PowerRoam

PowerRoam is the ground proof-of-concept of ComputeFollowsPower: instead of building power out to fixed compute, mobile compute roams to wherever power is cheapest, and the binding constraint is swapped from grid-lead-time to a communication-latency budget. For a batch of latency-tolerant jobs it answers:

On the ground, can dispatching jobs to the cheapest-power node within their latency budget actually beat a fixed grid-anchored site once the movement/comms cost is paid?

It dispatches each job to the cheapest-power node reachable inside its latency budget, computes the power arbitrage net of movement cost, and emits a posture verdict: viable, marginal, or blocked.

The Insight

A three-step transform on one counterfactual (L9: what if the grid scaled at software speed?):

• L1 direction reversal — flip "power → compute" to "compute → power": send compute to abundant power instead of waiting for substations.
• L9 scale shift — from a few campuses to a swarm of thousands of roaming nodes (LEO included).
• L17 constraint substitution — the binding constraint becomes the communication-latency budget, not the grid-lead-time.

This is the idea pool's most adventurous entry (pool-best surprise, lowest coherence), so PowerRoam is deliberately the feasibility beachhead — a ground-only, deterministic simulator of the one question everything else depends on.

Feasibility Discipline, Forced into Code

• Ground first (R1). Nodes carry an is_leo flag and the policy gate allow_leo defaults false, so LEO nodes are excluded from dispatch until feasibility is proven. Enabling allow_leo admits them.
• Latency-tolerant market only (R2). Every job carries a latency_budget_ms; a node is a candidate only if reachable within it (the L17 gate, made literal).
• The economic envelope is the verdict (R4). A reachable job whose power saving does not clear min_arbitrage_margin after movement cost is conditional — relocation doesn't pay — never silently dispatched.

Quickstart

python -m powerroam sample   --output examples/scenario_sample.json
python -m powerroam dispatch --input examples/scenario_sample.json
python -m powerroam report   --input examples/scenario_sample.json --output examples/powerroam_report.md

Scenario Spec

{
  "policy": {
    "min_arbitrage_margin": 0.1, "max_latency_ms": 1000.0,
    "allow_leo": false, "enforce_capacity": true
  },
  "nodes": [
    { "node_id": "solar-desert", "power_price": 0.01, "capacity_units": 800.0,
      "latency_ms": 120.0, "move_cost_per_unit": 0.008, "is_leo": false }
  ],
  "jobs": [
    { "job_id": "train-batch", "compute_units": 200.0,
      "latency_budget_ms": 500.0, "baseline_price": 0.05 }
  ]
}

How It Works

for each job (in input order):
  candidates = nodes where:
      not (is_leo and not allow_leo)                      # R1 LEO gate
      latency_ms <= min(latency_budget_ms, max_latency)   # R2 / L17 latency gate
      capacity_units >= compute_units                     # finite supply
  if no candidate                          -> blocked (leo_disabled / latency_exceeded / no_capacity)
  else best = cheapest  compute_units × (power_price + move_cost_per_unit)   # tie-break by node_id
       net_saving = compute_units × baseline_price − best_cost
       margin     = net_saving / baseline_cost
       margin < min_arbitrage_margin        -> conditional (margin_too_thin)   # R4 envelope fails
       otherwise                            -> realizable (dispatched; capacity consumed)

The swarm posture aggregates:

• jobs dispatched and aggregate net saving > 0 → viable (envelope holds)
• jobs reachable but the arbitrage doesn't pay → marginal
• nothing reachable → blocked (exit 1)

An unknown field, a negative magnitude, or a non-positive latency budget is a blocking issue with a nonzero exit, never a stack trace.

MVP Scope

PowerRoam produces a ground-PoC dispatch plan and economic-envelope verdict under a declared policy, not a production scheduler, an SLA commitment, or an orbital-deployment plan. It simulates and flags; it does not run jobs.

Provenance

PowerRoam is the concrete feasibility beachhead of IDEA-001 (Compute Follows Power: Orbital Load-Roaming Swarm — CIX rank 1, pool-best surprise), reached via the L1_DirectionReversal → L9_ScaleShift → L17_ConstraintSubstitute lens stack (CIX-20260607-001 ← IDX-20260607-001 ← TCX-20260606-001). It is a sibling of WasteStack, FailureFutures, RoboTrace, InferMesh, CertMesh, SettleMesh, SpendMesh, ReleaseMesh, ClimateMesh, SovMesh, FlowMesh, and SignalMesh. It is the directional inverse of WasteStack (fixed-site utilization ↔ roaming), so a roaming node can become a tenant of a WasteStack energized site.

License

MIT — see LICENSE.