Repository navigation
0.41 adds live ingest (Kshana Connect mapping files, a directory watch, MQTT, NATS and WebSocket inputs, an NTRIP client, live
receiver, clock, sensor, IQ and ADS-B/AIS recording), a headless receiver service (Kshana Station), Kshana Desktop, a C interface,
CAP 1.2 alerts and signed webhooks, one observation schema (kshana.obs/1), area trust from several reference stations, a
monitoring-network planner, location metrics in the referee, an environmental baseline, a fourth benchmark task (e) with the
TEXBAT spoofing scores, the flight-dynamics export and its adapters, and the tip-and-cue scenario kind. It records the scored
pre-registered rows of the 0.41 programme below, every NOT MET beside every MET; no bar was loosened. The verification matrix
grows from 401 to 413 rows: 176 VALIDATED (unchanged), 233 MODELLED (was 221) and 4 PARTNER. The twelve new rows are all MODELLED,
and no row was promoted to VALIDATED. The scenario kinds grow from 92 to 93 and the MCP server's tools from 49 to 53. The advisory
statement is unchanged: Kshana is advisory software, not type-approved navigation equipment; the operator remains responsible.
Known limits
- The B3a power-difference locator (
kshana::locate_pd) and the B3b TDOA locator (kshana::locate_tdoa) are Rust library
only, with no command line, Python or WebAssembly entry; both synthetic gates are NOT MET and neither is run on real data. - The broadcast-health rule reads GPS and QZSS health only; Galileo and BeiDou health is not read yet (their health bits are per
signal). - The benchmark leaderboard page renders tasks (a) to (d) only; task (e) results stay in
xval/bench-e/run/. - The Studio's IQ recording view works in Kshana Desktop only; on the site it says so.
Added
kshana ephem-export(Python, WebAssembly and MCPephem_export;kshana::ephem_export): a JPL Horizons state table
written as a CCSDS OEM 2.0, a CCSDS OPM 3.0 or an SPK kernel of type 13 or 9, then read back with the engine's own readers
(C7, step 1;docs/FLIGHT-DYNAMICS-EXPORT.md). Values and epochs are copied, not computed: on the mission thread's
CAPSTONE table all 4,321 states read back with the same bits from the OEM and the SPK, and from the SPICE Toolkit's
reading of the SPK too. Every surface writes the same bytes (pinned SHA-256).integrations/flight-dynamics/: the C7 adapters and notebooks, Orekit and Basilisk first. Orekit 13.1.9 reads the CAPSTONE OEM and
OPM Kshana wrote with its own CCSDS parsers; every value it holds is its own kilometre-to-metre conversion of the written
double and every epoch is the written one (4,321 states). The difference report names Orekit as rounding further than
the digits written: that conversion is a second rounding of at most one ulp. Theflight-dynamicsworkflow runs the
notebooks, checks that the command line and Python write the same bytes, and runskshana connect check. Basilisk
2.12.2, through the SPICE Toolkit it ships, reads the SPK's 4,321 stored records and epochs with the written bits.- ANISE 0.10.6, the fourth, file-based C7 tool (
integrations/flight-dynamics/anise.ipynb): its own kernel reader evaluates
Kshana's SPK, and its OEM conversion, at every node epoch no further from the written states than SPICE does (1.46e-11 km).
It reads SPK epochs as SPICE ET and OEMTDBlabels with hifitime's TDB model, 3.2 µs apart here; the notebook shows why
the scale matters. [design.carrier] fll_handover_mean_s(RustLoopConfig::fll_handover_mean_s), withfll_pull_in_max_s: a forced
FLL-to-PLL hand-over starts the PLL at the mean NCO Doppler of the pull-in's last this-many seconds. Opt-in, out of the
design hash when unset. Scored under D8a row 4.iq::acq::CfarReference::Row(CfarConfig::with_reference): CA-CFAR reference cells every stride-th lag round the
whole Doppler row. Opt-in; the 16-cell window stays the default. Scored under D8a row 3 (met).[monitors] roc_restart(default on): the detection ROC scores each C/N0 CUSUM window as restarted at its start
(D8b row 9, met); a campaign's automatic C/N0 stride is now one NWPR window (D8b row 8).- RTCM SSR reader (
kshana::rtcm_ssr) reads the IGS SSR format (RTCM 4076): orbit, clock, combined, high-rate clock, code
bias, phase bias with yaw and URA for GPS, GLONASS, Galileo, QZSS, BeiDou and SBAS, and the global ionosphere (201). Every
field of the 11 messages in pyrtcm 1.2.0's public IGS sample stream matches the values pyrtcm records; the message kinds
missing from that stream are checked against hand-built frames only. MODELLED. kshana.obs/1correction records, andkshana obs convert --format rtcm-ssr: an RTCM 3 stream's SSR corrections (RTCM
and IGS SSR) become one record per satellite of each message, witht_son GPS time. Every value of the 676 records from
the three public sample streams equals pyrtcm's after a round trip throughobs check;obs checkcounts corrections.
MODELLED; the limits are indocs/design/evidence/r1-1-rtcm-ssr/RESULTS.md.leo-signal: GPS L2C as a public reference preset (gps-l2c, from IS-GPS-200N: 1227.60 MHz, CM and CL time-multiplexed,
−160.0 dBW) and as a GNSS victim in the compatibility analysis, so a LEO-PNT design is scored against L2C and L2C against it.
