Skip to content
ORBITAL SEGMENT · CITED RCA · AUTHORITY LEVEL A0

Jantar Ops: Cited Anomaly Investigation

From a telemetry deviation to a root cause and a drafted recovery procedure. Every claim is cited or marked ungrounded, and a human reviews and authorises it.

WHAT IT DOES

From Deviation to a Cited Procedure

A telemetry trace spikes out of its band; a command log is checked; a knowledge graph points into a stack of documents; a drafted procedure links each line back to them; a person at a console decides
DetectDeviationTelemetry drifts out of family.
FilterCommand History FirstA commanded response is not mistaken for a fault.
InvestigateCited Root CauseCandidate causes come from the spacecraft’s own failure modes and documents.
AuthoriseHeld for a HumanA drafted recovery procedure, every step cited. A person reviews and authorises it.
CITED OR UNGROUNDEDA step the evidence does not support is marked UNGROUNDED, carries no citation, and blocks release.
DECLINING IS AN ANSWERWhere the evidence is absent the verdict is INSUFFICIENT.
AUTHORITY LEVEL A0At A0, the default, it constructs no command candidate; A1 staging is an opt-in, off unless a deployment enables it. No route transmits, and nothing an LLM writes can reach a transmitter.

Measured on the Reference Scenarios

Measured on seeded replay scenarios of JANTAR-REF-1, a synthetic reference spacecraft — not a real mission.

MeasureResult
Root-cause top-11.000
Citation integrity1.000
Investigations16/20
Grounded honesty (no fabrication)1.000
No unsafe release1.000

First three rows: Measured on synthetic · JANTAR-REF-1. mode=airgap · model=deterministic · detector=annotation. Source: make bench. Measured at commit 2bc4474. From this release’s scorecard. Last two rows: Measured on synthetic · JANTAR-REF-1. mode=airgap · model=deterministic · grounded_faults=16 · moat_holds=True. Source: make bench-moat. Measured at commit 2bc4474. From this release’s scorecard.

Reference Subsystem Coverage

The failure modes modelled on JANTAR-REF-1. An operator’s own craft is onboarded onto its own FMECA and documents.

Bus SubsystemModelled Failure Modes
EPS (Electrical Power)Battery cell degradation, solar array string interconnection loss, battery thermal runaway precursor
AOCS (Attitude and Orbit Control)Reaction wheel bearing lubricant starvation, rate gyro bias drift, wheel tachometer signal noise
TCS (Thermal Control)Survival heater relay contact fusion, radiator surface coating degradation
TTC (Telemetry and Command)Transponder power amplifier gain degradation
OBDH (On-Board Data Handling)SEU-induced memory bit-flip accumulation, processor thermal throttling load spike

Onboarding Your Own Spacecraft

In the console or over the API. Two people are involved, and nothing changes until the second one confirms.

OperatorValidateThe export is checked and every finding is named. Nothing is stored.
ApproverConfirmA second person confirms the validated export. The craft is live without a restart.
OperatorInvestigateA telemetry window is investigated, and its citations resolve against that craft’s own documents.

Each craft’s records are kept apart per tenant and per craft, and a citation resolves only within its own craft. Re-onboarding a changed export merges into the craft and never removes what the new export dropped.

Prior Art and What Jantar Adds

Grounded anomaly investigation is an active field, and parts of it are decades old. Each line below names a design difference; it is not a comparison of results.

ESA AISHGO: AIIA and SMAI

AIIA investigates ground-station incidents; Jantar investigates spacecraft telemetry anomalies, and every citation must resolve against the craft’s own records.

What its sources say

An ESA project with Solenix, Starion, FBK, Airbus and DLR, which Solenix reports delivered AIIA and SMAI as operational tools. AIIA, led by Starion, helps ground-station engineers investigate incidents: it retrieves from past incidents and documents, checks each answer against them with a judging model, links its sources, and has operators review its conclusions. SMAI matches telemetry patterns inside the telemetry tools at ESOC and DLR GSOC.

Sources: Solenix · Starion · AISHGO paper · DLR

Sift Agents (research preview)

Jantar holds its citation rule in both directions: a grounded claim without a citation that resolves is a violation, and so is an UNGROUNDED step that carries one. Either blocks release.

What its sources say

Agents that investigate hardware telemetry, test several hypotheses in parallel and surface root-cause candidates with their evidence. Sift states that every step links back to the underlying data, and that write operations require human approval.

Sources: Sift Agents · Sift AI

CelestCore

Jantar is not the only system that declines. Its domain is anomaly root cause, and its decline, an INSUFFICIENT verdict, is a releasable answer: declining is not withholding.

What its sources say

Decision support for collision-risk alerts. When confidence is insufficient it formally declines to recommend, and every recommendation passes a separate verification path that cannot be bypassed. It is advisory and does not execute manoeuvres.

Sources: CelestCore

OpenC3 COSMOS Enterprise

A design contrast: COSMOS lets the model an operator connects act with that operator’s access. Jantar is advisory, and there is no path from a model to a transmitter.

What its sources say

Ships an MCP server that gives a model the customer supplies access to the COSMOS API, normally authenticated as the logged-in user. Its built-in AI interface reads telemetry, sends commands and runs scripts through that server. COSMOS bundles no model of its own.

Sources: OpenC3 docs

Remote Agent and Livingstone (NASA Ames and JPL)

Diagnosis against a model of the spacecraft is decades old, and Jantar’s reasoning is in that line. Jantar runs on the ground and drafts a recovery procedure for a human to review, with the evidence for each step cited.

What its sources say

Model-based fault diagnosis and recovery, flight-validated on Deep Space One. Livingstone identified component faults by comparing sensor observations with models of the spacecraft, and Remote Agent issued the recovery commands on board.

Sources: NASA NTRS · JPL

SHINE (JPL)

SHINE encodes the knowledge of human experts. Jantar reasons over the operator’s own FMECA and documents, and each claim cites the record it rests on.

What its sources say

The Spacecraft Health Inference Engine, an expert system for monitoring, analysing and diagnosing real-time and non-real-time systems. NASA lists its use in diagnosing Voyager telecommunication anomalies at the Neptune encounter, and on the Mars Exploration Rovers.

Sources: NASA NTRS

Inductive Monitoring System (NASA Ames)

Jantar’s detectors are in this line, scoring deviation from a nominal profile. Detection is not what Jantar adds; the cited investigation after it is.

What its sources say

Learns clusters of nominal behaviour from archived data and measures how far new data falls from them. NASA reports it deployed as a tool in the ISS flight control room.

Sources: NASA NTRS · ISS deployment

What Jantar adds, narrowly

  • Citation integrity for anomaly root cause, enforced in both directions.
  • INSUFFICIENT is a releasable answer. Declining is not withholding.
  • Deterministic replay: with the deterministic reasoning model, a re-run reproduces the report byte for byte.
  • An air-gap mode that runs on the operator’s own infrastructure and refuses every network data source.
  • No path from a model to a transmitter.

Other systems are described as their own publications describe them, and those may lag their current versions.

Sovereign & Airgap Deployment Notice

Jantar Ops ships at Authority Level A0 (Advisory Only). In airgap mode it runs with the deterministic reasoning model on the operator’s own infrastructure and refuses every network data source.