Sati models

Nine models that Sati, the investigation agent, runs on — each named for one of the nine forms of Navadurga and given the job that fits it. Every model is built here and trained and tested on real public data, in the cloud.

Models

one per stage of the loop

Trained

9 of 9

each on real public data

Katyayani, unseen systems

88.4%

top-1 root cause · baseline 57.0%

Real incidents tested

735

RCAEval · pinned afeacb11

Where each model sits in the loop

Observe
Remember
- Brahmacharini
Understand
- Katyayani
Explain
- Kushmanda
Act
- Skandamata
Verify
- Mahagauri
Decide
- Siddhidatri

1Observetrained
Metric anomaly detection — learns what normal looks like for every series
Shailaputri — the steady foundation
NABSMDRP-60
Results ↓Try it on a held-out server →
On Kaggle:dataset ↗notebook ↗model ↗
2Remembertrained
Embeddings for memory — maps observed behaviour to the known technique it resembles
Brahmacharini — discipline and study
MITRE ATT&CK proceduresRP-64
Results ↓Ask it in Sati AI →
On Kaggle:dataset ↗notebook ↗model ↗
3Observetrained
Network intrusion alerting — flags attack traffic by family
Chandraghanta — the bell that warns
UNSW-NB15CIC-IDS2017RP-61
Results ↓Try it on a held-out flow →
On Kaggle:dataset ↗notebook ↗model ↗
4Explaintrained
Writes the structured hypothesis and explanation — Sati's own language model
Kushmanda — the creator
MITRE ATT&CK, CAPEC, CWECISA KEVNVD CVE descriptionsRP-65
Results ↓Not served publicly yet
5Acttrained
Proposes the protective response — Dharma still authorises it
Skandamata — the protecting mother
MITRE ATT&CK mitigationsRP-72
Results ↓Ask it in Sati AI →
On Kaggle:dataset ↗notebook ↗model ↗
6Understandtrained
Root-cause ranking — names the service at fault and the fault type, citing the metric series
Katyayani — the warrior who finds the adversary
RCAEvalRP-58
Top-1 root cause on systems never seen in training: 88.4% vs 57.0% without learning.
Results ↓Try it on a held-out incident →
On Kaggle:dataset ↗notebook ↗model ↗
7Observetrained
Log anomaly detection — what hides in the noise
Kalaratri — the night that dispels darkness
LogHub HDFSLogHub BGLRP-62
Results ↓Try it on a held-out session →
On Kaggle:dataset ↗notebook ↗model ↗
8Verifytrained
Verification — confirms a system has truly returned to normal after an action
Mahagauri — calm restored
SMD incidentsRP-71
Results ↓Try it on a held-out server →
On Kaggle:dataset ↗notebook ↗model ↗
9Decidetrained
Fusion — combines several detectors into one calibrated verdict
Siddhidatri — completion
SMD: Shailaputri + baselines, stacked out of foldRP-73
Results ↓Try it on a held-out server →
On Kaggle:dataset ↗notebook ↗model ↗

Shailaputri — anomalies in real metric series

Trained on 47 real series from NAB, each family of series held out in turn, and scored online by the same harness as the four statistical detectors on exactly the same series. F1 is read together with false alarms per 1000 points, never alone.

F1 on a family it never saw

vs the best statistical detector

trained modelno-learning baselinewhiskers: 95% interval

realAWSCloudwatchfamily never seen in training · best statistical: seasonal-naiven = 17
40.4%
33.4%
realAdExchangefamily never seen in training · best statistical: seasonal-naiven = 6
35.6%
29.5%
realKnownCausefamily never seen in training · best statistical: seasonal-naiven = 7
44.0%
30.3%
realTrafficfamily never seen in training · best statistical: seasonal-naiven = 7
49.6%
54.1%
realTweetsfamily never seen in training · best statistical: seasonal-naiven = 10
42.5%
25.9%

Episode recall and point precision combined; the threshold was chosen on training families only.

Mean across held-out families

Detector	F1	Recall	False alarms / 1000	Lead
Shailaputri	0.424	0.739	24.1	179.3
robust-z	0.325	0.911	78.7	234.1
ewma	0.261	0.900	14.2	229.0
Seasonal naive	0.346	0.930	74.7	240.0
Seasonal robust z-score	0.312	0.761	74.7	241.3

NAB’s labels are windows, and Wu & Keogh (2021) show several NAB series are trivial or mislabelled; these numbers inherit those limits.

