Begin with a decision—not a universe.
A route contract, an orbital-state sensitivity, or a provenance gap can be checked. “Solve satellite networks” cannot.
Open the research mapTrace acquiredRoute contract visible · deployment authority withheld
Field 01LEO / NTN · validity auditField 02Agent evaluation · holdField 03Agent × networkPolicyPublic by default · evidence before claim
Source ≠ reproductionFixture ≠ benchmarkLocal pass ≠ deploymentTrace before answer
One signal · four state changes
Watch a claim earn its weight.
Motion is not decoration here. Scroll changes the state of the same evidence trace while the written boundary stays readable.
Give every source a role.
A paper can report a mechanism. A repository can expose an artifact route. Neither becomes a Nimbus reproduction until inspected and executed.
Inspect curated sourcesRun the smallest check that can fail honestly.
Deterministic validators and route-aware fixtures test one mechanism while preserving alternate routes, unsafe side effects, and no-safe-action cases.
See the local fixtureEnd with what the result cannot authorize.
Fixture integrity is not model performance. A local pass is not deployment evidence. The boundary ships with the finding.
Read the release rules
Trace complete / 04
Four states retained.
The boundary remains attached when the next research field enters view.
Active research fields
Three scales of the same problem.
01
LEO / NTN
Orbital state, handover, RF/optical links, service continuity, and evidence-qualified operating envelopes.
↗ 02Agent evaluation
Tool-use traces, persistent state, false-green prevention, executable verifiers, and explicit DEMO boundaries.
↗ 03Agent × network
Telecom-specific action gates where physical constraints and provenance remain outside the language model.
↗Publication boundary
A pass is
not a claim.
Two public surfaces, two different promises.
Progress reports what Nimbus is reading, running, planning or holding; it is an activity log, not validation. Simulations is the validity-claim surface and must retain an explicit HOLD/PUBLIC/ARCHIVE boundary. Presentation may guide attention; it cannot raise the evidence tier.