MODELLED: checked against the CM and CL codes iniq::signals::gpsand the BPSK closed forms only.kshana::solid_tide: the IERS Conventions (2010) solid Earth tide displacement, Steps 1 and 2 (degree 2 and 3, the
latitude, out-of-phase and frequency-dependence terms), written from the Conventions text. MODELLED; not yet used by PPP. The
comparison with the IERS test cases waits for its pre-registration's sign-off (R1.4).kshana::bias_sinex: a Bias-SINEX 1.00 reader (DSB, ISB and OSB; code and phase; offsets, slopes and open intervals), and a
lookup of a satellite's observable-specific bias at a time. On two public files every data line is read with the values
printed. Both files' headers state a count of estimates that differs from their data lines. MODELLED; applying the biases
in PPP is item R1.3.kshana::tip_and_cue(row M2, library): Walker star RF and optical layers on 550 km sun-synchronous orbits over the 996
Baltic sea points (data/tip-and-cue/baltic_sea_points.csv, from Natural Earth), access intervals per point, and the
revisit, full-area refresh and RF-to-optical reaction figures; public helpers for the area grid, the sea mask, the primary
RF layer and one of its planes. Bar M2-B0 met: 1 283 access intervals equal the independent oracle's to the second. The
tip-and-cuescenario kind now runs it (scenarios/tip-and-cue-baltic-m2.toml,scenarios/tip-and-cue-short.toml,
docs/TIP-AND-CUE.md): 93 scenario kinds. MODELLED.kshana.obs/1, one observation schema for stations and systems (header, epochs with optional IQ references, alarms), with
its JSON Schema,kshana obs convert <log> --format <f> --station <station.toml>andkshana obs check <file>
(docs/OBS-SCHEMA.md). RINEX 4.00 observation files are read and give the same epochs as 3.04.- Kshana Connect mapping files.
kshana connect map <input> --mapping <connect.toml> --station <station.toml>turns any CSV,
TSV, JSON or JSON Lines export intokshana.obs/1through akshana.connect/1TOML file, with no code: time as RFC 3339, Unix
s or ms, GPS seconds, GPS week and time of week or a pattern, in UTC, GPS or TAI (leap seconds from the table); position with
units; per-satellite C/N0 in long or wide layout; gain control, interference indicator, the receiver's own flags and DOP; row
filters and JSON pointers. Every error names the line and the column. Python and WebAssemblyconnect_mapand the MCP tool
connect_mapgive the same bytes as the check vectors (docs/CONNECT.md). kshana connect check <vector.toml|dir> [--bless]: runs Connect check vectors (input, mapping file or built-in reader, station,
expectedkshana.obs/1output) and fails on any byte difference, naming the first differing line. The three mapping examples
(an Android raw-measurement CSV, a timing receiver's semicolon export, a monitoring service's nested JSON), the NMEA reader, a
RINEX 4.00 vector and an IEC 61162-450 vector are in the corpus; CI runs it.- Live ingest through a Connect mapping, each giving the epochs
connect mapgives on the whole log: `kshana connect watch ` (a directory of growing or rotating logs read as one continuous log; `--suffix`, `--interval`, `--once`), `kshana connect mqtt` (MQTT 3.1.1, QoS 0 and 1, credentials from the environment, keep-alive, reconnection, `--per-topic` for wildcard subscriptions), `kshana connect nats` (MSG and HMSG, PING/PONG, credentials or token from the environment, reconnection, `--per-topic`, station ids `/`) and `kshana connect ws` (an RFC 6455 client: checked upgrade, fragments, ping/pong, close). Memory is bounded; feeds trim after every epoch. kshana connect check-live <vector.toml|dir> [--json]: replays the Connect check vectors over MQTT (with a forced reconnection),
WebSocket and the directory watch, and checks order, holding, reconnection and latency.kshana connect load-testruns many
synthetic stations through the live feeds (default 1,000 at 1 Hz for 10 s); the recorded run gave 40,659 messages/s in memory
and a p99 of 28.5 ms over MQTT (release build, one core).- The adapter registry:
adapters/withkshana.adapter/1manifests (three examples),kshana connect adapters list|check
(manifest, mapping, the adapter's own check vector byte for byte, unique names), adapter-submission and integration-request
issue templates (docs/ADAPTERS.md). - Live receivers.
kshana ntrip <host:port>: an NTRIP client (NTRIP 1 ICY and NTRIP 2 with chunked encoding, Basic
authentication with the password from the environment, a GGA position, the source table; the RTCM 3 frame count does not
depend on chunking) (docs/NTRIP.md).kshana obs live <file|-|tcp:host:port> --format <f> --station <station.toml>reads a receiver stream as it arrives, with the epochs a replay gives and bounded memory (checked for SBF,
NovAtel and RTCM 3). kshana iq record: writes a capture tool's samples into SigMF as they arrive, receive only;iq monitor --followon the
growing file equals the replay.kshana interference-map record: records live ADS-B (readsbaircraft.json) and AIS (AIVDM) feeds into the CSV files the
locatereplays read.kshana clock-live --format <f> --from <-|tcp:host:port|file>: follows linuxptp, chrony, Linux PPS (the newpps-assert
clock-log format) and TICC logs and prints one JSON line per reading; the kept readings equal the replay series.kshana sensor-live: follows IMU logs (csv,euroc, and ROS 2sensor_msgs/Imufromros2 topic echo --csv, also
accepted in[log.imu]) and prints each sample with its integrated gyro heading, equal to a replay bit for bit;--kind n2k
turns NMEA 2000 frames from SocketCANcandumpinto NMEA 0183 (position, course, heading, time).- Live paths share one host time base with per-device latency, a watchdog per sensor (stale, frozen, out of range) and
reconnection with the gap marked, never filled. kshana device check <device.toml>andkshana device decode: thekshana.device/1sensor plug-in kit, a binary sensor
described by framing, messages and test vectors, decoded to JSON Lines for a Connect mapping;device checkruns the file's
own test vectors (docs/DEVICES.md).- Kshana Station (
kshana station run|check|service <station.toml>): a headless receiver service. Every configured receiver
stream (TCP, serial or USB device, file) is read as it arrives into rotatedkshana.obs/1files through the readerobs live
uses, so a connection's epochs are a replay's bit for bit; a stream that ends is reopened every second, the gap counted and never
filled. A local status port (GET /health,GET /status; another address needs a bearer token), an optional fleet report over
HTTP, andserviceprints a systemd unit, a launchd property list or a Windows start-up task. Read only: nothing is written to a
receiver.docs/STATION.mdis a field guide for one real receiver on a site. - Kshana Desktop (
apps/desktop, Tauri 2): the Studio in a native window with runs, charts, tables, the localised text and the
CCSDS and SP3 exports on the natively linked engine (the same calls as WebAssembly), files of any size read in blocks. The trust
page can score a receiver live from a serial or USB port, a TCP stream or a growing file (NMEA 0183, read only, every five
seconds). An IQ recording view (studio/iq/) shows power, a waterfall and the spectrum at a cursor for ci8, cu8, ci16 and cf32
recordings and SigMF metadata; it has no control that writes a file or transmits. The Desktop workflow builds the app on Linux,
macOS and Windows. - A C interface:
capi/kshana-capi, a separate crate so the library keepsforbid(unsafe_code), with five entry points
(version, run, validate, connect map, free) andkshana.h. C, C++, Python (ctypes) and Go (cgo) examples are built and run by the
tests against the same expected output; the C example also runs under AddressSanitizer and UndefinedBehaviorSanitizer.