Brahmacharini — from behaviour to the technique it is

Each test query is a real description of how an attacker behaved, written about one of 115 groups, malware families or tools never seen in training. The model ranks all 706 ATT&CK techniques; the same queries go to BM25 and to the embedder Smriti uses today.

Held-out behaviour descriptions

n = 1727

Retriever	Recall@1	Recall@5	Recall@10	MRR
Brahmacharini	0.614	0.802	0.854	0.699
BM25	0.268	0.478	0.580	0.372
TF-IDF + logistic regression	0.646	0.832	0.869	0.730
Brahmacharini + TF-IDF classifier, fused	0.673	0.859	0.896	0.758
Hashing embedder (Smriti's previous retriever)	0.120	0.226	0.283	0.177
Brahmacharini, as served	0.556	0.735	0.797	0.638
Brahmacharini as served + TF-IDF classifier, fused	0.650	0.847	0.887	0.738

A plain TF-IDF classifier beats Brahmacharini on its own (Recall@1 0.646 against 0.614). Combined, with the mix chosen on the validation sources and never on the test, the two reach 0.673, better than either. The classifier can only name techniques it saw in training and cannot retrieve passages, so Sati AI still grounds its answers with Brahmacharini.

Chandraghanta — attack traffic in real labelled flows

Two lab datasets of labelled network flows. On CIC-IDS2017 the model trains on Monday to Wednesday and is tested on Thursday and Friday, so every attack family it is tested on is one it has never seen — and the random-split row shows what the usual evaluation would have claimed instead.

Try it

runs live in the API · research use only

Loading the examples…

UNSW-NB15: 82,332 held-out flows

55.1% attacks · official training and testing files

Detector	Precision	Recall	F1	False-positive rate
Chandraghanta, best-F1 threshold	0.821	0.988	0.897	0.2634
Chandraghanta, best-F1 threshold, without TTL fields	0.827	0.981	0.897	0.2522
Chandraghanta	0.952	0.926	0.939	0.0578
PCA, trained on benign flows	0.581	0.993	0.733	0.8782

Attack family	Flows	Seen in training	Chandraghanta, best-F1 threshold recall	Chandraghanta, best-F1 threshold, without TTL fields recall	Chandraghanta recall	pca (benign-only) recall
Analysis	677	yes	0.975	0.972	0.916	0.994
Backdoor	583	yes	1.000	1.000	0.990	0.997
DoS	4,089	yes	0.999	0.997	0.987	0.988
Exploits	11,132	yes	0.996	0.992	0.973	0.995
Fuzzers	6,062	yes	0.919	0.881	0.531	0.979
Generic	18,871	yes	1.000	1.000	0.999	1.000
Reconnaissance	3,496	yes	0.999	0.997	0.990	0.984
Shellcode	378	yes	0.992	0.992	0.921	0.989
Worms	44	yes	1.000	1.000	0.977	1.000

Research use only, not for commercial use (CC BY-NC-SA 4.0): trained on research-use-only data; owner decision, 27 Oct 2026. UNSW-NB15 and CIC-IDS2017 are research-use only.

CIC-IDS2017: 1,162,213 held-out flows

25.1% attacks · by day: trained on Monday to Wednesday, tested on Thursday and Friday

Detector	Precision	Recall	F1	False-positive rate
Chandraghanta, best-F1 threshold	0.989	0.294	0.453	0.0011
Chandraghanta	0.927	0.448	0.604	0.0118
PCA, trained on benign flows	0.440	0.135	0.207	0.0575
random split (reference only)	0.998	0.996	0.997	0.0005

Attack family	Flows	Seen in training	Chandraghanta, best-F1 threshold recall	Chandraghanta recall	pca (benign-only) recall
Bot	1,966	no	0.000	0.000	0.028
DDoS	128,027	no	0.636	0.975	0.215
Infiltration	36	no	0.000	0.000	0.722
PortScan	158,930	no	0.013	0.020	0.074
Web Attack - Brute Force	1,507	no	0.908	1.000	0.000
Web Attack - Sql Injection	21	no	1.000	1.000	0.000
Web Attack - XSS	652	no	0.974	1.000	0.000

Research use only, not for commercial use (CC BY-NC-SA 4.0): trained on research-use-only data; owner decision, 27 Oct 2026. UNSW-NB15 and CIC-IDS2017 are research-use only.

