time walk

A slow time spoof. GPS is dragged away a few milliseconds per second — gently enough that no single reading looks wrong. Watch the gap open between the GPS trace and the consensus band, and note which detector catches it: the step threshold never trips, but CUSUM accumulates the drift until it does.

ALARM

last agreed 157s before the end

glonass_cn0_spread: OKgps_cn0_spread: OKgps_position_error: OKgps_time_offset: ALARMntp_time_offset: OKroughtime_time_offset: OK

What each chart shows

Every formula below is the one the code runs, with its source location. Thresholds are not repeated here — they come from the daemon's own published settings.

`gps_time_offset`

Variable: gps_time_offset — seconds
Source: The time field of each $GPRMC sentence, against the receiving host's own clock.
Description: How far GPS-reported UTC sits from local time. A time spoof drags this away while the independent references stay put — which is the disagreement the monitor exists to notice.
Formula: gps_time_offset = gps_utc − local_unix_time (collectors/gps.py:54)

`glonass_cn0_spread`

Variable: <constellation>_cn0_spread — dB-Hz
Source: The per-satellite carrier-to-noise values in each $GPGSV cycle, one channel per constellation.
Description: Spread of signal strength across the satellites in view. A real sky fades with elevation, so healthy spread is wide; a simulator transmitting every channel at equal power collapses it toward zero.
Formula: cn0_spread = pstdev(per-satellite C/N₀) (collectors/cn0.py:106)

`gps_cn0_spread`

Variable: <constellation>_cn0_spread — dB-Hz
Source: The per-satellite carrier-to-noise values in each $GPGSV cycle, one channel per constellation.
Description: Spread of signal strength across the satellites in view. A real sky fades with elevation, so healthy spread is wide; a simulator transmitting every channel at equal power collapses it toward zero.
Formula: cn0_spread = pstdev(per-satellite C/N₀) (collectors/cn0.py:106)

`gps_position_error`

Variable: gps_position_error — metres
Source: The latitude/longitude of each $GPRMC fix, against the site's surveyed antenna position.
Description: Distance between where the receiver says it is and where it actually is. The antenna does not move, so anything but noise here is the fix being wrong.
Formula: gps_position_error = haversine_m(fix, surveyed_site) (collectors/gps.py:55, geo.py:10)

`ntp_time_offset`

Variable: ntp_time_offset — seconds
Source: An SNTP exchange with one or more network time servers (RFC 4330). With two or more, a median consensus corroborated by a quorum.
Description: Network time minus local time — a reference that does not come from the sky. A GPS spoofer would also have to own the network to move this.
Formula: offset = ((t2 − t1) + (t3 − t4)) / 2 (collectors/ntp.py:153)

`roughtime_time_offset`

Variable: roughtime_time_offset — seconds
Source: A Roughtime server's signed response, verified with Ed25519 and a Merkle inclusion proof for our nonce.
Description: Cryptographically verified time minus local time. An attacker cannot forge this without the server's private key, which is what separates it from plain NTP.
Formula: roughtime_time_offset = verified_midpoint − local_unix_time (collectors/roughtime.py)

How every channel is judged

The same three detectors run on every channel above. Thresholds are per channel and are the ones the daemon actually applied — the dotted bands on each chart are drawn from them.

Detector	Rule	Source
step	alarm when abs(x − target) > threshold	discrepancy/step.py:31
cusum	S⁺ = max(0, S⁺ + (x − target − k)); S⁻ = max(0, S⁻ + (target − x − k)); alarm when max(S⁺, S⁻) > h, then reset	discrepancy/cusum.py:55-58
staleness	alarm when now − last_sample.timestamp > max_sample_age_s	discrepancy/engine.py:126