DebrisGuard. Multi-modal Space Domain Awareness.

Why DebrisGuard wins

The only SDA platform your insurer can re-verify.

Every alert ships with its inputs, model version, and a public verification URL. That is the evidence chain an underwriter would demand, built in from day one.

100%

Independently verifiableHMAC-signed dossier on every alert; recompute it on your own laptop.

0.91

BSTAR-drift AUC · 45-day horizonn = 522 · 95% CI 0.88–0.92 · audited 12 May 2026.

$0 → $299

Self-serve in < 5 minThe free tier is the real product, not a demo.

Built for the three questions insurers ask

"How did you know?" Fused score breakdown by detector.
"On what data?" The exact TLE pair, by reference and timestamp.
"Is the chain intact?" HMAC signature, recomputable.

Dossiers are retained for two years, so they still verify when a claim is adjusted.

2 yrforensic retention on Mission Control

1 % of the price of a sensor prime

Sensor primes start at $50 k+/year, sales-led. DebrisGuard starts at $0 → $299/mo, self-serve, on the same public-catalog inputs. On what most operators actually need, the capability gap is close to zero.

99%cost-gap vs the sensor primes (apples-to-apples)

DebrisGuard forensic dossier: a JSON alert panel showing norad_id, fused_score, model_version, an HMAC-SHA256 signature and a verify_url, sealed with a signature mark and linked to an 'independently verified' check — The artefact behind every alert: inputs, model version, HMAC signature, public verify URL. Recompute it yourself.

Capability matrix · what an SDA buyer actually asks for

Live numbers, not asserted ones · see the benchmark →

Capability

LeoLabs · Slingshot

Space-Track · Celestrak

DebrisGuard

Manoeuvre detection on public TLEs

✓Premium add-on

✕DIY in Python

✓Built-in · published rate

Early-decay watch with published ROC-AUC

~Implicit · not published

✕No

✓0.91 AUC · n=522 · reproducible

HMAC-signed forensic dossiers

✕PDFs only

✕No

✓Share by URL · re-verifiable

Public, reproducible benchmark

✕No

✓Cosmos 1408, Iridium-Cosmos

Self-serve sign-up (no sales call)

✕6-week cycle

~Form + approval

✓Cognito + Stripe · <5 min

Outbound alert webhooks

~Custom integration $$

✕No

✓Built-in (Operator+)

Entry price · 5-satellite operator

✕$50 k+ / yr

✓$0 · no UX

✓$0 → $299 / mo

Four detectors. One score.

Multi-modal anomaly detection.

Four detectors run in production tonight, fused so corroborating signals reinforce each other and lone false alarms get dampened. One of them is unique to DebrisGuard: decay-by-silence, which reads the disappearance of a low-flying object from the public catalogue as the strongest sign it is about to come down. The RF, optical, and ViT models are trained and join the score in v1.1.

In production tonight all four fused into every score

Decay-by-silence Live · only here

A low-flying object that stops being catalogued is usually an object about to re-enter. We learn each object's own tracking cadence, and when the updates stop at low perigee the silence itself raises the alarm. Weeks of warning from the one signal everyone else ignores.

backtested on 4,349 real re-entries · per-object cadence gates · signed evidence on every alert

Catalog anomaly (snapshot) Live

Per-orbit-class Mahalanobis distance over the six classical elements. Catches out-of-family launches, fragmentations, and gross orbit departures.

30 000+ ref objects · per-class χ² · nightly

Manoeuvre / burn detection Live

Per-element z-scores between TLE pairs, gated by per-class noise floors. Every burn is confirmed or retired against the next post-burn TLE.

14-day rolling window · per-class noise floors · post-TLE confirmation

BSTAR-drift early-decay watch Live · validated

Tracks BSTAR and mean-motion slopes to catch drag ramping up, the early signal of re-entry, weeks before a public decay warning.

45-day AUC 0.91 [0.88-0.92] · n_pos = 522 · May 2026 · verify on benchmark

Roadmap · v1.1 models trained; join the fused score when the capture pipelines ship

RF carrier classification Roadmap v1.1

1-D CNN over 24 modulation classes, trained on RadioML 2018.01A. Joins the fused score when the capture pipeline ships.

top-1 48 % · macro-F1 0.43 · 24 classes

Optical streak detection Roadmap v1.1

YOLOv8 streak detection on star-fixed exposures with sub-pixel localisation. Ships with the v1.1 optical pipeline.

mAP50 0.995 (synthetic) · P 0.99 · R 0.99

ViT object anomaly Roadmap v1.1

A frozen DINO ViT flags objects that don't look like anything in the known fleet. Ships alongside the optical pipeline.

AUC 0.995 (held-out) · recall@1 %FPR 100 %

Worked example

60 seconds with DebrisGuard.

Here's a real burn DebrisGuard caught overnight on LANDSAT 8 (NORAD 39084). What the catalog showed, what we flagged, and the advice we shipped.

Public catalog

15.51

Mean motion (rev/day) on the latest TLE looks nominal. Decay roster: not listed. CDM screen: clear.