Kushmanda — Sati's own language model

16.9M parameters, trained from scratch on 32.2M tokens of security text. It answers from the passages it is given, cites one, and says so when none holds the answer — tested on 1,200 questions about techniques, CVEs and weaknesses it never saw in tuning. It writes answers well and copies exact IDs badly, so IDs are checked against the citation. A model twice its size, on a larger corpus, copies IDs better but declines less often when no passage holds the answer, and declining is what Sati depends on, so the smaller one is served.

Held-out questions

higher is better

Question	Measure	Kushmanda	Kushmanda, 34M	Word overlap	Chance
Define a technique, CWE or CVE	rouge l	0.785	0.827	0.683	0.321
What to do about a CVE	rouge l	0.938	0.896	0.361	0.258
Mitigations for a technique	rouge l	0.998	0.946	0.613	0.264
No passage holds the answer	abstained	0.543	0.417	0.000	0.000
Technique ID by name	exact match	0.528	0.717	0.962	0.396
Behaviour → technique	exact match	0.423	0.567	0.690	0.305

Can a simulator teach root cause? Not yet.

The same ranker, trained only on 50,000 synthetic incidents, scores nearly perfectly on synthetic tests — and below a no-learning baseline on real fault-injection cases. Katyayani, trained on real data, reaches 88.4% on a system it never saw. This is why Sati's root-cause model learns from real incidents.

Zero-shot on real RCAEval cases

RCA@1 — the right service ranked first

System	Ranker	RCA@1	RCA@3
Online Boutique	Trained on simulator	0.490	0.657
Online Boutique	loudest service	0.608	0.914
Online Boutique	earliest anomaly	0.098	0.380
Sock Shop	Trained on simulator	0.265	0.576
Sock Shop	loudest service	0.747	0.967
Sock Shop	earliest anomaly	0.139	0.335

Skandamata — the defences for what was seen

Given a behaviour, rank ATT&CK's 44 mitigations. Tested on 1,269 behaviours written about 115 attackers never seen in training; the right answers are the ones ATT&CK itself links. It recommends — Dharma decides.

Held-out behaviours

higher is better

Recommender	Hit@1	Recall@5	Precision@5	MRR	nDCG@10
Skandamata	0.787	0.824	0.436	0.849	0.839
Popularity (the same list for everyone)	0.232	0.398	0.206	0.443	0.433
BM25, behaviour to mitigation text	0.195	0.259	0.128	0.344	0.276
BM25, behaviour to technique to its mitigations	0.537	0.646	0.335	0.659	0.652
TF-IDF + logistic regression	0.827	0.873	0.464	0.880	0.881
Skandamata + TF-IDF classifier, fused	0.822	0.882	0.465	0.877	0.880
TF-IDF + logistic regression, pruned as served	0.780	0.842	0.450	0.850	0.850

A plain TF-IDF classifier beats Skandamata here (Hit@1 0.827 against 0.787). When Brahmacharini finds the technique, Sati AI shows only the mitigations ATT&CK links to it, in Skandamata's order, so its first pick is always one ATT&CK recommends.

Mori can communicate using HTTP over IPv4 or IPv6 depending on a flag set.
ATT&CK says: Filter Network Traffic, Network Intrusion Prevention
Skandamata: Network Intrusion Prevention, Filter Network Traffic, Network Segmentation, Disable or Remove Feature or Program, Data Loss Prevention
Qilin can employ several code obfuscation methods, including renaming functions, altering control flows, and encrypting strings.
ATT&CK says: Antivirus/Antimalware, Behavior Prevention on Endpoint
Skandamata: Behavior Prevention on Endpoint, Antivirus/Antimalware, User Training, Audit, Execution Prevention
GoldMax has used scheduled tasks to maintain persistence.
ATT&CK says: Audit, Operating System Configuration, Privileged Account Management, User Account Management
Skandamata: User Account Management, Audit, Operating System Configuration, Privileged Account Management, Multi-factor Authentication

Try Shailaputri, Siddhidatri and Mahagauri

The three models that watch servers, run live on incidents from machines none of them trained on.

