Separate on purpose
Radar route leaks have their own queue and dossier, outside Proofline. KEV and EPSS guide vulnerability collection; they do not label routing incidents. Country context does not enter routing features.
AGENT9 / ROUTING INTELLIGENCE / FIELD SCHOOL 02
From a route
to a defensible
judgment.
Learn what BGPKit adds to Signal Atlas.
Replay the evidence. Challenge the model.
Know exactly where the claim must stop.
7 guided modules · interactive exercises · no installation required
CASE REPLAY / SYNTHETIC
More data ≠ more proof
BGPKit adds reach into history.
Observer lineage determines independence.
07guided missions
05Atlas claim gates
16existing kNN features
00new observers from tooling alone
One lab. Three decisions.
START WITH THE QUESTION01 / INVESTIGATE
What changed?
Inspect an advertised path. Separate prepending from a non-adjacent repeat. Replay an origin change.
Open the workbench ↗02 / CHALLENGEWho actually saw it?
Add six receipts and watch independent observer counts refuse to follow the number of tools.
Test the lineage ↗03 / INTEGRATEWhat should ML do?
Extend the existing model with replay features, honest evaluation and explicit abstention.
Design the experiment ↗The finding that changes the design
Signal Atlas already solves source-overlap accounting. BGPKit belongs upstream of that contract as historical acquisition and parsing—not beside it as a new confidence vote.
CONFIRMED Reviewed in sourcePROPOSED Design, not deployed integrationSYNTHETIC Teaching data, not a live incident
01 / LEARNING PATH
Learn by making
the call.
Seven short missions. Read the objective, work the example, then defend your answer. Progress is saved in this browser.
0 of 7 completed
02 / ACTUAL REPOSITORY
Build on what
already exists.
A source-grounded map of Signal Atlas at 16ba3e2. Select a stage to inspect its behavior and implementation anchors.
Do not mistake scaffolding for coverage
RIPE Atlas has an approved-request lifecycle, but result receipt ingestion is not built. Planet catalog is default-off. Mission and intelligence governance need identity/provider completion; OAP keeps those claims locked.
Source decisions
KEEP CAPABILITY / REMOVE DOUBLE COUNTING“Current” means implemented in the reviewed repository. Deployed configuration, source freshness and production parity are separate checks. Source file names are review anchors, not public repository links.
03 / PATH WORKBENCH SYNTHETIC
A path is a clue.
Test the clue.
Inspect a plain AS sequence before turning it into a feature. These calculations run locally; this is not an MRT parser or an attack detector.
Replay the case
DOCUMENTATION PREFIX / PRIVATE ASNs203.0.113.0/24
Expected origin AS64500 · one synthetic collector-peer view
BaselineCandidateWithdrawalRestored
Keep three clocks separate
Observation time: when the collector saw it. Fetch time: when the adapter retrieved it. Validation time: when authorization was assessed. A current ROA lookup cannot silently become historical proof.
04 / EVIDENCE INDEPENDENCE SYNTHETIC
Six receipts.
How many witnesses?
Toggle the inputs. The exercise applies Atlas’s conservative peer-AS counting rule; it does not run the complete production Proofline evaluator.
COUNTING BOARD
Radar’s exposed peer set in this exercise is {AS64510}. Unknown overlap is not a certificate of independence. All direct examples describe the same prefix, origin and time window.
Two keys, two jobs
A receipt digest protects provenance. A semantic observation key links replay and live transports. Then peer-AS lineage bounds independent support. Keep all three layers; collapsing them into one hash loses either provenance or the duplicate relationship.
05 / LIVE DISCOVERY
Meet the source.
Read the limits.
Two existing read-only data paths. Nothing runs until you request it. Discovery finds archives; Radar supplies provider-reported candidate events.
BGPKit BROKER
API reference ↗Find a 15-minute UPDATE slice
UTC timestamps only. Results are capped at 10 files; this is not an exhaustive search. An empty result is an archive-discovery gap, not proof of no routing activity.
Awaiting your query.
CLOUDFLARE RADAR
Inspect provider-reported hijack candidates
Minimum provider confidence filter: 8. Scores remain provider values, not probabilities. Radar may share underlying peers with your direct collectors; independence requires case-specific overlap checks.
Awaiting your query.
From an archive to a parser job
Broker returns the MRT file URL. The external Parser then streams route elements for a bounded investigation. Keep the file hash, record position, peer identity, parser version and original path. The browser exercise above is a separate teaching tool.
import bgpkit
# Use an archive URL returned by Broker.
parser = bgpkit.Parser(url=archive_url)
for elem in parser:
# Preserve the native record before normalizing.
print(elem)API illustration based on the current official Python bindings; not an installed or executed parser job. Pin and validate the package version on the replay runner before use.
06 / ML DESIGN STUDIO
Make the analyst faster.
Keep the evidence in charge.
The first model decision is architectural: extend and evaluate the existing pipeline before introducing a second anomaly score.
16 features. 8 exercise analogues.
deterministic-knn-v1 compares bounded cases. The committed baseline is four benign and four suspicious synthetic analogues. It abstains without enough eligible reviewed examples. Its index is uncalibrated.