SECURITY.mdandSAFETY.mdstate the boundary. - Alerts. OASIS CAP 1.2 messages from the trust stream (Alert, Update and all-clear, linked by references, with
operator-set status and scope; valid against the CAP 1.2 schema), and webhook payloads signed with HMAC-SHA256 overt.body
(Kshana-Signature, a replay window, key rotation; checked against the RFC 4231 vectors). On the command line:
kshana trust-telemetry --cap-config/--cap-dir/--webhook --webhook-secret-env. - Deployment kits: a Compose stack (monitor, Prometheus, provisioned Grafana; an ingest profile with MQTT) and a Helm chart;
the image runs as UID 10001; an end-to-end workflow feeds both the bundled replay.
docs/guides/MONITORING-STATION.mdwalks through a GNSS interference-monitoring station
(observations, trust, CAP alerts, the live stack, a signed webhook);tests/guide_monitoring_station.rsruns its steps as
written. kshana bench-speed [--repeats <n>] [--json <out>] [--baseline <old.json>]: times the live paths against a same-run
calibration and prints the station sizing table; with a baseline it fails on a slowdown over 10 %
(docs/SIZING.md). CI runs it on three platforms.- Readers and checks for other sectors: IEC 61162-450 network datagrams unwrapped to NMEA with TAG-block checksums (reader
iec61162-450); IEEE C37.118.2 PMU time-sync health (kshana::c37118: STAT sync error, time quality, unlocked time, FRACSEC
flags, CRC; checked against an independent encoder); business-clock divergence from UTC against the RTS 25 Annex limits
(kshana::rts25: excursions, gaps, within, beyond or not assessable); ASTERIX CAT021 ADS-B reports (edition 2.6 UAP) to the
interference map's ADS-B CSV (kshana::asterix_cat021; checked against an independent encoder); gps-sdr-sim motion files
(time,x,y,zandtime,lat,lon,h) imported into the test bench (round trip exact in ECEF, within 0.2 mm in LLH). - Clock-log formats: ntpd
loopstatsandpeerstats, White Rabbit PTP core stat lines (clock offsetckoin ps; the pull-in
before the firstlock:1left out), and CGGTTS 2E track files (kshana::cggtts: checksums, all-in-view series and common-view
differences, read on a public BIPM file). The CGGTTS comparison against Circular T (R6.1) is pre-registered as proposed and not
run (docs/CLOCK-LOG-FORMATS.md). - RTCM 3 SSR corrections (1057 to 1068, 1240 to 1263) for GPS, GLONASS, Galileo, QZSS, SBAS and BeiDou, and the 1264 ionosphere,
inkshana::rtcm_ssr, checked field by field against the values pyrtcm 1.2.0 records for its public sample streams. MODELLED;
no pre-registration (docs/design/evidence/r1-1-rtcm-ssr/RESULTS.md). kshana scenario migrate <scenario.toml> [--in-place]andkshana scenario version: an optionalschema_versionline in a
scenario file, checked and stripped before any kind reads the file (results byte-identical), and refused when newer than the
engine (docs/SCENARIO-VERSIONS.md).KSHANA_THREADS: campaigns, sweeps, the N-D sweep ensembles, the LEO PPP ensemble,iq track, and the clock-holdover and
inertial Monte Carlo use every core by default, with the same bytes on one core and many (kshana::par).KSHANA_MAX_MEMORY: a cap on what one command plans to hold;iq trackrefuses an over-cap--jsonbefore it starts and names
--epochs, which streams in a fixed amount of memory (docs/MEMORY.md).kshana area-trust <stations.json>(Python and WebAssemblyarea_trust_timeline;kshana::area_trust): one trust state for an
area from several reference stations, per second, with alerts, as JSON Lines (B5;docs/AREA-TRUST.md).