Try it

runs live in the API

Loading the incidents…

Mahagauri — is it really over?

327 real server incidents. Once a minute, from onset until 60 minutes after the end, the verifier says whether the server has recovered, and declares it over once recovery has held long enough. The error that matters is a premature all-clear. It now stays within its 5% budget on servers it never saw, but it confirms few recoveries: most incidents still go back to a person.

Held-out machine groups

premature all-clear: lower is safer

Verifier	Premature all-clear	Confirmed	Left to a human
Mahagauri	0.021	0.095	0.884
Mahagauri + trajectory features	0.025	0.110	0.866
back within 3σ	0.165	0.248	0.587
fixed timer	0.480	0.520	0.000

Siddhidatri — one verdict from several detectors

28 real servers (SMD (Server Machine Dataset)). Siddhidatri stacks Shailaputri and two statistical detectors, trained out of fold, and each machine group is held out in turn. It matches Shailaputri alone on F1 but raises far more false alarms, so Shailaputri's alarm stays the one Sati acts on and the fusion is advisory.

Mean over held-out machine groups

false alarms per 1000 minutes: lower is quieter

Detector	Precision	Episode recall	F1	False alarms / 1000	Lead (min)
Siddhidatri	0.558	0.774	0.537	47.8	98
Shailaputri alone	0.642	0.597	0.539	3.8	122
Robust z-score, largest across metrics	0.168	0.778	0.253	20.4	98
PCA	0.448	0.724	0.416	84.5	99

Lead is how many minutes before an anomaly's end the first flag came: higher is earlier.

Kalaratri — anomalies hidden in real logs

Two real logs, each split by time — trained on the earliest part, tested on the latest — against two classic detectors on exactly the same test data. HDFS is the easy benchmark; BGL, a supercomputer's log over seven months, is the hard one.

Try it

runs live in the API · research use only

Loading the examples…

LogHub HDFS_v1: 115,013 held-out HDFS blocks

2.9% anomalous · later in time than anything trained on

Detector	Precision	Recall	F1	False-positive rate
Kalaratri	1.000	0.934	0.966	0.0000
Kalaratri + novelty features	1.000	0.939	0.968	0.0000
PCA (Xu et al. 2009)	0.766	0.996	0.866	0.0045
unseen-template rule	1.000	0.432	0.604	0.0000

Research use only, not for commercial use (CC BY-NC-SA 4.0): trained on research-use-only data; owner decision, 27 Oct 2026. LogHub is research-use only.

LogHub BGL: 9,427 held-out windows of 100 lines

10.2% anomalous · later in time than anything trained on

Detector	Precision	Recall	F1	False-positive rate
Kalaratri	0.513	0.271	0.355	0.0244
Kalaratri + novelty features	0.778	0.220	0.343	0.0059
PCA (Xu et al. 2009)	0.092	0.829	0.166	0.7716
unseen-template rule	0.098	1.000	0.179	0.8696

Research use only, not for commercial use (CC BY-NC-SA 4.0): trained on research-use-only data; owner decision, 27 Oct 2026. LogHub is research-use only.

Katyayani — root cause, on real failures

Trained and tested on 735 real fault-injection incidents from RCAEval across three microservice systems. Each row holds something out of training entirely, and both bars are scored on exactly the same incidents. Hover a row for its interval.

Try it

runs live in the API

Loading the incidents…

A system it has never seen

top-1 accuracy

trained modelno-learning baselinewhiskers: 95% interval

Online Boutiquenever seen in training · 430 training incidentsn = 215
87.9%
57.2%
Sock Shopnever seen in training · 430 training incidentsn = 215
94.9%
76.7%
Train Ticketnever seen in training · 430 training incidentsn = 215
82.3%
37.2%

Share of incidents where the true root-cause service is ranked first.

