Strange LabSign in

excytin_ash_train_benign_ah_byhost

1,450,520events in the training captures
3models serving questions
2024 → 2024span of the captures
12results
cyber_defense · excytin_ash_train_benign_ah_byhost · World model · public
01 · Data

Dataset versions

A dataset version is a capture of the source as of a date: its event spine and the files training reads.

1 dataset version

excytin_ash_train_benign_ah_byhost excytin_ash_train_benign_ah_byhost @ asof-2026-08-11

Recorded where it is (vei-gdrive; spine also on Modal rezendi vei-oss-data); see REPORT.md

llm_report · REPORT.md

excytin_ash_train_benign_ah_byhost@asof-2026-08-11 — data stage report

Recorded 2026-10-13 as part of the line's migration to the factory. The data already existed and nothing was rematerialized.

Inputs, by identity. ExCyTIn-Bench, tenant alphineskihouse, the advanced-hunting device world. This stream is derived from excytin_alphineskihouse_ah (derivation.json):

  • 3,038,988 source events over the benign training windows from

2024-06-20 to 2024-07-22;

  • grouped by host (24 hosts);
  • 342,701 incident-activity events stripped over 73 windows, padded by

1,800 s.

The raw benchmark and the parent streams are external: they stay on the archive drive and Google Drive.

Output, by identity.

  • Spine: canonical_events.jsonl, 1,450,520 benign events.
  • Resource: behaviour_profile.json.
  • Receipts: derivation.json, lag_grid.json, and the ingest-time

drift check (drift.json, drift.txt).

Decisions. Recorded where it is: at its Google Drive copy (vei-gdrive:ingest-data/cyber_defense/), which holds every file. The spine is also on the rezendi workspace's vei-oss-data volume, where the v10 runs read it. Nothing was uploaded except this report and the drift receipts.

Drift. Ingest-time: 2 monthly periods, no verb or volume flags. The check has little to compare over two periods. The retained models train on the whole benign stream (no test side) and score separate incident windows, so a split-time check inside this stream does not apply.

For the next stage. Atlas device_v1. The retained runs device_v9, device_v10 and device_aux_v10 read this stream.

Receipts (2)
drift.json
{
  "era": {
    "n_periods": 2,
    "n_events": 1450520,
    "structural_types": [
      "dvc_misc:antivirusdetection",
      "dvc_misc:appcontrolpolicyapplied",
      "dvc_misc:asrlsasscredentialtheftblocked",
      "dvc_misc:controlledfolderaccessviolationblocked",
      "dvc_misc:directoryserviceobjectcreated",
      "dvc_misc:firewalloutboundconnectionblocked",
      "dvc_misc:ntmapviewofsectionremoteapicall",
      "dvc_misc:sensitivefileread",
      "dvc_misc:useraccountmodified",
      "dvc_misc:writetolsassprocessmemory"
    ],
    "verb_flags": [],
    "volume_flags": [],
    "trailing_periods": 6,
    "spike_ratio": 4.0
  },
  "split": null
}
drift.txt
[drift] 1,450,520 events over 2 periods | structural types excluded: dvc_misc:antivirusdetection, dvc_misc:appcontrolpolicyapplied, dvc_misc:asrlsasscredentialtheftblocked, dvc_misc:controlledfolderaccessviolationblocked, dvc_misc:directoryserviceobjectcreated, dvc_misc:firewalloutboundconnectionblocked, dvc_misc:ntmapviewofsectionremoteapicall, dvc_misc:sensitivefileread, dvc_misc:useraccountmodified, dvc_misc:writetolsassprocessmemory
[drift] no verb moved >=4x its own trailing median — clean
7 files, 1.74 GB
  • Event spine · 1 file1.71 GB
  • canonical_events.jsonl36b9e45d0cb41.71 GB
  • Resource · 1 file22.2 MB
  • behaviour_profile.json74095d5c278f22.2 MB
  • Receipt · 4 files9 kB
  • derivation.jsonf65c3f15b7361 kB
  • drift.jsoneb43f4304441661 B
  • drift.txtd964c2cbfa99502 B
  • lag_grid.json6c0129152b566 kB
  • Phase_report · 1 file2 kB
  • REPORT.md759305bf6c3b2 kB
Events
1,450,520 in 24 streams
Span
20 Jun 2024 → 22 Jul 2024
Source
ExCyTIn-Bench alphineskihouse, advanced-hunting device world, benign windows grouped by host; derived from excytin_alphineskihouse_ah, as of 11 Aug 2026
Published
13 Oct 2026
02 · Features

Featurization

An Atlas recipe declares how events become model inputs.

1 feature recipe

device_v1 pipeline/featurize/cyber_defense/recipes/device_v1.yaml

Used by
device_v9
1 runtrained over it
03 · Training

Recipes and runs

A Praxis recipe fixes the model, the objective and the split.