"area_rule": "v2"selects a two-station floor for areas of six stations or fewer;"v3"and akshana.area-measurements/1
document add a common-mode satellite-fault screen on the stations' fixes, and a loss-of-lock alarm alone no longer makes a station
Degraded(it stays in the station's record). Without the field the rule is v1, with v1's bytes.AreaLiveruns the rule on a
live stream, closing each second as soon as every station has reported. MODELLED (VM-79c6dba3); B5 v1, v2 and v3 are NOT MET.kshana planner map|place <spec.json>(Python and WebAssemblyplanner;kshana::planner): a monitoring-network planner. Per
site cell, each station's detection probability, the network's, the probability that enough stations detect to locate, and the
TDOA DRMS and 95 % ellipse;placechoosesmstations from candidates (greedy,greedy+exchangeand
greedy+exchange2+restarts, selected by name). Cells with a condition number above 1e8 carry anill_conditionedflag
(docs/PLANNER.md). MODELLED (VM-14a961c9), free space only.kshana jammertest locate-score(Python and WebAssemblyreferee_location_score, MCPreferee_location_score): location
metrics in the referee, namely error, region coverage and size, time to first location, time to locate and false locations (B2).
kshana::geodesic: the WGS 84 inverse geodesic distance (Karney 2013, a port of GeographicLib, MIT), within 1e-8 m of
GeographicLib on 3,616 pairs.kshana env-baseline <site-hours.json>(Python and WebAssemblyenv_baseline;kshana::env_baseline): an environmental baseline
per site and hour from Kp and F10.7, a score from 0 to 1 andquiet,elevatedordisturbed(B8). MODELLED.kshana data-pack <out-dir> [--seed <n>] [--logs <n>] [--iq <n>]: synthetic labelled data packs, NMEA logs with exact drag-off
spoofing labels and IQ scenes with natural impairments (obstruction, scintillation, multipath); the same seed gives the same bytes
(docs/DATA-PACKS.md).- Benchmark task (e), four-class interference labelling on receiver logs (spec 2.1.0; tasks (a) to (d) keep 2.0.0): truth from
the JammerTest 2024 folders (Zenodo record 15911589, pinned by SHA-256), the scorer, and the B-log baseline (kshana bench baseline --task e);validateaccepts taske.docs/guides/BENCHMARK-CLASSES.md
scores an interference classifier on it step by step. - TEXBAT spoofing scores (
VM-bd8626f0):kshana bench texbat-windows(the fixed B-iq baseline in 1 s windows, acquired once per
recording then tracked continuously throughsdr::track_continue),kshana bench texbat-pd(power-and-distortion window scores)
andkshana bench texbat-score(the registered AUC from the committed window-score files). The recordings stay on the user's
computer; only window scores are written.kshana bench verify-manifestchecks an external source's files against pinned sizes
and SHA-256 (hashed in 1 MiB blocks) and refuses a file that differs; all ten TEXBAT recordings are now pinned in
bench/manifests/b-iq.json. MODELLED. kshana::interference_map::almanac_hdop: a YUMA almanac reader, the IS-GPS-200 almanac orbit, HDOP at a position with a stated
mask, the NACp 95 % bound and an EPU-against-HDOP flag with the mask, UERE and multiple stated by the caller (D18 item 1). Built
from public sources as Kshana's own method, not a reproduction of the cited paper; no real data scored.kshana::space_geoloc(M1): space RF geolocation bounds for a surface emitter seen from low Earth orbit (a 550 km sun-synchronous
orbit with secular J2, passes, single-satellite frequency of arrival, two- and three-satellite TDOA/FDOA formations, link-budget
noise per emitter class, several passes fused). Library. MODELLED (VM-46168417).kshana::m4_dark_vesselandkshana::m4_logic(M4): a synthetic Baltic truth-set generator and dark-vessel and AIS-integrity
decisions from RF fixes against AIS reports (NORMAL,DARK,GAP,POSITION-OFFSET,IDENTITY-DUPLICATE;M4Params::v1,
v2,v3). Every output describes data consistency, never intent. Library. MODELLED (VM-a66a88b7).kshana::m5_hybrid(M5): interference-source location from orbit, from a ground network and from both combined, with link
budget and a Wilson-Hilferty energy detector. Library. MODELLED (VM-473e1da1).kshana::locate_pd(B3a,kshana-locate/power-difference/v1, session-level shadowing in v2) andkshana::locate_tdoa(B3b):
fixed-station interference-source locators from power differences and from time differences of arrival, each with its synthetic
session generator and gate runner. Library only. MODELLED (VM-db8ed956,VM-09cb4375); both gates NOT MET.kshana::opnav(C5a): the optical-navigation measurement model, the apparent direction of a body centre (converged Newtonian
light time and stellar aberration) or a catalogue star, projected through a pinhole camera with Brown-Conrady distortion in the
OpenCVprojectPointsconvention, with the partials of the pixel with respect to position and velocity. MODELLED
(VM-3c8cdeb7); the camera values are illustrative.kshana::c3_force(C3): the A2 lunar force model with the Sun and Earth from a DE440 kernel plus Venus and the Jupiter system
barycentre (M1), and optionally cannonball radiation pressure with Earth and Moon conical shadows (M2); with no kernel and no
radiation pressure it is the A2 model bit for bit. MODELLED (VM-a074a1db).- The
tli-and-cruisephase's optionalimpulseslist (t_sfrom the phase start,dv_mpsin GCRF), and
mission_thread::impulses::detect, which flags hours whose one-hour velocity residual exceeds 0.01 m/s (C2). Without impulses
every output is unchanged. MODELLED (VM-22f8a601). BocTracking::exact_inverseandboc11_dll_inverse: the BOC(1,1) DLL read through the exact inverse of each discriminator's
response on the ideal peak (closed forms inside and outside |ε| = d/2, clamped to the main peak). Opt-in; the linear scaling
stays the default. Scored under D8a row 5 (met on re-run).[campaign_fit] cn0_lag_s(D8a row 6) and[campaign_fit] cn0_model = "estimator-response"withcn0_windows,
cn0_window_sandcoherent_s(D8a row 6 v2;cn0::estimator_reading): the lab fit models each C/N0 bin at the stated J/S a
lag earlier, or as the M2M4 estimator's own reading along the stated profile. Defaults unchanged. Both NOT MET.stats::m2m4_step_readingandstats::cusum_delay_through_m2m4(D8b row 8 v2): the noise-free M2M4 estimate of a C/N0 step and
the drop CUSUM's delay through it.examples/iono_h2h_klobuchar.rs: the broadcast Klobuchar model as an entry for a pre-registered ionosphere comparison
(iono_h2h/1: VTEC at nodes, STEC along given lines of sight; no F2-peak model).- The GMAT R2026a adapter and notebook (
integrations/flight-dynamics/gmat.ipynb), its own workflow job with
continue-on-error: GMAT'sSPKpropagator runs Kshana's kernel end to end but does not carry it (within 1.19 mm of the written
state, a GMAT time-representation finding); itsCCSDS-OEMpropagator refuses an OEM 2.0 outside GMAT's TESTING mode. - MCP: four tools,
interference_emitter_regions,connect_map,referee_location_scoreandephem_export(53 in all), each
with a round-trip test; the translated READMEs list all 53. docs/STABILITY.md: the public surface, versions, deprecation, the conditions for 1.0, and 12 months of
long-term support from 1.0.- Studio: the two
tip-and-cuescenarios and seven more join the library (clock-fitted-log,clock-measured-log,lunar-descent,
lunar-surface,lunar-ascent,lunar-orbit-rendezvous,lunar-orbit-rendezvous-elfo), with a Kshana Connect card in the tools
hub and the Connect, adapter, flight-dynamics and Station documentation pages.