A kind of failure it has never seen

top-1 accuracy

trained modelno-learning baselinewhiskers: 95% interval

cpufault type never seen in trainingn = 120
96.7%
58.3%
delayfault type never seen in trainingn = 120
79.2%
54.2%
diskfault type never seen in trainingn = 120
100.0%
64.2%
lossfault type never seen in trainingn = 120
66.7%
36.7%
memfault type never seen in trainingn = 120
100.0%
68.3%
socketfault type never seen in trainingn = 45
100.0%
66.7%
code-level faults (RE3)a whole suite never trained onn = 90
80.0%
48.9%

Each fault type held out of training in turn; RE3 never trained on at all.

Pooled results

Wilson 95% intervals

Held out	Model	Top-1	Top-3	Avg@5
leave one system out	Katyayani	88.4%	96.7%	95.3%
leave one system out	baseline	57.0%	87.1%	81.3%
held out fault	Katyayani	89.3%	97.0%	95.9%
held out fault	baseline	57.0%	87.1%	81.3%
held out repetition	Katyayani	91.5%	98.2%	97.1%
held out repetition	baseline	61.2%	90.3%	84.6%
unseen system, without series count	Katyayani (ablated)	88.7%	97.0%	95.4%

The ablated row removes a feature that could let the model guess the kind of service rather than read what changed. It scores the same — the model is reading the metrics.

What a ranking cites

re1ob_adservice_cpu_1

Every ranking names the metric series behind it, so the answer can be checked — the same rule as for any hypothesis Sati makes. Root cause here: adservice.

shift_cpufromadservice_cpu+3.5509
shift_anyfromadservice_cpu+2.444
rank_shift_anyfromadservice_cpu+1.7611

The incident’s injection time is given to the model, which is RCAEval’s standard setting; a deployed system has to detect onset itself. Model sha256 5cff5f7588cbf46f…. These models were designed, written and trained by an AI assistant under the owner’s direction. Evaluation

Sati models

Models

one per stage of the loop

Trained

9 of 9

each on real public data

Katyayani, unseen systems

88.4%

top-1 root cause · baseline 57.0%

Real incidents tested

735

RCAEval · pinned afeacb11

Where each model sits in the loop

Observe
Remember
- Brahmacharini
Understand
- Katyayani
Explain
- Kushmanda
Act
- Skandamata
Verify
- Mahagauri
Decide
- Siddhidatri

1Observetrained
Metric anomaly detection — learns what normal looks like for every series
Shailaputri — the steady foundation
NABSMDRP-60
Results ↓Try it on a held-out server →
On Kaggle:dataset ↗notebook ↗model ↗
2Remembertrained
Embeddings for memory — maps observed behaviour to the known technique it resembles
Brahmacharini — discipline and study
MITRE ATT&CK proceduresRP-64
Results ↓Ask it in Sati AI →
On Kaggle:dataset ↗notebook ↗model ↗
3Observetrained
Network intrusion alerting — flags attack traffic by family
Chandraghanta — the bell that warns
UNSW-NB15CIC-IDS2017RP-61
Results ↓Try it on a held-out flow →
On Kaggle:dataset ↗notebook ↗model ↗
4Explaintrained
Writes the structured hypothesis and explanation — Sati's own language model
Kushmanda — the creator
MITRE ATT&CK, CAPEC, CWECISA KEVNVD CVE descriptionsRP-65
Results ↓Not served publicly yet
5Acttrained
Proposes the protective response — Dharma still authorises it
Skandamata — the protecting mother
MITRE ATT&CK mitigationsRP-72
Results ↓Ask it in Sati AI →
On Kaggle:dataset ↗notebook ↗model ↗
6Understandtrained
Root-cause ranking — names the service at fault and the fault type, citing the metric series
Katyayani — the warrior who finds the adversary
RCAEvalRP-58
Top-1 root cause on systems never seen in training: 88.4% vs 57.0% without learning.
Results ↓Try it on a held-out incident →
On Kaggle:dataset ↗notebook ↗model ↗
7Observetrained
Log anomaly detection — what hides in the noise
Kalaratri — the night that dispels darkness
LogHub HDFSLogHub BGLRP-62
Results ↓Try it on a held-out session →
On Kaggle:dataset ↗notebook ↗model ↗
8Verifytrained
Verification — confirms a system has truly returned to normal after an action
Mahagauri — calm restored
SMD incidentsRP-71
Results ↓Try it on a held-out server →
On Kaggle:dataset ↗notebook ↗model ↗
9Decidetrained
Fusion — combines several detectors into one calibrated verdict
Siddhidatri — completion
SMD: Shailaputri + baselines, stacked out of foldRP-73
Results ↓Try it on a held-out server →
On Kaggle:dataset ↗notebook ↗model ↗

