Veldris
Products · the realm

One enclosure. Six products.

Every Veldris product runs inside the Midgard enclosure and inherits its guarantees: post-quantum sealing, signed lineage, local inference, zero required outbound calls. Each is independently useful. Together they compound.

Try every console in your browser →

§ 1 · Midgard Core

The enclosure

Midgard Core is the enclosure layer: identity, the object fabric, storage, compute, governance and post-quantum cryptography inside one defended trust boundary. It is the part our competitors call the substrate, and the part they will not show you.

Explore the platform →

§ 2 · The six

Standing on it

A first deployment is Midgard Core with Mimir and Forseti: AI on sensitive data, with the evidence to prove what it did. The rest expand from there.

§ 3 · In motion

The software, recorded

Screen recordings of Forge, Mimir, Heimdall and the Midgard console, captured in October 2026 from development builds of the public source, grouped by product, newest first within each: the code behind them is in the midgard-app repository and anyone can build it. All are silent and served from veldris.io, with no third-party player. Counts, tracks, alerts, checks and audit entries are those on screen at the time of recording.

Try it yourself. The Midgard playground runs all seven consoles in your browser on sample data, with a guided tour: nothing to install, and nothing you do there leaves the page. Open the playground →

Midgard · the enclosure · 1 recording · about Midgard

Midgard · object fabric and governance · 29 Oct 2026 · 1:05 · no audio

One fabric, every record governed and signed

The recording opens on the Midgard console's object fabric: 22 object types holding 41,278 instances, with 100% of their lineage signed. Each type lists its properties, instance count and marking, from live air-quality sensors and aircraft to flood warnings, grid frequency, Heimdall alerts and operators. Airport is marked restricted; most types are internal.

Selecting Airport shows what the type allows. Its link types tie an airport to a region and an operator, and its actions (open or close an airport, set the alert level, issue an advisory, divert inbound traffic) each carry a human gate: none runs until a person authorises it. A search for Heathrow returns London Heathrow Airport, and the audit lineage beneath fills in live as the operator reads, every entry signed ML-DSA-65.

The operator then follows the links: a METAR weather report for Heathrow, reviewed through an authorised action; Heathrow's runway 09R/27L, where open or close is a human-gated action; and a Heimdall power-cut alert raised from the distribution network operator's report, whose triage is authorised and recorded, linked to the power cut it reports.

Governance closes the recording. Policies and markings lists 24 enforced policies across 4 marking levels, each shown as a declarative rule, and a dry-run evaluates a request against them, denying by default. The audit lineage holds every record from the session, from the operator's reads to Heimdall's record syncs, append-only and verified, with each entry chained to the one before it and exportable as a chain.

On screenWhere to look
Object types22 types · 41,278 instances · 100% lineage signed
Human gatesThe Action types list for Airport and Runway
Linked recordsWeather report, runway and power-cut panels on the right
Policies24 policies · 4 marking levels · dry-run
Audit lineageappend-only · ML-DSA-65 signed · chain export

Forge · data operations · 1 recording · about Forge

Forge · connectors, expectations and quarantine · 29 Oct 2026 · 1:23 · no audio

Forge: a new source in, a bad file stopped at the gate

The recording opens on Forge's pipeline view for aircraft: ADS-B positions for the UK and Ireland pulled over HTTPS from adsb.lol every 15 seconds, parsed into 19 typed columns, kept as a versioned dataset of 326 rows and projected into Midgard as 326 live aircraft objects. The side panel shows how it is protected: AES-256-GCM with ML-KEM-1024 at rest, X25519 with ML-KEM-768 in transit, and lineage signed with ML-DSA. Run now re-runs the pipeline and reports it unchanged: nothing has altered since the last sync, so nothing is synced again.

The connector catalogue lists 11 connectors, all healthy, and 0 outbound paths. Each one's trust boundary is spelled out: inbound pull only, no egress.

The operator then adds a source of their own: a CSV of 18 River Thames monitoring stations. Forge previews the file, proposes a type for every column, and asks for an object type name, a key column, the latitude and longitude, a schedule and a marking, which is required. One click creates the connector and runs it. The file is sealed in the worker, the new object type joins the ontology, and the run writes a sealed version and projects 18 station objects into Midgard over post-quantum mutual TLS, with nothing quarantined.

Next, a rule. In data health, 24 checks are passing. The operator adds an expectation that every station's latitude lies between -90 and 90, and marks it blocking: if it fails, the projection stops and Midgard keeps the last good version.

Then the file is replaced with one carrying a bad latitude. Forge still seals the new version, but the check fails on one row, that row is quarantined, and nothing is synced: the connector shows projections blocked, data health shows 25 checks with 1 failing at a 94.4% pass rate, and the quarantine preview names the row, Radcot Lock, with a latitude of 95.699764. Replacing the file with the correct one clears the block, and Forge reports the data unchanged from the version Midgard already holds.