Changed
pvt::select_ephemerisskips broadcast records marked unhealthy (GPS: health not 0; QZSS: any bit but the L1C/A-only flag), and
a satellite with no healthy record withinMAX_EPH_AGE_Sis left out of the epoch. On by default, with no option. One pinned
expectation moved: a constellation built fromBRDC00WRD_R_20181330000_01D_GN.rnxnow has 31 satellites, not 32 (G04's only
record has health 63). Found by B5 v3 (docs/design/evidence/b5-area-trust/ENGINE-FINDING-HEALTH.md).- The live decoder window is 256 frames, not 4096, which makes the live decoder about 12 times faster (measured by
kshana bench-speed). - The planner and the environmental baseline use the engine's portable maths, so every target gives the same bytes. The first
WebAssembly check of the planner had differed from the command line; the scored outcomes are unchanged (B1a P1 to P4 identical,
B8 scores within 2.4e-16, no level changed). - The
labfit-iqreport omitssettings.cn0_windows,cn0_window_sandcoherent_sat their defaults (50, 0.02 s, 0.001 s),
as it omitscn0_model, so stated-mode reports keep their bytes; the field-units schema gains four rows. - The published crate stays under crates.io's 10 MiB (10 402 479 bytes):
apps/,deploy/,skills/, the Connect check-vector
folder andREADME.mdare left out; the library embeds none of them, and the bench manifest thatsrc/bench/texbat.rsembeds is
packaged. deny.tomlrecords RUSTSEC-2026-0457 and RUSTSEC-2026-0462, both in rmcp client code (OAuth and the Streamable HTTP client
transports) that the server-onlykshana-mcpdoes not compile.- Release workflows: only the newest version tag may publish; the MCP registry step waits until npm and PyPI serve the version;
every remaining tolerated failure states its reason; a winget pull request already open for the version counts as success; the
OCI image builds get 180 minutes. A monthly Apple signing check fails when a Developer ID certificate has fewer than 30 days
left. - The Studio's 184 recorded runs are gzip-compressed (
studio/recorded/<name>.json.gz): 43 MB of JSON becomes 6.9 MB, and the
deployed site 498.1 MB becomes 458.1 MB.
Fixed
- An SBF log cut inside its last block, or with a stray
$@sync whose declared length runs past the end, no longer makes
kshana obs convertand the receiver-trust SBF reader loop forever: the blocks are framed by Kshana and the unfinished one
counts as one skipped record (tests/receiver_trust_sbf.rs). - The RINEX navigation reader takes each record's length from the file, so RINEX 3.05 five-line GLONASS records no longer lose
the records after them. The BKG merged daily files of 2025-01-08, 14 and 15 mix 4- and 5-line GLONASS records, and the fixed
count had dropped most GPS ephemerides on those days. Found by B5 T4; the B5 T4 scored result stands, and the same run after
the fix is recorded beside it (still NOT MET). - The
fusion-pnttiming bound, in both the decoupled and the[coupled]suites, adds the flicker phase variance accumulated
since the last anchor (defect D11c, under D17). Exactly zero without a flicker floor;fusion-pntandfusion-pnt-coupledare
byte for byte as before.
Revisions
- Mars short-arc orbit determination, run 2: since portable maths (0.36), the fit takes 11 Jacobian evaluations, B3 is 0.359 km
and B4 is 0.523 km (were 12, 0.362 and 0.527). Every verdict is unchanged; B2 stays NOT MET and the row stays MODELLED. The
data-gated test was missed by the 0.36 re-baseline. Seedocs/revisions/CAP30-run2-portable-math-rebaseline.md. - Version-bearing pins re-blessed at 0.41.0 (21 telemetry hashes, 38 skyline rows, two
iq_campaigndigests, three README demo
charts, the SBOM fixture hash):docs/revisions/STUDIO15-pins-0.41.0-version.md. No claim changed.
Evidence
Each scored row below was pre-registered and signed off before its code or data, and scored once; no bar was changed after a
result was seen. C7, D1 and D18 have no bar, and the R items are not scored. Disclosures (amendments, development runs,
restarts) are in each results file. Everything here stays MODELLED.