Shailaputri — anomalies in real metric series

F1 on a family it never saw

vs the best statistical detector

trained modelno-learning baselinewhiskers: 95% interval

realAWSCloudwatchfamily never seen in training · best statistical: seasonal-naiven = 17
40.4%
33.4%
realAdExchangefamily never seen in training · best statistical: seasonal-naiven = 6
35.6%
29.5%
realKnownCausefamily never seen in training · best statistical: seasonal-naiven = 7
44.0%
30.3%
realTrafficfamily never seen in training · best statistical: seasonal-naiven = 7
49.6%
54.1%
realTweetsfamily never seen in training · best statistical: seasonal-naiven = 10
42.5%
25.9%

Episode recall and point precision combined; the threshold was chosen on training families only.

Mean across held-out families

Detector	F1	Recall	False alarms / 1000	Lead
Shailaputri	0.424	0.739	24.1	179.3
robust-z	0.325	0.911	78.7	234.1
ewma	0.261	0.900	14.2	229.0
Seasonal naive	0.346	0.930	74.7	240.0
Seasonal robust z-score	0.312	0.761	74.7	241.3

NAB’s labels are windows, and Wu & Keogh (2021) show several NAB series are trivial or mislabelled; these numbers inherit those limits.

Brahmacharini — from behaviour to the technique it is

Held-out behaviour descriptions

n = 1727

Retriever	Recall@1	Recall@5	Recall@10	MRR
Brahmacharini	0.614	0.802	0.854	0.699
BM25	0.268	0.478	0.580	0.372
TF-IDF + logistic regression	0.646	0.832	0.869	0.730
Brahmacharini + TF-IDF classifier, fused	0.673	0.859	0.896	0.758
Hashing embedder (Smriti's previous retriever)	0.120	0.226	0.283	0.177
Brahmacharini, as served	0.556	0.735	0.797	0.638
Brahmacharini as served + TF-IDF classifier, fused	0.650	0.847	0.887	0.738

Chandraghanta — attack traffic in real labelled flows

Try it

runs live in the API · research use only

Loading the examples…

UNSW-NB15: 82,332 held-out flows

55.1% attacks · official training and testing files

Detector	Precision	Recall	F1	False-positive rate
Chandraghanta, best-F1 threshold	0.821	0.988	0.897	0.2634
Chandraghanta, best-F1 threshold, without TTL fields	0.827	0.981	0.897	0.2522
Chandraghanta	0.952	0.926	0.939	0.0578
PCA, trained on benign flows	0.581	0.993	0.733	0.8782

Attack family	Flows	Seen in training	Chandraghanta, best-F1 threshold recall	Chandraghanta, best-F1 threshold, without TTL fields recall	Chandraghanta recall	pca (benign-only) recall
Analysis	677	yes	0.975	0.972	0.916	0.994
Backdoor	583	yes	1.000	1.000	0.990	0.997
DoS	4,089	yes	0.999	0.997	0.987	0.988
Exploits	11,132	yes	0.996	0.992	0.973	0.995
Fuzzers	6,062	yes	0.919	0.881	0.531	0.979
Generic	18,871	yes	1.000	1.000	0.999	1.000
Reconnaissance	3,496	yes	0.999	0.997	0.990	0.984
Shellcode	378	yes	0.992	0.992	0.921	0.989
Worms	44	yes	1.000	1.000	0.977	1.000

Research use only, not for commercial use (CC BY-NC-SA 4.0): trained on research-use-only data; owner decision, 27 Oct 2026. UNSW-NB15 and CIC-IDS2017 are research-use only.