What an operator would see on Space-Track alone.

DebrisGuard fused

0.70 amber

Two detectors fired. Burn: +2.6×10⁻⁴ rev/d, z = 50, surfaced on the next nightly cycle.

Three-detector ensemble fused via noisy-OR.

Recommended action

Phasing burn. Re-screen conjunctions for the next 48 h.

Routed into the daily digest, with the full HMAC-verifiable evidence chain one click away.

Every alert ships with this advice.

Why it matters: the public catalog stays silent on burns this size for 12–48 hours. DebrisGuard surfaces them on the next nightly cycle, so you re-screen before the next CDM lands. Open the signed dossier →

Pricing

Start free. Scale to a constellation.

Every tier ships only features we operate today. No aspirational line items, no surprise upcharges. All prices in USD, billed monthly, cancel any time.

Watcher

Hobbyists, students, curious minds.

$0

Forever free

Live anomaly dashboard (read-only)
10 catalog object lookups per day via REST API
Latest daily anomaly report (HTML + JSON)
1 NORAD ID on personal watchlist
Read-only forensic dossiers (HMAC-signed)
Community support via GitHub issues

Three calls to a fused score.

Token-authenticated, rate-limited per tier, deployed on api-debrisguard.deepfieldlabs.dev/v1/*. Python SDK on the v1.1 roadmap.

Bearer JWT or X-API-Key. Cognito-issued tokens for browsers, hashed API keys for machines.
Same fused score the dashboard uses. No "API tier" dilution. What you see is what you get.
Daily report endpoint. /v1/reports/daily/latest returns the same JSON the dashboard renders, ready for your pipeline.

# 1. Score one catalog object end-to-end
curl https://api-debrisguard.deepfieldlabs.dev/v1/objects/25544 \
  -H "X-API-Key: $DG_KEY"

# 2. Pull today's anomaly report (JSON + HTML)
curl https://api-debrisguard.deepfieldlabs.dev/v1/reports/daily/latest \
  -H "X-API-Key: $DG_KEY"

# 3. Add a NORAD ID to your watchlist
curl -X POST https://api-debrisguard.deepfieldlabs.dev/v1/fleet \
  -H "X-API-Key: $DG_KEY" \
  -H "Content-Type: application/json" \
  -d '{"norad_cat_id":"25544","label":"ISS","threshold":0.65}'
    

FAQ

Questions you'll ask anyway.

What sources do you ingest?

The nightly production loop runs on Space-Track (GP, CDMs, decay roster, SATCAT) and Celestrak (TLEs). SatNOGS (RF captures) and DISCOSweb (physical properties) feed the v1.1 RF and enrichment work; the optical capture pipeline lands in v1.1.

How fresh is the data?

Catalog refit nightly at 00:00 UTC, scoring every active object. Per-asset scoring on demand from the API. Per-tier rate limits apply. Alert webhooks to your own endpoint ship today on Operator and above; sub-hourly streaming beyond the nightly cycle sits on the v1.1 roadmap.

Can I see why an alert fired?

Every alert ships the full per-modality breakdown: catalog d², manoeuvre delta-mean-motion plus per-element z-scores, BSTAR-drift slope fit, and the silent-gap evidence, with RF and optical to follow in v1.1. Your analysts can audit every decision, and each forensic dossier is HMAC-signed for audit trails.

What about false positives?

The fused score requires consensus across detectors before firing. A single noisy modality cannot trigger an alert on its own. On Operator tier and above you also get per-asset thresholds so you can dial precision and recall to your own playbook.

Do you store our orbital data?

Watchlist NORAD IDs and your alert history, yes, retained per your tier's window (7-day on Watcher, 30-day on Analyst, 90-day on Operator, 2-year on Mission Control). Raw third-party catalog data is not redistributed; you get the score, not the source feed.

Where does DebrisGuard run?

Entirely on AWS in the Sydney region (ap-southeast-2): a serverless API, a containerised nightly pipeline, and versioned artifact storage with tamper-evident hash chains. Data stays resident in Australia, every endpoint is TLS, and the audit trail is HMAC-signed at write time. Nothing runs on hardware we could lose in an office move.

Can we run it on-prem?

Institutional tier supports deployment in your own AWS account or sovereign region, on bespoke commercial terms. Air-gapped container delivery is in scope but planned per engagement (contact us for the security review).

Working with operators and institutions

Independent. Deployable wherever your mandate lives.

DeepFieldLabs is an independent space-intelligence lab. We work with civil space programmes, national regulators, defence and security organisations, research groups, and commercial operators on bespoke terms, always under their own governance. Where required, DebrisGuard can be deployed inside your own AWS account or sovereign region, with the full source of the detection engine reviewable under NDA.

What happens next

We reply within one business day, AEST hours.
For institutional deployments, we can run a 30-minute scoping call before any commitment.
Submissions land in our shared mailbox via SNS, no public inbox is exposed on this site.
We do not add you to any marketing list. Single-thread reply only.

Watch every object in low Earth orbit. Catch the anomaly before it catches you.

Proven on real production traffic, not slideware.