- M1 (space RF geolocation bounds;
docs/design/evidence/m1-space-rf-geolocation/RESULTS.md): MET. B1 against an
independent recomputation over 60 cases, worst relative difference 3.0e-10 (bar 1e-6); B2 reproduces Cao et al. 2015, Fig. 1
(25 plot readings) within +5.5 % (bar 20 %); B3 RMS over DRMS 0.979 to 1.019 (bar 0.9 to 1.25), 1000 of 1000 converged; B4 and
B5 met. The bounds are optimistic: no clock or orbit-knowledge error, representative link budgets. - M2 (revisit and tip-and-cue latency, a cross-check of published simulation figures;
docs/design/evidence/m2-revisit-tip-and-cue/RESULTS.md): not consistent with the published figures. M2-B0 MET (1 283
access intervals, worst difference 0 s); M2-B1 one of six figures consistent (16-satellite revisit 23.97 min against 20 min,
window 15 to 25); M2-B2 none of the three 16-to-32 ratios consistent (1.046, 1.146 and 1.778 against 2.00, 2.17 and 2.14). - M4 (dark vessels and AIS integrity, synthetic;
docs/design/evidence/m4-dark-vessel-ais-integrity/): v1 NOT MET: B2
dark recall 0.848 (bar 0.95), B4 offset recall 0.807 at 5 km (bar 0.85) and 0.846, 0.825, 0.871 at 10, 20, 50 km (bar 0.98);
B1 (FA 0.0050), B3, B5, B6, B7 met. v2 NOT MET: B2 0.877, B4 0.396 at 5 km and 0.689, 0.650, 0.646; B1 (FA 0.00006), B3,
B5, B6, B7 met. v3 NOT MET on B4: 0.489 at 5 km (137 of 280) and 0.646, 0.675, 0.661; B2 MET at 0.994 (1 392 of 1 400,
precision 0.867); B1 (FA 0.00075), B3, B5, B6, B7 met. - M5 (interference sources from orbit and a space-plus-ground hybrid, synthetic;
docs/design/evidence/m5-jammers-from-orbit-hybrid/): v1 NOT MET on B1 to B3 (success: space 0.630, ground 0.000, hybrid
0.630 against 0.680; medians hybrid 56.08 m, space 56.07 m against 44.86 m; p95 hybrid 359 km against 263 km); B4, B5, B6 met.
v2 NOT MET on B1 to B3 (success: space 0.990, ground 0.00025, hybrid 0.990 against 1.040; medians 7.352 m and 7.351 m
against 5.881 m; p95 71.300 m against 71.297 m, missed by 3.5 mm); B4, B5, B6 met. The hybrid equals the space layer. - B1a (monitoring-network planner, part 1;
docs/design/evidence/b1a-planner/): v1 NOT MET on P1 (4 of 619 line-layout
cells on the degenerate axis miss 1e-9 relative, by 1.3e-9 to 2.6e-9), P2 (52 of 54 cells within ±5 %) and P4's greedy part
(0.2754 against the best of 1 000 random subsets, 0.2790); P3, P4 exhaustive, P5 (Linux, macOS, Windows, command line, Python, WebAssembly) and P6
met. v2 NOT MET on P4 (greedy+exchange 0.350982 against the best of 1 000 random subsets, 0.353361); P1, P2 (46 of 46 in-regime cells), P3,
P6 met, P5 met on Linux. v3 MET on Linux: P4 met (greedy+exchange2+restarts 0.440809 against the best of 1 000 random subsets, 0.410470 on 48
candidates,m= 8); P1 to P3 and P6 met; P5 met on Linux, its macOS and Windows half pending, and the v3 placement's
consistency check did not complete (the container restarted; the printed values are the scored ones). - B2 (location metrics in the referee;
docs/design/evidence/b2-referee-metrics/RESULTS.md): R1 to R5 MET. R1 0 misses
over 10 054 reports, worst distance difference 2.5e-9 m (bar 1e-6 m); R3 coverage 0.9455 at 95 %, 0.8963 at 90 %, 0.4961 at
50 % (n = 10 000); R4 on the command line, Python, WebAssembly and MCP (MCP compared value for value). This checks the metric,
not any locator. - B3a (power-difference locator, synthetic gate;
docs/design/evidence/b3a-power-difference-locator/): v1 NOT MET: S1
coverage 0.976 met, S2 median 95 % region 52.5 % of the domain (bar 25 %), S3 centroid 2.55 × the station spacing (bar 0.5 ×);
S4 not run. v2 (session-level shadowing) NOT MET: S1 0.890 (bar 0.90), S2 26.1 %, S3 1.45 ×; S4 platforms only, the
command line, Python and WebAssembly half not assessable (no such surface). The row stops; no real run is made with this
method. - B3b (TDOA locator, synthetic gate;
docs/design/evidence/b3b-tdoa-locator/RESULTS.md): NOT MET. T1 coverage 0.044
(bar 0.90), T2 median area share 1.000 (bar 0.25), T3 centroid 3.35 × the spacing (bar 0.5); 872 of 1 000 sessions not
located (the χ² consistency test failed). T4 met on Linux (debug equals release); macOS and Windows recorded when the
b3b-determinismjob has run. The independent oracle agrees on 19 of 20 sessions and on the scored outputs of the twentieth.