CIC-IDS2017: 1,162,213 held-out flows

25.1% attacks · by day: trained on Monday to Wednesday, tested on Thursday and Friday

Detector	Precision	Recall	F1	False-positive rate
Chandraghanta, best-F1 threshold	0.989	0.294	0.453	0.0011
Chandraghanta	0.927	0.448	0.604	0.0118
PCA, trained on benign flows	0.440	0.135	0.207	0.0575
random split (reference only)	0.998	0.996	0.997	0.0005

Attack family	Flows	Seen in training	Chandraghanta, best-F1 threshold recall	Chandraghanta recall	pca (benign-only) recall
Bot	1,966	no	0.000	0.000	0.028
DDoS	128,027	no	0.636	0.975	0.215
Infiltration	36	no	0.000	0.000	0.722
PortScan	158,930	no	0.013	0.020	0.074
Web Attack - Brute Force	1,507	no	0.908	1.000	0.000
Web Attack - Sql Injection	21	no	1.000	1.000	0.000
Web Attack - XSS	652	no	0.974	1.000	0.000

Research use only, not for commercial use (CC BY-NC-SA 4.0): trained on research-use-only data; owner decision, 27 Oct 2026. UNSW-NB15 and CIC-IDS2017 are research-use only.

Kushmanda — Sati's own language model

Held-out questions

higher is better

Question	Measure	Kushmanda	Kushmanda, 34M	Word overlap	Chance
Define a technique, CWE or CVE	rouge l	0.785	0.827	0.683	0.321
What to do about a CVE	rouge l	0.938	0.896	0.361	0.258
Mitigations for a technique	rouge l	0.998	0.946	0.613	0.264
No passage holds the answer	abstained	0.543	0.417	0.000	0.000
Technique ID by name	exact match	0.528	0.717	0.962	0.396
Behaviour → technique	exact match	0.423	0.567	0.690	0.305

Can a simulator teach root cause? Not yet.

Zero-shot on real RCAEval cases

RCA@1 — the right service ranked first

System	Ranker	RCA@1	RCA@3
Online Boutique	Trained on simulator	0.490	0.657
Online Boutique	loudest service	0.608	0.914
Online Boutique	earliest anomaly	0.098	0.380
Sock Shop	Trained on simulator	0.265	0.576
Sock Shop	loudest service	0.747	0.967
Sock Shop	earliest anomaly	0.139	0.335

Skandamata — the defences for what was seen

Held-out behaviours

higher is better

Recommender	Hit@1	Recall@5	Precision@5	MRR	nDCG@10
Skandamata	0.787	0.824	0.436	0.849	0.839
Popularity (the same list for everyone)	0.232	0.398	0.206	0.443	0.433
BM25, behaviour to mitigation text	0.195	0.259	0.128	0.344	0.276
BM25, behaviour to technique to its mitigations	0.537	0.646	0.335	0.659	0.652
TF-IDF + logistic regression	0.827	0.873	0.464	0.880	0.881
Skandamata + TF-IDF classifier, fused	0.822	0.882	0.465	0.877	0.880
TF-IDF + logistic regression, pruned as served	0.780	0.842	0.450	0.850	0.850

Mori can communicate using HTTP over IPv4 or IPv6 depending on a flag set.
ATT&CK says: Filter Network Traffic, Network Intrusion Prevention
Skandamata: Network Intrusion Prevention, Filter Network Traffic, Network Segmentation, Disable or Remove Feature or Program, Data Loss Prevention
Qilin can employ several code obfuscation methods, including renaming functions, altering control flows, and encrypting strings.
ATT&CK says: Antivirus/Antimalware, Behavior Prevention on Endpoint
Skandamata: Behavior Prevention on Endpoint, Antivirus/Antimalware, User Training, Audit, Execution Prevention
GoldMax has used scheduled tasks to maintain persistence.
ATT&CK says: Audit, Operating System Configuration, Privileged Account Management, User Account Management
Skandamata: User Account Management, Audit, Operating System Configuration, Privileged Account Management, Multi-factor Authentication

Try Shailaputri, Siddhidatri and Mahagauri

The three models that watch servers, run live on incidents from machines none of them trained on.

Try it

runs live in the API

Loading the incidents…

Mahagauri — is it really over?