The dataset browser closes the recording: 12 datasets, every one sealed on its main branch, with the station dataset's schema, its lineage from upload to 18 objects, and both versions kept. Time travel compares them by station reference: nothing added, nothing removed, one value changed, the latitude that was blocked.

On screenWhere to look
Pipelineadsb.lol → aircraft.csv → aircraft → aircraft_live
Trust boundaryinbound pull only · 0 outbound paths
New connectorUpload a CSV, in the connector catalogue
Blocking expectation-90 ≤ lat ≤ 90 · 1 row quarantined
Projection gateData health: 94.4% pass rate, projection blocked
Time travel+0 added · -0 removed · ~1 changed

Mimir · local agentic AI · 4 recordings · about Mimir

Mimir and Heimdall · conflict airspace · 29 Oct 2026 · 0:42 · no audio

Asking the picture: which airspace is closed because of conflict

In Mimir, the Picture analyst is an agent that answers questions about the live Heimdall picture: power cuts, floods, earthquakes, aircraft, ships and alerts. It runs on Granite 3.3 2B (Apache-2.0) inside the enclosure, on CPU, with no egress. For each question it is given up to 20 items from the picture that are most relevant to it, plus an area summary. It answers only from those items and cites them, and every cited id is checked against what it was given. It is advisory: it suggests, and never acts.

Asked which airspace is closed or restricted because of conflict right now, it lists the conflict-zone advisories in the picture, each with its reference: Ukraine, Sudan, the Russian Federation, Iran, Iraq, Lebanon, the Persian Gulf and Gulf of Oman, Jordan, Saudi Arabia, Afghanistan, Libya, Mali, Pakistan, Somalia, Syria and Yemen. All 13 citations are found in the picture.

The operator checks the answer on the map. In Heimdall, with the airspace and events layers on, the restricted volumes over the Middle East show where the advisories apply.

Session trace S-756 then shows the call end to end: the request over post-quantum mutual TLS from heimdall-advisor with the purpose “investigation”, a PERMIT decision written as signed ledger entry #756, a prefill of 3,323 prompt tokens, and 537 tokens generated, 18 seconds in all. The model plane closes the recording: one model served, its weights verified by digest and ML-DSA signature on load, and no egress.

On screenWhere to look
QuestionWhich airspace is closed or restricted because of conflict right now?
Citations13 of 13 found in the picture
SessionS-756 · ledger #756 · 537 tokens · 18 s
Modelgranite-3.3-2b-instruct-q4km · CPU · local
Mimir and Heimdall · A400M or C-130 · 29 Oct 2026 · 0:46 · no audio

Asking the picture: are any A400M or C-130 aircraft flying?

Asked whether any A400M or C-130 aircraft are flying right now, and where, the Picture analyst answers that C-130s are, naming five airborne by their object ids and a sixth in the vicinity. It names no A400M. All five citations are found in the picture.

The operator then looks one of them up. Heimdall's search finds 02B26B in the North American picture, listed as a military aircraft, and its recorded flight history opens beside the map.

Session trace S-730: signed ledger entry #730, a prefill of 3,072 prompt tokens, and 205 tokens generated, 9.93 seconds in all.

On screenWhere to look
QuestionAre any A400M or C-130 aircraft flying right now, and where?
Citations5 of 5 found in the picture
CheckHeimdall search for a cited aircraft, then its flight history
SessionS-730 · ledger #730 · 205 tokens · 9.93 s
Mimir and Heimdall · Gulf watch and Hormuz · 29 Oct 2026 · 0:46 · no audio

Asking the picture: is shipping through Hormuz normal?

Asked what has changed in the Gulf watch and whether shipping through Hormuz is normal, the Picture analyst reads the Gulf watch indicators. Ships moving through the Strait of Hormuz remain at their normal baseline; GPS interference around the Gulf is well above its historical average; and recent GDELT news mentions Hormuz and Gulf shipping without signs of abnormal activity. It concludes that shipping through Hormuz appears normal, citing three items, all found in the picture: the Hormuz ships indicator, the GNSS interference cells and the GDELT feed.

On the map, with the maritime and events layers on, a medium GPS-interference alert sits in the Gulf among the ship tracks.

Session trace S-662: signed ledger entry #662, a prefill of 1,037 prompt tokens, first token in 914 ms, and 287 tokens generated, 6.51 seconds in all.

On screenWhere to look
QuestionWhat has changed in the Gulf watch, and is shipping through Hormuz normal?
CitationsGW-HORMUZ-SHIPS · GW-GNSS-CELLS · GDELT-hormuz
Map checkMaritime and events layers, GPS interference medium
SessionS-662 · ledger #662 · 287 tokens · 6.51 s
Mimir and Heimdall · GPS interference · 29 Oct 2026 · 0:42 · no audio

Asking the picture: where is GPS interference affecting aircraft?

Asked where GPS jamming or interference is affecting aircraft right now, the Picture analyst lists the affected aircraft types (C130, C30J, C208, TUCR, B350, PC21 and F100) and grades the interference as high or medium. It cites 16 items, all found in the picture: four GNSS interference cells and the aircraft inside them.