New ledger rowVM-09cb4375. The row stops; no real run is made with this method. - B4a (benchmark task (e);
docs/design/evidence/b4a-benchmark-classes/): E1 and E2 MET, so task (e) ships. The B-log
baseline's bars are NOT MET: E3 not scoreable (one jamming folder intest), E4 every spoofing epoch labelledjamming
(share 1.00 against 0.50), E6 balanced accuracy 0.259 against 0.80; v1's E5 not assessable (no TEXBAT file). On all ten TEXBAT
recordings (1 s windows, scored once on 11 November 2026, recomputed independently to 1e-12): E5 v2 NOT MET, AUC 0.6941
against 0.70 (2 729 spoofing and 922 nominal windows); E5-PD MET, AUC 0.7410 (2 729 and 802 windows). Both miss the
matched-power attacks (ds1, ds3, ds6). The detectors stay MODELLED (VM-bd8626f0). - B5 (area trust;
docs/design/evidence/b5-area-trust/): v1 NOT MET on T4: 23 alerts toDEGRADEDover 13 area-days in
Madrid (1.77 per area-day; bar 1), none toUNTRUSTED; T1, T2, T3 (p95 1.004 s, maximum 1.015 s, 220 events at 100 stations)
and T5 met. v2 NOT MET on T4: 16 alerts toDEGRADEDor worse over 13 area-days at Goddard (1.23; bar 1), 9 of them to
UNTRUSTED(0.69; bar 0.1); T1, T2, T3 (p95 1.004 s, maximum 1.013 s) and T5 met. v3 NOT MET on T2: T1 MET (520 epochs, 125
flagged, equal to the independent oracle through the library and the command line); T2 NOT MET (4 of 127 single-satellite
faults not flagged; 2 of 67 two-satellite epochs flagged with one satellite); T3 MET (220 events, webhook and MQTT p95 1.003 s and 1.004 s, maximum 1.006 s); T4 MET (1 alert toDEGRADEDor
worse and 1 toUNTRUSTEDover 26 area-days, 0.038 each; bars 1 and 0.1; Goddard 0, Wettzell 1, 17 to 30 May 2026); T5 met on
the command line (all four documents; Python and WebAssembly check the same documents in CI). - B8 (environmental baseline;
docs/design/evidence/b8-environmental-baseline/RESULTS.md): MET. N1 1 295 disturbed
station-hours from all three bands; V1 AUC 0.846 (bar 0.75); V2 100 % of 1 422 quiet mid-latitude station-hoursquiet(bar
95 %); V3 99.4 % of 621 disturbed storm high-latitude station-hourselevatedordisturbed(bar 80 %); V4 AUC 0.872 at low
latitude (bar 0.70); V5 the same bytes on Linux, macOS, Windows, the command line, Python and WebAssembly. - C2 (the Artemis I cruise with impulsive steps modelled;
xval/mission-thread-c2/RESULTS.md): NOT MET on W1: 17.3 km
against 2 km, and 0.0177 m/s against 0.01 m/s; W2 not run (the published burn times leave 42.8 h, not 72 h). Four of the five
steps in the reconstruction carry position jumps no velocity step within an hour can produce. C2 v2 was not run: the fresh window
its registered rule names does not exist (the longest gap after W1 is 105.9 h against 120 h). C1 stays NOT MET. - C3 (the CAPSTONE NRHO with a corrected force model;
docs/design/evidence/c3-capstone-nrho-force-model/RESULTS.md): C3-R
MET on W0 (21.66 km, 4.24 km); C3-H NOT MET on E, H1 and H2 (37.25, 36.23, 49.65 km end position against 25 km); C3-S
NOT MET (H1 12.92 km met, H2 57.13 km). A2 is therefore not recorded as met on re-run and stays NOT MET. C3-S v2 (akper
window, fresh span G) MET: end position 0.21, 0.60, 0.73 km and periapsis radius 11, 4, 87 m at +156 h; G1 is
low-information by the registered sensitivity rule (18.1 km) and its verdict counts. M1 alone also meets both bars on G, so G is
a less demanding span; H2 is still missed with its ownk(57.8 km, no bar). - C5a (the optical-navigation measurement model;
docs/design/evidence/c5a-optical-navigation-model/RESULTS.md): v1 NOT
MET on O5 (258 of 432 partial-derivative elements fail against central differences at the registered steps); O1 (0.0016 mas),
O2, O3, O4 and O6 met. v2 MET: the partials agree with an independent complex-step oracle to 2.3e-11 of each row's largest
element on all 72 measurements; O1 to O4 met on 36 fresh measurements. v1's O5 stays NOT MET as recorded. - C7 (flight-dynamics adapters; no bar;
docs/design/evidence/c7-flight-dynamics-adapters/RESULTS.md): Orekit 13.1.9,
Basilisk 2.12.2 and ANISE 0.10.6 carry the files; GMAT R2026a does not carry the SPK (1.19 mm, a time-representation finding)
and refuses the OEM. Kshana's type 13 SPK is byte for byte the kernel the SPICE Toolkit'sspkw13writes from the same states
(245,760 bytes). - D1 (the Artemis I and CAPSTONE gaps;
docs/design/evidence/d1-artemis-capstone/RESULTS.md): done (both registered re-runs
made once and recorded). Neither original row (C1, A2) is met. - D8a (IQ tracking and C/N0 re-runs;
docs/design/evidence/d8a-iq-tracking-cn0-rerun/RESULTS.md): row 3 (VM-f5e2ce4e,
the CFAR reference set over the whole row) MET on re-run (C4: 0, 3, 4, 0, 0 of 2495 off-peak cells above, bar 1 %; C2 and
C3 0.969 to 1.033 %). Row 4 (VM-bdfec2e4, the seeded hand-over) NOT MET: at 33 dB-Hz 20 of 20 hold on seen and fresh