Held-out machine groups

premature all-clear: lower is safer

Verifier	Premature all-clear	Confirmed	Left to a human
Mahagauri	0.021	0.095	0.884
Mahagauri + trajectory features	0.025	0.110	0.866
back within 3σ	0.165	0.248	0.587
fixed timer	0.480	0.520	0.000

Siddhidatri — one verdict from several detectors

Mean over held-out machine groups

false alarms per 1000 minutes: lower is quieter

Detector	Precision	Episode recall	F1	False alarms / 1000	Lead (min)
Siddhidatri	0.558	0.774	0.537	47.8	98
Shailaputri alone	0.642	0.597	0.539	3.8	122
Robust z-score, largest across metrics	0.168	0.778	0.253	20.4	98
PCA	0.448	0.724	0.416	84.5	99

Lead is how many minutes before an anomaly's end the first flag came: higher is earlier.

Kalaratri — anomalies hidden in real logs

Try it

runs live in the API · research use only

Loading the examples…

LogHub HDFS_v1: 115,013 held-out HDFS blocks

2.9% anomalous · later in time than anything trained on

Detector	Precision	Recall	F1	False-positive rate
Kalaratri	1.000	0.934	0.966	0.0000
Kalaratri + novelty features	1.000	0.939	0.968	0.0000
PCA (Xu et al. 2009)	0.766	0.996	0.866	0.0045
unseen-template rule	1.000	0.432	0.604	0.0000

Research use only, not for commercial use (CC BY-NC-SA 4.0): trained on research-use-only data; owner decision, 27 Oct 2026. LogHub is research-use only.

LogHub BGL: 9,427 held-out windows of 100 lines

10.2% anomalous · later in time than anything trained on

Detector	Precision	Recall	F1	False-positive rate
Kalaratri	0.513	0.271	0.355	0.0244
Kalaratri + novelty features	0.778	0.220	0.343	0.0059
PCA (Xu et al. 2009)	0.092	0.829	0.166	0.7716
unseen-template rule	0.098	1.000	0.179	0.8696

Research use only, not for commercial use (CC BY-NC-SA 4.0): trained on research-use-only data; owner decision, 27 Oct 2026. LogHub is research-use only.

Katyayani — root cause, on real failures

Try it

runs live in the API

Loading the incidents…

A system it has never seen

top-1 accuracy

trained modelno-learning baselinewhiskers: 95% interval

Online Boutiquenever seen in training · 430 training incidentsn = 215
87.9%
57.2%
Sock Shopnever seen in training · 430 training incidentsn = 215
94.9%
76.7%
Train Ticketnever seen in training · 430 training incidentsn = 215
82.3%
37.2%

Share of incidents where the true root-cause service is ranked first.

A kind of failure it has never seen

top-1 accuracy

trained modelno-learning baselinewhiskers: 95% interval

cpufault type never seen in trainingn = 120
96.7%
58.3%
delayfault type never seen in trainingn = 120
79.2%
54.2%
diskfault type never seen in trainingn = 120
100.0%
64.2%
lossfault type never seen in trainingn = 120
66.7%
36.7%
memfault type never seen in trainingn = 120
100.0%
68.3%
socketfault type never seen in trainingn = 45
100.0%
66.7%
code-level faults (RE3)a whole suite never trained onn = 90
80.0%
48.9%

Each fault type held out of training in turn; RE3 never trained on at all.

Pooled results

Wilson 95% intervals

Held out	Model	Top-1	Top-3	Avg@5
leave one system out	Katyayani	88.4%	96.7%	95.3%
leave one system out	baseline	57.0%	87.1%	81.3%
held out fault	Katyayani	89.3%	97.0%	95.9%
held out fault	baseline	57.0%	87.1%	81.3%
held out repetition	Katyayani	91.5%	98.2%	97.1%
held out repetition	baseline	61.2%	90.3%	84.6%
unseen system, without series count	Katyayani (ablated)	88.7%	97.0%	95.4%

The ablated row removes a feature that could let the model guess the kind of service rather than read what changed. It scores the same — the model is reading the metrics.

What a ranking cites

re1ob_adservice_cpu_1

Every ranking names the metric series behind it, so the answer can be checked — the same rule as for any hypothesis Sati makes. Root cause here: adservice.

shift_cpufromadservice_cpu+3.5509
shift_anyfromadservice_cpu+2.444
rank_shift_anyfromadservice_cpu+1.7611