Provenance

Every measurement on this site was produced offline, validated against a schema before it was written, and committed. The console reads those files and renders them. It runs no model, calls no service and holds no credentials, so there is nothing between the number in the receipt and the number on the page.

The snapshot

Snapshot

snap-20260817-stage1

frozen before any model was fitted

Queue built

2026-08-22 04:07

seed 42, fixed

Receipts

11.0 MB, all committed

Waterfalls shipped

of 25 requested, 0 degraded

Splits

Four ways of holding data back, because a chronological split alone cannot tell you whether the model learned the task or learned the stations.

One scale across all four rows, so a split that keeps fewer observations is drawn shorter. The number at the right of each row is that split's total, and every total is the same corpus.

All four are defined in SPLIT_MANIFEST.json, digest c0f8cd3a5dd5fbf9…
Split	Train	Calibration	Test	Excluded
chronological	1909	408	410	—
cold_station	2031	293	403	—
cold_transmitter	2235	139	353	—
cold_combined	945	110	183	1489

The cold combined split excludes rather than assigns the observations that would break its own guarantee. An observation whose station is held out but whose transmitter is not cannot sit in a partition that promises both are cold, so it sits in neither, and the count is stated rather than absorbed.

Receipts

One per stage, each with the digest of the file as it stood when this payload was built.

File	SHA-256	Bytes	As of
AGENT_RECEIPT.json	7927ca92bfb7f3ad…	29,230	this build
ATTRIBUTION_AUDIT.json	661213c8fed4b152…	85,390	the build before the sign-off
AXIS_READER_AUDIT.json	01592786df3cb33c…	15,640	this build
BASELINE_RECEIPT.json	f2796822ec53b590…	8,803	this build
CIRCULARITY_RECEIPT.json	8e6f893a005e3bcc…	11,376	this build
CLEAN_CLONE_TRANSCRIPT.json	312237a9855220eb…	14,794	this build
DATASET_MANIFEST.json	7029e2a8b3bef5fd…	2,309,369	this build
EXPLAIN_RECEIPT.json	49f9a2f902c2b1fb…	24,954	this build
FONT_PAINT_RECEIPT.json	cf9a8a36448599c1…	6,377	this build
FUSION_RECEIPT.json	57c379eff2f2ffdb…	270,666	this build
GATE3_POOL.json	dc6529a2eb4a22ff…	3,017,898	this build
GATE3_POOL_A_RECEIPT.json	80d1fa57f0cc3276…	2,064,646	this build
GATE3_RECEIPT.json	fd52c0dbf08ef188…	2,048,382	this build
GATE3_SENSITIVITY.json	eac05c8bfb49a420…	2,725	this build
GATE4_BUNDLE.json	4a4411de5b4ea76c…	1,509	this build
GATE4_RECEIPT.json	16ca34b7ed81bd41…	74,928	this build
GATE4_WORKSHEET.json	44cc11bd27d624f1…	11,481	this build
GATE_POWER_RECEIPT.json	5912a7c8cb50b329…	11,333	this build
HERO_NULLS.json	6c5ed7b02bf449a1…	24,728	this build
LANGCHAIN_RECEIPT.json	043c8239286dee16…	4,069	this build
LANGFLOW_RECEIPT.json	71cc7dfe1826e3cb…	3,018	this build
LEAKAGE_AUDIT.json	ecb4aff23de89169…	10,942	this build
LIVE_SHELF_RECEIPT.json	26c7c431441925aa…	2,653	this build
MOTION_AA_CONTROL.json	144edd40ad9380f9…	11,217	this build
MOTION_RECEIPT.json	7848414e2567859d…	13,006	this build
NETWORK_SCALE.json	5dcb941cf37d8d1f…	3,153	this build
OFFLINE_RECEIPT.json	e869ab82c1b7bdae…	3,727	this build
OPERATOR_SESSION.json	69564bde019b6783…	6,829	this build
PHYSICS_VALIDATION.json	6dc5302b24363528…	63,045	this build
PRECEDENT_RECEIPT.json	7cc6b2fde7059111…	9,549	this build
QUEUE_RECEIPT.json	4204193634349fa8…	259,290	this build
REPO_WEIGHT.json	b0a80699b259dd2d…	4,927	the build before the sign-off
SATELLITE_NAMES.json	196ad682c56350b8…	12,328	this build
SECOND_TRACE_SURVEY.json	ef43f2ad5bc160be…	334,073	this build
SECRET_SCAN.json	c9b9f0296c922f16…	1,053	the build before the sign-off
SIGNOFF_RECEIPT.json	dc2280012b556543…	3,829	the build before the sign-off
SPLIT_MANIFEST.json	c0f8cd3a5dd5fbf9…	223,247	this build
THROUGHPUT_RECEIPT.json	8434514de99fd625…	3,112	this build
TRIAGE_RECEIPT.json	5884eb5dc9b6d1a7…	3,775	this build
WATSONX_RECEIPT.json	86c9fbc9266c249c…	2,108	this build
corridor_features.json	748ac151f0873df0…	527,163	this build

Every digest here is of the file as it stood when this payload was built. Four are marked, because scripts/signoff.py rewrites them after the console is built: it re-runs the release audit and then writes its own receipt, and each of those records the commit it ran at. A commit cannot record its own hash, so those four digests are one generation behind by construction and are marked rather than quietly wrong. Every unmarked digest is checked against the file on every gate run by tests/test_published_digests.py.

