RelayProof.METHOD / LIMITS / CREDITS
A transparent model
is a better starting point.
RelayProof is an early planning tool for students, community networks and field sensor teams. It combines an explicit RF budget with a separately loaded, portable clearance network. All example inputs are hypothetical, and no field measurements or customer adoption are claimed.
1. Radio budget
We use the exact SI free-space expression, with c = 299,792,458 m/s, distance d in metres and frequency f in hertz. Inputs in MHz and km are converted before calculation.
FSPL [dB] = 20 log10(4π d f / c) EIRP [dBm] = Pt + Gt − Lt Pr [dBm] = EIRP − FSPL + Gr − Lr − Lextra Margin [dB] = Pr − receiver sensitivity Headroom [dB] = margin − target reserve
The radio screen passes only when the margin meets the entered target and EIRP is at or below the entered design cap. A 2× distance or frequency change adds approximately 6.0206 dB loss. See ITU-R P.525-5.
Receiver sensitivity must match the intended bandwidth, modulation, coding and data rate. The EIRP cap is supplied by the user: band, jurisdiction, antenna restrictions, spectral density, duty cycle, indoor/outdoor use and equipment certification are not determined here. No example value grants permission to transmit.
2. A separate physical clearance screen
ClearPath is a companion agent website included in this project. It can be co-hosted or served on a different origin. Its graph is fetched and validated; its declared, SHA-384 pinned code runs in the caller’s isolated worker. It is not a real external inspection organization.
λ = c / f r1 [m] = √(λ d1 d2 / (d1 + d2)) Earth bulge [m] = d1 d2 / (2 k R), R = 6,371,000 m Line height = hA + (hB − hA) × obstruction fraction Available clearance = line height − obstruction top − Earth bulge Required clearance = selected clear-radius fraction × r1 Equal mast raise = max(0, required − available)
All heights must use the same elevation datum. “Antenna elevation” means the elevation of the antenna phase centre, not just mast height above unrelated local ground. The effective Earth radius factor k defaults to 4/3 as a hypothetical nominal case; actual refraction varies. The 60% threshold is an initial obstruction screen. It is not the complete ITU-R P.530 antenna-height and availability procedure. See ITU-R P.526 and ITU-R P.530.
Only one sampled obstruction is modeled. Other terrain points may dominate. We do not convert a blocked Fresnel screen into a guessed diffraction loss or combine it with the RF margin as a reliability percentage. Wind loading, structural design, vegetation growth, rain, multipath, interference, atmospheric variability and long-path terrain profiles need further engineering work.
3. Search and comparison
The finite candidate search minimizes equal antenna gain at both ends, then conducted transmitter power. It checks every unique candidate pair against the radio margin and EIRP cap. It does not optimize monetary cost, hardware availability, physical antenna size, mounting feasibility or network availability. If no pair passes, the result says so. A feasible radio candidate still needs a separate clearance screen.
The comparison holds unspecified inputs fixed. The distance sweep holds all other RF inputs fixed and has no changing terrain geometry. Numerical feasibility decisions use full precision with a 1e−9 comparison tolerance; display values are rounded. Construction decisions need additional reserve beyond an exact threshold.
4. Agent networks and inference
The public artifacts follow the submission-start MVP Profile 0.2. JSON is a valid resolved HOCON serialization. Both networks have a front agent, a specialist and typed portable coded tools. Public artifacts include manifests, resolved definitions and publisher-owned links.
The keyless default is structured-task-v1: a deterministic symbolic inference adapter that searches the published graph for an allowed path to an explicit typed capability. It is a real graph and tool execution path, not a mock LLM or free-form language model. It intentionally cannot interpret arbitrary conversation or arbitrary agent instructions. That limitation is visible in the interface and manifests.
A visitor may opt into an OpenAI-compatible, tool-capable inference service for general reasoning. Router URL, model and optional credential remain a caller choice. The credential lives only in a private object in tab memory, is attached only to the chosen inference endpoint, and is absent from prompts, manifests, tool inputs, traces and peer requests. The optional provider may charge for inference and must support browser CORS. Provider availability is outside this static site’s control. Use “Forget & use local planner” to remove the key.
Ordinary peer instructions can be interpreted by a visitor-selected model; this optional path has protocol-fixture tests but no real paid model call has been performed in this preparation package. Full public-host HTTPS/CORS validation is a deployment gate, still pending entrant approval. These are verification limits, not completed production claims.
5. Trust and execution
External artifact fetches use credentials: omit and reject redirects. A new third-party contact needs a visible origin, URL and purpose with explicit approval. This strict policy asks per fetch; it does not remember blanket consent. Public instruction text remains untrusted. SRI detects an altered tool relative to its manifest; it does not authenticate the publisher or establish instruction safety.
Executable tools use the official worker sandbox with a restrictive worker CSP. The host rejects malformed schemas and graphs, wrong exact network names, undeclared actions, SRI failures, graph cycles and exceeded budgets. The default maximums are depth 6, 24 invocations, 16 artifact contacts, 12 model steps per agent, one concurrent task and 20 seconds total. A stop control cancels in-progress requests and terminates the worker. Private context is deny-by-default at peer boundaries, with only destination-specific caller allow rules; router credentials never enter private context.
Every local write is reviewed before execution, including reversible saves and deletes. The exact reviewed arguments, origin and request ID are bound to the receipt. A write timeout is “unknown”, with no automatic retry. Only the last five scenarios are stored locally, and storage failures are reported. No analytics or third-party scripts are loaded.
6. Evidence and privacy
An exported JSON record contains inputs, model results, assumptions, selected network metadata and a SHA-256 digest. The digest detects accidental changes to a record; anyone who edits a record can compute another digest, so it is not a signature or certificate. Imports validate the digest and schemas, discard imported claims, and recalculate the radio budget. The path must be audited again.
Default numeric tasks run locally. A consented peer serves public artifacts; the numerical invocation itself runs in the caller browser. Optional inference sends the request and tool results to the selected router. Public agent instructions and source are intentionally inspectable. The browser tab must remain open for execution.
7. Open source and AI assistance
Original interface, calculations, graph host, agent definitions, tests and documents are licensed Apache-2.0. The official Cognizant agentic-web runtime is retained under its Apache-2.0 license; Ajv is MIT licensed. See third-party notices and license. RelayProof is not endorsed by Cognizant or ITU.
OpenAI Codex was used for ideation, code, documentation and tests. The entrant’s stated preparation workflow also includes Anthropic Claude for topic selection and review advice; no separate Claude review of this final implementation is claimed. The entrant must review, understand and finalize the work and explain it personally in the demonstration. No synthetic narration is included.