Use history to explain novelty.
Add origin novelty against a pre-event baseline, effective path-length change, prepend ratio, withdrawals, visibility loss and new adjacencies. Version the feature contract; retain missingness and collector coverage.
Language assistance, bounded.
The model may rewrite fixed narrative slots. Code owns facts, citations, hypotheses and actions. The OAP branch does not currently consume kNN output. A model never releases an alert.
Design a leakage-resistant experiment
This is a design check, not a trained model or measured accuracy result.
What to measure before claiming improvement
PROPOSED ACCEPTANCE CONTRACTUtility over spectacle
- Precision at a fixed analyst review budget, with adjudicated labels.
- Recall on independently labeled incidents, with missed-case analysis.
- Detection delay, false alerts per prefix-day and analyst time per case.
- Abstention, missing coverage and stability across collector outages.
No target accuracy is claimed. Establish the current deterministic/kNN baseline first, then report held-out differences and uncertainty.
Ablations that expose skew
- Remove Radar score; test circular supervision.
- Remove duplicate transports; test false volume gains.
- Hold out time, incident families and observer coverage.
- Drop one source; test graceful degradation and retained counterevidence.
Fit scalers on training folds only. Keep unknown labels out of the benign class. Retraining and probability calibration require a real reviewed corpus.
Where a second model earns its place
Evaluate Isolation Forest as a shadow novelty-ranking challenger on clean pre-event numeric features. Compare against the existing kNN and deterministic rules. Retain it only if analyst utility improves under the same holdout and review budget. No inference or training is running in this lab.
07 / INTEGRATION BRIEF
One bounded slice.
A credible next step.
The lab is implemented. The production adapter below is a proposal, anchored to the real receipt contract—not a claim that Atlas already ingests BGPKit.
Integration sequence
SHADOW FIRST / NO NEW ALERT TRIGGERA reproducible case bundle
Exact prefix and event window, prior RIB, ordered updates, collector inventory, bounded downloads, raw file hashes, parser version and gap ledger. Execute on existing compute, outside the request-serving Worker.
A versioned replay receipt
Extend ControlPlaneSource and SQL source CHECK constraints deliberately. Retain underlying collectorFamily, collectorLineage and peerLineage. Preserve AS sets; add message type and historical mode. Do not masquerade replay as live.
Prove non-inflation
Replay the same bundle twice: same semantic observations, no new independent peers. Change fetchedAt only: no new historical claim. A withdrawal, set, stale validator or missing segment must not become fabricated evidence.
Earn production admission
Compare historical features with the current model in shadow mode. Review failures and counterevidence. Promote only after receipt, freshness, lineage, retention, licensing and governed access tests pass.
Real schema mapping
| Replay input | Atlas contract | Required treatment |
|---|---|---|
| Collector + peer ASN/IP | collectorFamily / collectorLineage / peerLineage | Retain RIS or RouteViews; peer-as:<ASN> remains the boundary. |
| Prefix + path segments | watchedPrefix / observedPrefix / asPath / originAsns | Exact case scope. Preserve sets; do not invent one origin. |
| Record time + retrieval time | observedAt / fetchedAt | Historical mode is explicit; retrieval time cannot refresh evidence age. |
| Archive + record position | sourceNativeId / sourceUrl / raw | Add file digest, position, message type and parser version. |
| Transform + run provenance | observationDigest / transformVersion / retrievalRunId / licenseClass | Keep integrity digest; add cross-transport semantic linkage separately. |
Production source union currently permits only ripe-ris-live and routeviews-rib. A new identifier requires an explicit migration and corresponding readers/tests. No schema changes have been applied by this lab.
The five-minute hackathon demonstration
- 0:00 / Frame the case. Explain the actual Atlas pipeline and label the exercise synthetic.
- 1:00 / Replay the change. Show baseline, candidate announcement, withdrawal and recovery.
- 2:00 / Defeat duplicate confidence. Add a replay copy; show unchanged observer count.
- 3:00 / Bound the judgment. Separate routing, authorization and impact; explain missing evidence.
- 4:00 / Show the ML contract. Reject a leaked split, export the brief, and state the shadow-adapter acceptance tests.
Evidence and reference desk
Source review: fetched Forgejo origin/main on 7 September 2026; clean checkout at 16ba3e2ff66866870e06ee701460340ad0bc6c5b. Inspected ingestion, receipt types/persistence, source-lineage rules, Proofline, graph features, model assurance and deployment configuration. This is a focused integration review, not a full security audit or proof of deployed/source parity.
- Repository findings and implementation anchors
- BGPKit Broker — archive metadata discovery ↗
- BGPKit Parser — MRT/BGP/BMP processing ↗
- Python bindings and lookup capabilities ↗
- Radar BGP hijack investigation ↗
BGPKit’s lookup tools can add ASN, IP and community context. They are enrichment aids—not evidence of physical path or an actor. Historical ROA lookup is an integration candidate requiring event-time and validation checks.