You don't need more AI tools. You need to solve problems at the speed of the fight.
Two days. 2 prototypes, 3 MVPs, 8 PoCs, 8 conceptualized wireframes. The AWS infrastructure they all can share. The tamper-evident ledger tracking every human-machine interaction.
We didn't buy more AI tools, we learned how to use the ones we had.
The operating model IS the accelerator.
The frame · how to read every mission on this site
Your AI isn't a tool. It's your newest employee.
The biggest question in AI isn't which tools are best. It's the one everyone keeps asking: how do we use this? The answer is not more AI tools. It is leading the ones you already have.
Huge speed and efficiency gains in digital problem solving. The operating model took a full production platform from nothing to launched in 266 logged hours, counted, not estimated. The same scope, built the conventional way, models to 8,000 to 21,000 engineer-hours (Brooks team-sizing). Here you watched the same model stand up 21 mission builds, five live and governed on one engine, in two days.
One shared brain, as code
Every agent works from the same context, memory, and rules, kept as code. Train it once; it stays trained.
Rules that keep it honest
Every claim traces to an artifact. Nothing unsourced ships. A health check pulls any agent that drifts.
A human leads, all logged
A human approves every irreversible call. The decision signs into one hash-chained ledger you can replay.
Three doors
The sequence-of-eyes for judges: README, the 90-second receipts, the live consoles, the credibility receipts. Lands the load-bearing claims without missing the receipts.
Open the judge walkthrough →Switch between #01 / #02 / #04 / #07 / #08 on the mission rail. Same gate, same signed ledger, same always-live abort underneath. Per-mission SCOPE OF CLAIM panel travels with the demo.
Open the C2 console →The Counter-UAS floor, rendered as a live 3D battlespace. Real RF / radar / EO-IR detection geometry, drones on real tracks, an LLM-free threat call on each, and a recommend-only DEFEAT held behind a two-person gate. Same gate, same signed ledger as the C2 console — no separate build, no fork.
Open the C-UAS 3D →The thesis
The deep proof is Counter-UAS #08, the 1 km² defended area, real C-UAS track fusion through a live stream, the graduated-authority gate with positive hostile-ID at weapons-tight and a two-person rule, effector tasking under BSI Flex 335 / MITRE CoT (simulated). #07 FOB UAV/UGV is the same engine, the discrimination win: a real ADS-B Cessna 172 (ICAO a24c5a) correctly classified cooperative against the synthetic drone threat. #01 seabed, #02 convoy, and #04 swarm-recon are honest width on the shared spine: real corpora, governed gate, signed ledger, mission posture stated on each card.
The rubric front-loads Working Prototype. One running thing beats four mocks. We win on the depth of the governed loop, not the breadth of the graph. The line we will not cross: selling maritime, seabed, and swarm at parity to C-UAS is overclaim. We do not.
The five missions
Area Counter-UAS
The 1 km² defended area. Radar + RF + EO/IR fused tracks, IMM trajectory prediction, NINJA-EW + LEONIDAS-HPM under BSI Flex 335 / MITRE CoT (simulated effector). Two-person auth required at engage; RF NO-EMIT sector; graduated ROE.
FOB UAV/UGV Defense
Same engine as #08, no fork. Real ADS-B cooperative anchor (Cessna 172, ICAO a24c5a, adsb.lol open API) correctly classified cooperative against the synthetic drone threat. Real cooperative aircraft, distinguished from a threat, on the same gate.
Seabed Cable Corridor
Surface-AIS picture over a public submarine-cable landing area (San Pedro Channel). The shared gate and ledger under UNCLOS recommend-only posture. Subsurface acoustics deliberately excluded; the metamorphic harness proves no answer-key leak.
Convoy Escort
Real Marine Cadastre AIS (San Pedro Bay), same governed gate, signed ledger. Cooperative AIS gets HOLD_FRIENDLY; hostile or suspect gets RECOMMEND under UNCLOS posture. Same engine as #01, different mission posture.
Swarm Reconnaissance
Mapping mosaic, not threat tracking. Real SkyFusion overhead-imagery detections (Apache-2.0, annotations only) driving a mosaic-fusion variant on the same gate and ledger spine. COP-commit is human-gated; the gate ladder differs from #08 by design.
The architecture
- AWS Snowblade (DoD / JWCC) for the forward node, or
- IoT Greengrass v2 + SageMaker Neo small models on the device
- Sensor fusion, detection, and the gate run with no reachback; deltas buffer for reconnect
- Bedrock + AgentCore are cloud-only; reached when the link is back
- S3 Object-Lock WORM anchor for the ledger; reconcile on reconnect
- Cross-site intelligence and after-action records consolidate here
Team and lanes
How to start
$ git clone <wolfberg-peraspera-hackathon> $ cd wolfberg-peraspera-hackathon $ ./onboarding/bootstrap.sh # corepack + pnpm install + typecheck + contract loop # expect: OK — contract loop verified. $ pnpm --filter @wolfberg/c2-mission demo # #08 floor: real drone-track fusion + gate + signed ledger $ pnpm --filter @wolfberg/c2-mission demo:07 # #07 FOB reuse: same engine, swap config $ pnpm --filter @wolfberg/c2-width demo # #01 + #02 maritime: real NOAA AIS + governed gate + ledger $ pnpm --filter @wolfberg/c2-fusion test # #08 12/12 + #07 13/13 answer-key invariants
Schedule and constraints
Constraints (hard)
- AWS Kiro mandatory for the build
- Real corpus + disclosed synthetic scenario per mission. No classified or PII inputs (MMSI hashed at the adapter; SkyFusion annotations only)
- Commercial AWS, us-east-1. Not GovCloud; classified and PII are out of scope
- Event: AWS Global Government Hackathon. Judging Wed 2026-07-01
Ship-Reality (non-negotiable)
- Running prototype on real open-source data with the governed loop. Not a deployment, not an ATO.
- No prompt-caching cost claim. Not wired; do not put a caching number on any surface
- NIST 800-53 = Rev 3-class self-assessment, r5 mission mapping. Never say "compliant"
- Effectors are simulated. Nothing actuates. The ledger records what would have been dispatched