seeds with σ ratios 1.003 and 1.006, but at 31 dB-Hz 15 and 6 of 20 hold (D2) and σ ratios are 1.815 and 2.815 (D3). Row 5
(VM-35b27a4f, the exact BOC inverse) MET on re-run (sampled-signal slopes 0.9947 to 1.0106). Row 6 (VM-c9318c63, the
lag correction) NOT MET: (a′) +0.5015 dB seen against 0.5 dB, (b′) 0 of 50 on both sets, (c) and (d) one set each. Row 6 v2
(the estimator's own response) NOT MET: (a′) met on all three sets (+0.165, +0.098, +0.099 dB), (b′) 36, 32 and 27 of 50
(bar 45), (c) missed on v1-fresh (time to loss 1.000), (d) missed on v1-registered (2.195 dB). - D8b (IQ timing and campaign re-runs;
docs/design/evidence/d8b-iq-timing-campaign-rerun/RESULTS.md): row 7
(VM-e6a6348d, acquisition hysteresis on a fresh power ladder) MET on re-run. Row 8 (VM-59bb0091, M3 at a one-window
stride) NOT MET on M3's timing clause: 18 of 18 detected, median time to detect 0.221 s and 0.201 s against 0.0068 s and
0.0064 s (was 1.001 s against 0.433 s). Row 8 v2 (the bound with the estimator's delay) NOT MET on every seed set: 0.2208 s,
0.2008 s and 0.2207 s against median bounds of 0.0868 s, 0.0864 s and 0.0660 s; the detection clause and M4 hold. Row 9
(VM-6254a75d, the window restart) R1 MET on re-run:cn0_dropAUC 1.0000 on seen and fresh seeds. - D17 (
docs/design/evidence/d17-defect-fixes/RESULTS.md): D11c criteria 1 (integrity mean 0.99988 / 0.99989 for the
classical clock, was 0.77101 / 0.75865), 2 and 4 MET; criterion 3 NOT MET (pooled RMS of error over the bound 0.768 /
0.783 classical and 0.590 / 0.592 quantum, band 0.8 to 1.2): the bound is conservative, opened as a new defect. D12 closed
(grid-free acquisition). - D18 (no bar;
docs/design/evidence/d18-bench-leftovers/RESULTS.md): item 1, the almanac-HDOP scorer, built and passing its
acceptance tests (HDOP within 1.6e-15 of numpy at 20 points); item 2, the bench checks external files against pinned hashes, and
the ten TEXBAT recordings were pinned from the owner's fetches on 9 and 11 November 2026. - R items: R1.1 (RTCM SSR) has no pre-registration and no bar; it is MODELLED (
docs/design/evidence/r1-1-rtcm-ssr/RESULTS.md).
R1.4 (tides) has no scored comparison: its pre-registration is proposed and awaits sign-off. The R6.1 CGGTTS comparison is
pre-registered as proposed and not run. The other R pre-registrations committed in this release (R1.2, R1.3, R3.2 to R3.5,
R4.3, R5.1, R6.2, R7.4) are proposed and not run.
Get this release
Download — attached below, prebuilt (no toolchain needed); each artifact carries
SLSA build-provenance (verify with gh attestation verify <file> --repo AshfordeOU/kshana):
kshana— the simulator CLI / engine (Linux x86-64)kshana-aarch64-apple-darwin·kshana-x86_64-apple-darwin— the same CLI for
macOS on Apple silicon and on Intel (releases after 0.27.2)kshana-x86_64-pc-windows-msvc.exe— the same CLI for Windows x86-64 (releases after 0.27.2)kshana-mcp— the Model Context Protocol server (Linux x86-64)kshana-sbom.cdx.json— CycloneDX SBOMkshana-validation-summary.html— the per-release validation summarykshana-channels-sbom.cdx.json— CycloneDX SBOM for the OpenCPN plugin, the Signal K plugin,
the reference-build image and the Grafana dashboard (the engine's crates are inkshana-sbom.cdx.json)kshana_pi-0.41.0-1_ubuntu-wx32-24.04-x86_64.tar.gz·kshana_pi-0.41.0-ubuntu-wx32-x86_64-24.04.xml—
the OpenCPN trust-panel plugin (GPL-3.0-or-later; Ubuntu 24.04, wxWidgets 3.2, x86-64) and its plugin-catalogue
metadata;kshana-opencpn-plugin-0.41.0-source.tar.gzis its source. Install the tarball with
OpenCPN's Import Plugin button. Not yet in OpenCPN's own catalogue.signalk-kshana-trust-0.41.0.tgz— the Signal K server plugin, the same file as the npm packagekshana-grafana-*.json— the Grafana dashboard for the GNSS trust telemetrySHA256SUMS— a SHA-256 (Secure Hash Algorithm) checksum of every file above, so a
download can be checked withsha256sum -c SHA256SUMS --ignore-missing(releases after 0.27.2)
The PyPI wheels, the npm/WASM package and the Docker image below are cross-platform too.
Install from a package registry:
| Channel | Get it |
|---|---|
| crates.io | cargo install kshana · cargo install kshana-mcp |
| PyPI | pip install kshana |
| npm | npm install kshana |
| ghcr.io | docker run -i ghcr.io/ashfordeou/kshana-mcp:0.41.0 |
| npm | signalk-kshana-trust — the Signal K plugin (Signal K app store, keyword signalk-node-server-plugin) |
| ghcr.io | docker pull ghcr.io/ashfordeou/kshana-reference-build:0.41.0 — the trust gate service image (amd64, arm64) |
| MCP registry | io.github.ashfordeOU/kshana-mcp (auto-discovered by MCP clients) |
| JetBrains Marketplace | search "Kshana" in your IDE → Plugins |
No install: run it in your browser at kshana.dev · Cite: DOI 10.5281/zenodo.20528627
Full changelog: CHANGELOG.md · Docs: README