On the map, with the air and events layers on, the interference alerts sit along the cells, among the aircraft they affect.

Session trace S-557: signed ledger entry #557, a prefill of 3,358 prompt tokens, and 340 tokens generated, 13 seconds in all. The model plane shows 24 calls served and none failed since it started, at 43.4 tokens a second on 10 CPU cores.

On screenWhere to look
QuestionWhere is GPS jamming or interference affecting aircraft right now?
Citations16 of 16 found in the picture
SessionS-557 · ledger #557 · 340 tokens · 13 s
Model plane24 calls · 0 failed · 43.4 tok/s · CPU

Heimdall · operational fusion · 3 recordings · about Heimdall

Heimdall · London traffic cameras · 28 Oct 2026 · 0:47 · no audio

London traffic: from the road network to one vehicle

The operating picture opens over the south of England with the land layer on, then closes in on London. The road network is coloured green, amber and red by traffic state, and the city's traffic cameras sit on it as objects in the picture.

Selecting the A12 Dyers Hill camera opens its record: the camera's details and a current still. Opening the clip brings up AI analysis of the footage. Every vehicle is boxed and labelled, and the panel beside it counts what is moving and what is stationary, which way the traffic flows, and how the counts have trended across recent results. People are pixelated.

The operator picks one car and follows it. At the next camera along the A12, Church Lane, Heimdall matches the same car as probable, with a confidence of 0.95, and the follow carries across both cameras.

The clip then opens in the enhancement studio, with a before-and-after split. Presets choose between fidelity, classical sharpening and AI enhancement, and each effect, from levels and contrast to upscaling and motion deblur, can be switched on or off and tuned on its own.

On screenWhere to look
Road trafficThe road network over London, coloured by traffic state
Camera recordA12 Dyers Hill, in the right-hand panel
AI analysismoving · stationary · flow · trend
FollowingCAR 6 · probable (0.95) · 2 cameras
EnhancementThe studio, with before and after side by side

Traffic camera imagery: Transport for London JamCams. Powered by TfL Open Data.

Heimdall · Gulf watch · 27 Oct 2026 · 0:47 · no audio

Gulf watch: indicators that explain themselves

The recording opens on a public social media post. Heimdall treats social channels as cueing signals: a prompt to look somewhere, which the operator then checks against the governed feeds in the picture. Behind it, the UK and Ireland carry weather and flood warnings pinned to the map alongside their validity windows. The operator follows the cue and moves the focus to the Gulf and brings up the air, events and land layers over satellite imagery.

The Gulf watch panel lists a set of indicators, each measured against a baseline Heimdall has learned for it: civil aircraft over Iraq is rated elevated, GPS interference around the Gulf is still learning, and shipping through the Strait of Hormuz reads normal. Selecting an indicator opens its record: the count now against the normal count and the window it was learned over, its sources, its first sighting and its observations. Nothing on the panel is an unexplained dot.

Asked what has changed in the Gulf watch, Mimir answers in plain language beside the map, from the governed objects in the picture and on local inference.

On screenWhere to look
Social cueThe post that opens the recording
Regional focusUK and Ireland, then the Gulf, from the Regions rail
Baseline indicatorselevated · learning · normal
Indicator recordKind, status, baseline window, sources, observations
Decision supportMimir, in the Live picture panel
Heimdall · 3D picture and airspace · 27 Oct 2026 · 0:18 · no audio

From the UK to Western Russia, in 3D

The operating picture tilts into 3D over the UK and Ireland, dense with live air, rail and maritime tracks, while the classification rail counts every object by type and the airborne, on-ground and median altitude figures update with it.

The operator then brings Western Russia into focus. Satellite imagery resolves under the cloud layer, tracks fill in from the newly relevant feeds, and restricted, prohibited and danger airspace rises from the map as 3D volumes, so a track's position against a boundary reads at a glance. The fused feed count in the header climbs from 15 to 20 as the region comes in, and feed authentication stays post-quantum throughout.

On screenWhere to look
3D pictureThe 3D control on the map toolbar
Airspace volumesExtruded over Western Russia
Feed statusfeeds fused · feeds PQ-auth · sync local
ClassificationThe right-hand rail, click to filter
§ 4 · The map

If you know their stack

They sellWe publishThe difference
Foundry substrate · OntologyMidgard CoreOpen source, self-hosted, post-quantum sealed, air-gap native.
FoundryForgeDeclared, replayable pipelines with signed lineage on every object.
AIPMimirLocal inference only. Models pinned by hash. No cloud, ever.
ApolloBifrostReproducible builds, signed SBOM per release, offline delivery bundles.
Maven · GothamHeimdallEvery fused object carries provenance you can open and read.
Offline modellingGungnirDecision effects projected against live, governed data.
Vendor attestationForsetiSigned provenance and evidence packs anyone can verify.
Next

Each stands alone. Together they compound.