qa_rows_2026-10-24/method/qa_judge_prompt.txt

@@ -0,0 +1,15 @@
+You review question-and-answer training rows written about one telecom document (real or synthetic records,
+described in its header). Rows that pass will train a telecom model, so a wrong fact or a wrong explanation must be removed.
+
+For each row decide:
+- grounded: every observed fact in "What the records show" (value, time, identifier, message, outcome, count or
+  difference) is present in or correctly computed from the document;
+- domain_correct: "What it means" is correct standard 3GPP / IETF / software behaviour, does not overclaim, and respects
+  the header's provenance (simulated or uncalibrated values, synthetic documents);
+- question_ok: the question is natural, answerable from the document (or, for type "unanswerable", the answer correctly
+  says what is missing without guessing) and matches its question type; a question on a label-stripped document does
+  not reveal the scenario.
+List concrete problems with severity "major" (must not be learned) or "minor". Verdict "keep" only with no major problem.
+Answer with one JSON object only:
+{"rows": [{"n": 1, "verdict": "keep" | "remove", "grounded": true, "domain_correct": true, "question_ok": true,
+           "problems": [{"severity": "major" | "minor", "issue": "..."}]}]}

deliveries/phaseC_qa_rows_2026-10-24/reports/distribution.md

@@ -0,0 +1,20 @@
+# Q&A row distribution
+
+| category | origin | rows | extraction | procedure | comparison | concept | cause_effect | anomaly_detection | unanswerable | stripped docs | train | validation | test |
+|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| 5g-core-signalling | real | 531 | 136 | 87 | 59 | 45 | 113 | 62 | 29 | 173 | 434 | 39 | 58 |
+| 5g-core-signalling | synthetic | 599 | 159 | 140 | 61 | 57 | 116 | 37 | 29 | 153 | 599 | 0 | 0 |
+| aaa-radius | real | 100 | 25 | 21 | 5 | 11 | 20 | 10 | 8 | 36 | 85 | 10 | 5 |
+| aaa-radius | synthetic | 108 | 35 | 25 | 7 | 14 | 16 | 9 | 2 | 24 | 108 | 0 | 0 |
+| faults-ground-truth | real | 323 | 79 | 78 | 36 | 32 | 54 | 30 | 14 | 87 | 264 | 23 | 36 |
+| faults-ground-truth | synthetic | 654 | 163 | 117 | 50 | 73 | 128 | 89 | 34 | 227 | 654 | 0 | 0 |
+| lte-signalling | real | 331 | 92 | 66 | 37 | 28 | 61 | 29 | 18 | 93 | 298 | 21 | 12 |
+| lte-signalling | synthetic | 546 | 147 | 127 | 42 | 61 | 91 | 53 | 25 | 139 | 546 | 0 | 0 |
+| network-performance | real | 186 | 51 | 37 | 18 | 23 | 30 | 19 | 8 | 50 | 138 | 18 | 30 |
+| network-performance | synthetic | 296 | 68 | 52 | 32 | 26 | 71 | 32 | 15 | 103 | 296 | 0 | 0 |
+| ran-telemetry | real | 621 | 157 | 91 | 49 | 56 | 159 | 76 | 33 | 239 | 518 | 48 | 55 |
+| ran-telemetry | synthetic | 480 | 128 | 108 | 49 | 34 | 93 | 45 | 23 | 142 | 480 | 0 | 0 |
+| rf-signal-quality | real | 241 | 65 | 57 | 22 | 20 | 50 | 18 | 9 | 56 | 219 | 16 | 6 |
+| rf-signal-quality | synthetic | 357 | 86 | 64 | 29 | 39 | 81 | 34 | 24 | 125 | 357 | 0 | 0 |
+
+Overall question mix (AT&T's suggested mix in brackets): extraction 26% [25%], procedure 20% [20%], cause_effect 20% [20%], comparison 9% [10%], anomaly_detection 10% [10%], concept 10% [10%], unanswerable 5% [5%]

deliveries/phaseC_qa_rows_2026-10-24/reports/judge_calibration.json

@@ -0,0 +1,22 @@
+{
+ "unmodified": {
+  "n": 12,
+  "correct": 12,
+  "rate": 1.0
+ },
+ "count": {
+  "n": 7,
+  "correct": 7,
+  "rate": 1.0
+ },
+ "nf": {
+  "n": 1,
+  "correct": 1,
+  "rate": 1.0
+ },
+ "negate": {
+  "n": 11,
+  "correct": 10,
+  "rate": 0.91
+ }
+}

deliveries/phaseC_qa_rows_2026-10-24/reports/judge_calibration_phaseB_documents.json