3 training recipes
3 runs

device_v9 excytin_ash_train_benign_ah_byhost__transformer__drivers_ah1__ctx16__ash_v9r_heldout_ah1_lag300__20260820-005405

20 files, 1.65 GB
  • Latent matrices · 1 file1.49 GB
  • latents.npy1.49 GB
  • Sample and event ids · 1 file103.0 MB
  • latent_event_ids.json103.0 MB
  • Metrics · 1 file27.1 MB
  • metrics.json27.1 MB
  • Checkpoints · 2 files21.6 MB
  • model.pt5.4 MB
  • train_checkpoint.pt16.2 MB
  • Auxiliary · 8 files8.4 MB
  • ARCHIVED.md926 B
  • metrics.jsonl288 B
  • scores_l1_ho.parquet2.6 MB
  • scores_n1_std_ho.parquet2.3 MB
  • scores_nl300_std_ho.parquet1.9 MB
  • scores_nl3600_std_ho.parquet1.6 MB
  • tiers_inc134_heldout.json21 kB
  • tiers_inc134.json16 kB
  • Configuration · 5 files503 kB
  • asof_config.json190 B
  • config.json414 B
  • praxis_import.json213 B
  • praxis_paths.json509 B
  • praxis.json502 kB
  • Evaluation outputs · 1 file14 kB
  • gnomon/tiers_inc134_v1/results.jsond587957283bc14 kB
  • Split · 1 file869 B
  • split.json869 B
Status
imported
Trained on
excytin_ash_train_benign_ah_byhost @ asof-2026-08-11
Origin
manifest inferred from the directory
Published
13 Oct 2026
1.65 GB20 artifacts

ash_v10_auxnexttype_ah1_lag300 ash_v10_auxnexttype_ah1_lag300

22 files, 1.65 GB
  • Auxiliary · 16 files1.61 GB
  • asof_config.json190 B
  • config.json414 B
  • latent_event_ids.json103.0 MB
  • latents.npy1.49 GB
  • metrics.jsonl545 B
  • readout_heads.pt45 kB
  • scores_l1_ho.parquet2.6 MB
  • scores_n1_std_ho.parquet2.3 MB
  • scores_nl300_std_ho.parquet1.9 MB
  • scores_nl3600_std_ho.parquet1.6 MB
  • scores_provenance.json716 B
  • scores_vtl300_ho.parquet1.6 MB
  • split.json869 B
  • tiers_inc134_heads.json5 kB
  • tiers_inc134_heldout.json21 kB
  • train_checkpoint.pt16.3 MB
  • Metrics · 1 file27.1 MB
  • metrics.json27.1 MB
  • Checkpoints · 1 file5.4 MB
  • model.pt5.4 MB
  • Configuration · 3 files502 kB
  • praxis_import.json218 B
  • praxis_paths.json299 B
  • praxis.json502 kB
  • Evaluation outputs · 1 file13 kB
  • gnomon/tiers_inc134_v1/results.jsondc616a575e9713 kB
Status
imported
Trained on
excytin_ash_train_benign_ah_byhost @ asof-2026-08-11
Origin
manifest inferred from the directory
Published
13 Oct 2026
1.65 GB22 artifacts

ash_v10_noaux_ah1_lag300 ash_v10_noaux_ah1_lag300

21 files, 1.65 GB
  • Latent matrices · 1 file1.49 GB
  • latents.npy1.49 GB
  • Sample and event ids · 1 file103.0 MB
  • latent_event_ids.json103.0 MB
  • Metrics · 1 file27.1 MB
  • metrics.json27.1 MB
  • Checkpoints · 2 files21.6 MB
  • model.pt5.4 MB
  • train_checkpoint.pt16.2 MB
  • Auxiliary · 8 files8.4 MB
  • metrics.jsonl545 B
  • remote.json147 B
  • scores_l1_ho.parquet2.6 MB
  • scores_n1_std_ho.parquet2.3 MB
  • scores_nl300_std_ho.parquet1.9 MB
  • scores_nl3600_std_ho.parquet1.6 MB
  • scores_provenance.json577 B
  • tiers_inc134_heldout.json22 kB
  • Configuration · 6 files513 kB
  • asof_config.json190 B
  • config.json414 B
  • praxis_import.json214 B
  • praxis_paths.json287 B
  • praxis.json502 kB
  • run_manifest.json10 kB
  • Evaluation outputs · 1 file13 kB
  • gnomon/tiers_inc134_v1/results.jsona3631dcb6c0813 kB
  • Split · 1 file869 B
  • split.json869 B
Status
imported
Trained on
excytin_ash_train_benign_ah_byhost @ asof-2026-08-11
Origin
not recorded
Published
13 Oct 2026
1.65 GB21 artifacts
04 · Evaluation

How it compares to other methods

uses the world modelother methodcoin flip

AUC is the chance a method ranks a case that happened above one that didn't: 0.5 is a coin flip, 1.0 is perfect.

The models of record and their headline evaluations. Every result includes the ablations and earlier variants.

Trained only on a tenant's normal device activity, does the model's surprise single out the attack when it comes? Higher means the attack's events or time windows rank above the normal ones.

Events that belong to the attack

World model, 1 step0.905
World model, 5 min0.790
World model, 1 hour0.600
Linear control0.665
0.500.600.700.800.90

Time windows that contain the attack

World model, 1 step0.741
World model, 5 min0.708
World model, 1 hour0.679
Linear control0.784
0.500.600.700.800.90
How it is graded: 1 eval config

tiers_inc134_heldout eval config · revision 4 · pipeline/infer/cyber_defense/recipes/migrated/tiers_inc134_heldout.yaml

Targets
incident 134, event tier
incident 134, window tier
Arms
L1_ho, N1_std_ho, NL300_std_ho, NL3600_std_ho
12 cellsover 1 run
05 · Serving

Ask the model

If the query service is cold, it will need a few minutes to warm up.

Serving

idle

Published, not yet served.

This line has no query surface and no published report yet, so its models are published but answers nothing here.