The queue and evaluation pages read QUEUE_RECEIPT.json (4204193634349fa8…) and FUSION_RECEIPT.json (57c379eff2f2ffdb…).

Contracts

A schema is ratified before the script that writes against it runs, so a receipt that violates its contract never reaches disk.

Contract	Version	Status	SHA-256
annotation_record.schema.json	0.1.0	ratified	0f25ccfb0faca625…
dataset_manifest.schema.json	0.2.1	ratified	13e1805625ea0f28…
fusion_receipt.schema.json	0.1.0	ratified	6f1da46ce3f05513…
queue_receipt.schema.json	0.3.0	ratified	0a5cdbdc7aa59ac5…
source_observation.schema.json	0.2.1	ratified	851c4db620bc3bce…
split_manifest.schema.json	0.5.0	ratified	daca6d6992ea30b0…
triage_receipt.schema.json	0.3.0	ratified	5d491ab7fe86a9b2…
waterfall_geometry.schema.json	0.2.2	ratified	1ac826b1a1610843…

The documents these numbers were promised against

Each was written before the measurement it governs, and each is served here as the file itself rather than as a summary of it.

Document	What it fixes
KILL_GATE.md	The gates, their wording, and what each one would have killed. Written before the pipeline that answers them.
CLAIM_REGISTER.md	Every claim this project makes, with the evidence behind it and the claims that were withdrawn.
C2_PREREGISTRATION.md	The concentration caps, the grouping keys and the decision rule, all committed before the numbers on this site were computed.

What it costs to run

One core, timed over a real batch rather than a benchmark loop.

One observation

1.26 s

at the slowest measured stage

One core, one day

68,702

observations, at that rate

The network, one day

6,380

captures with a waterfall, over 9.4 hours

Headroom

10.77×

0.093 of a core keeps up

One core processes 68702 observations a day at the dominant measured stage, against 6380 the network produced in the span this snapshot covers. That is 10.77 times the load on a single core, so the pipeline is not compute-bound at network scale.

Both stages were timed over the same batch. The slower one sets the rate above, because a pipeline runs no faster than its slowest stage.
Stage	Observations	Wall clock	Per observation	One core, one day
corridor fit	743	934 s	1.26 s	68,702
second trace survey	743	867 s	1.17 s	74,077

Ingestion costs more per observation than the corridor fit does, and it is bound by a courtesy interval and a 1.7 MB image download rather than by anything this project computes.

The deployment that removes it is the obvious one: run at the station, where the waterfall is already local and there is no API to be polite to.

The capture span is 9.4 hours inside one day. A day rate from it is one observation of the network's rate, not a long-run average.elapsed_s covers the corridor fit and the second-trace survey only. SGP4, the fusion forward pass and the queue sort are cheaper per observation and are not in it, and Granite is not per observation at all.Both stages were timed on one machine, single-threaded. The core count is a division and not a measured parallel speed-up.No figure here is a claim about latency. The queue is a batch reading order, and nothing in this project needs to answer inside a pass.

What this console does not do

Stated positively, because an absence is easy to claim and hard to notice.

No request for data to any origin but its own, before or after load. Two licensed display faces are the one exception and they carry no data; the content security policy in vercel.json names the only hosts a browser is permitted to reach, and every digit of every measurement is set in IBM Plex served from this site: artifacts/FONT_PAINT_RECEIPT.json.
No model runs in the browser. The probabilities shown were fitted offline on the training partition of the split named beside them.
No number on any page is computed by the console, except the map from stored intensities to screen colours, which the waterfall viewer states on every card.
No analytics about you, no cookies, no storage, and nothing to consent to.

Degraded states

What the console shows when something is missing, and whether any observation it ships actually puts the console into that state.

Every count is zero, which is worth saying rather than leaving in a table: the observations this console ships are the top of a queue, so they have the cleanest geometry. To see a degraded path here, block the waterfall image and the page will say what it lost.

Every state, and the count in this corpus

Counted over the 25 observations with imagery. A zero means the path is covered by the offline suite and by a forced check, not by this corpus.
When	What the console shows	Shipped cards in this state
The observation is not in the snapshot	The card page says so and links back to the queue. No blank frame, no zeroes.	0
No frequency information, so no centre pixel	The waterfall renders and the corridor overlay is withheld, with the reason and the share of records it affects.	0
The TLE will not propagate, so there is no pass geometry	The same withheld overlay, carrying the physics module's own degraded reason rather than a generic one.	0
The corridor fit ran into the edge of its search range	The offset is shown with a greater-or-equal marker and a note that it is a lower bound. The observation is excluded from the stale-catalogue conflict criterion.	0
The browser has no WebGL2, or loses the context	The same image as a plain img, the reason in a note, and the contrast controls removed rather than left dead.	not from data
The waterfall image will not decode	The same note, naming the decode as the cause.	not from data
JavaScript is off	The waterfall, the corridor overlay and the top of the queue all still render. The filter controls are hidden and a line says why.	not from data
A filter matches nothing	A stated empty result with the full queue size beside it, so an empty table cannot be read as a broken one.	not from data

Data and attribution

Waterfall imagery from the SatNOGS Network, contributed by volunteer ground stations, under CC BY-SA 4.0. See DATA_LICENSE.md.

Waterfall imagery is the unweighted mean of the R, G and B channels, which is the quantity the corridor fit operates on. A SatNOGS waterfall is false coloured, so its luma is not its intensity and two different intensities can share a luma.