@@ -0,0 +1,29 @@
+{
+ "unmodified": {
+  "n": 24,
+  "correct": 24,
+  "rate": 1.0
+ },
+ "implausible-scale": {
+  "n": 3,
+  "correct": 2,
+  "rate": 0.667
+ },
+ "swap-order": {
+  "n": 9,
+  "correct": 8,
+  "rate": 0.889
+ },
+ "identity-switch": {
+  "n": 3,
+  "correct": 3,
+  "rate": 1.0
+ },
+ "timeline-contradiction": {
+  "n": 5,
+  "correct": 5,
+  "rate": 1.0
+ },
+ "planted_catch_rate": 0.9,
+ "false_remove_rate_real": 0.0
+}

deliveries/phaseC_qa_rows_2026-10-24/reports/run_report.md

@@ -0,0 +1,44 @@
+# Phase C Q&A rows: run report
+
+Counts are from the generation run, before near-duplicate questions and the rows on the 45 removed synthetic packets were taken out; the delivered count is in the README.
+
+units 2070, kept rows 5461, spec-check errors 0
+
+| category | origin | kept | rejected (code) | rejected (judge) | dropped (too long) | dropped (synthetic conflict) |
+|---|---|---:|---:|---:|---:|---:|
+| 5g-core-signalling | real | 531 | 1 | 9 | 6 | 0 |
+| 5g-core-signalling | synthetic | 651 | 2 | 3 | 0 | 29 |
+| aaa-radius | real | 100 | 0 | 1 | 0 | 0 |
+| aaa-radius | synthetic | 114 | 0 | 1 | 0 | 9 |
+| faults-ground-truth | real | 324 | 0 | 10 | 0 | 0 |
+| faults-ground-truth | synthetic | 656 | 0 | 8 | 0 | 1 |
+| lte-signalling | real | 331 | 0 | 3 | 0 | 0 |
+| lte-signalling | synthetic | 551 | 0 | 8 | 1 | 3 |
+| network-performance | real | 186 | 7 | 6 | 0 | 0 |
+| network-performance | synthetic | 302 | 12 | 7 | 0 | 3 |
+| ran-telemetry | real | 624 | 18 | 10 | 127 | 0 |
+| ran-telemetry | synthetic | 487 | 6 | 5 | 0 | 5 |
+| rf-signal-quality | real | 243 | 76 | 24 | 4 | 0 |
+| rf-signal-quality | synthetic | 361 | 75 | 15 | 3 | 2 |
+
+## Question-type mix of kept rows
+
+- 5g-core-signalling: extraction 299 (25%), cause_effect 246 (21%), procedure 229 (19%), comparison 124 (10%), anomaly_detection 112 (9%), concept 105 (9%), unanswerable 67 (6%)
+- aaa-radius: extraction 60 (28%), procedure 52 (24%), cause_effect 36 (17%), concept 25 (12%), anomaly_detection 19 (9%), comparison 12 (6%), unanswerable 10 (5%)
+- faults-ground-truth: extraction 244 (25%), procedure 195 (20%), cause_effect 182 (19%), anomaly_detection 119 (12%), concept 105 (11%), comparison 87 (9%), unanswerable 48 (5%)
+- lte-signalling: extraction 239 (27%), procedure 194 (22%), cause_effect 152 (17%), concept 89 (10%), comparison 82 (9%), anomaly_detection 82 (9%), unanswerable 44 (5%)
+- network-performance: extraction 122 (25%), cause_effect 103 (21%), procedure 90 (18%), anomaly_detection 51 (10%), comparison 50 (10%), concept 49 (10%), unanswerable 23 (5%)
+- ran-telemetry: extraction 287 (26%), cause_effect 253 (23%), procedure 201 (18%), anomaly_detection 123 (11%), comparison 98 (9%), concept 92 (8%), unanswerable 57 (5%)
+- rf-signal-quality: extraction 151 (25%), cause_effect 131 (22%), procedure 123 (20%), concept 60 (10%), anomaly_detection 54 (9%), comparison 52 (9%), unanswerable 33 (5%)
+
+## Synthetic documents flagged by Q&A answers as internally inconsistent: 45 (see synthetic_docs_flagged.json)
+
+
+## Final rejection reasons (after the repair attempt)
+
+- 5g-core-signalling: judge x3; The answer first says there are 5 downlink app_iperf_start e x1
+- faults-ground-truth: judge x6
+- lte-signalling: judge x4
+- network-performance: judge x7
+- ran-telemetry: judge x8; label-leak x2
+- rf-signal-quality: judge x28; This document is a label-stripped variant. The question name x1

deliveries/phaseC_qa_rows_2026-10-24/reports/synthetic_docs_flagged.json

@@ -0,0 +1,47 @@
+[
+ "local-radius-synthetic__syn2-12214cccf2c6",
+ "local-radius-synthetic__syn2-2a60fac37fce",
+ "local-radius-synthetic__syn2-a977f8d97c9f",
+ "local-radius-synthetic__syn2-d00fd18bc2a3",
+ "local-radius-synthetic__syn2-ff64480fe470",
+ "ota-5g-b210-synthetic__syn2-6a9822fd171c",
+ "ota-5g-b210-synthetic__syn2-a30aeed13452",
+ "ota-5g-b210-synthetic__syn2-b2bb1f2b3584",
+ "ota-5g-b210-synthetic__syn2-df1663338209",
+ "ota-5g-b210-synthetic__syn2-e2c3f76392f3",
+ "ota-5g-b210-synthetic__syn2-e4f44fd16691",
+ "ota-5g-b210-synthetic__syn2-e91dc26134ec",
+ "ota-5g-b210-synthetic__syn2-fb03cad0d5fc",
+ "ota-5g-b210-synthetic__syn2-fc54fc53846e",
+ "ota-5g-b210-synthetic__syn2-fd0acf537bfd",
+ "ota-lte-b210-synthetic__syn2-1348ae38bfdd",
+ "ota-lte-b210-synthetic__syn2-618417d4d6af-r1",
+ "ota-lte-b210-synthetic__syn2-702113c71443",
+ "powder-lte-v1-synthetic__syn2-3ca633672e30",
+ "powder-lte-v1-synthetic__syn2-78b9b3e1ac69",
+ "powder-lte-v1-synthetic__syn2-7aa49b0d046f",
+ "powder-lte-v1-synthetic__syn2-ebe42a986b7d",
+ "zmq-5g-core-sim-v1-synthetic__syn2-06cbbc874b35",
+ "zmq-5g-core-sim-v1-synthetic__syn2-1452d20fdad5",
+ "zmq-5g-core-sim-v1-synthetic__syn2-3652d33a0bd2",
+ "zmq-5g-core-sim-v1-synthetic__syn2-373e8201ae51",
+ "zmq-5g-core-sim-v1-synthetic__syn2-46410d599550-r1",
+ "zmq-5g-core-sim-v1-synthetic__syn2-58e4b4761add",
+ "zmq-5g-core-sim-v1-synthetic__syn2-5a703eaeae21",
+ "zmq-5g-core-sim-v1-synthetic__syn2-6fc237f3b3ae",
+ "zmq-5g-core-sim-v1-synthetic__syn2-86482b6ab4b6",
+ "zmq-5g-core-sim-v1-synthetic__syn2-93a2841cd651",
+ "zmq-5g-core-sim-v1-synthetic__syn2-965fbca4f96e",
+ "zmq-5g-core-sim-v1-synthetic__syn2-9aa641dda4d7",
+ "zmq-5g-core-sim-v1-synthetic__syn2-9ffa3f617069-fx",
+ "zmq-5g-core-sim-v1-synthetic__syn2-b5cd622a78ac",
+ "zmq-5g-core-sim-v1-synthetic__syn2-bb4430024200-r1",
+ "zmq-5g-core-sim-v1-synthetic__syn2-bc396ebeb1eb",
+ "zmq-5g-core-sim-v1-synthetic__syn2-be27d2bb0301",
+ "zmq-5g-core-sim-v1-synthetic__syn2-c765ead73bb7",
+ "zmq-5g-core-sim-v1-synthetic__syn2-ceaf2ee94876",
+ "zmq-5g-core-sim-v1-synthetic__syn2-e3582eb2e55e",
+ "zmq-5g-core-sim-v1-synthetic__syn2-e600847724b6",
+ "zmq-5g-core-sim-v1-synthetic__syn2-f369e603c3bd",
+ "zmq-5g-core-sim-v1-synthetic__syn2-fbb19cae0dfa"
+]