The endpoints you actually use

A shared, rate-limited Free-tier key for evaluation. Not for production — it is shared with every other evaluator.

{
  "api_key": "or_…",
  "note": "This is a shared demo key with Free tier limits (100 routing decisions/day). Use it in the X-API-Key header to evaluate; sign up at /signup for your own key."
}

Your key's tier, daily limit and consumption. Pass the key in the header — never in a query string.

{"owner": "demo-user", "tier": "free", "daily_limit": 100,
 "usage_today": 0, "usage_this_week": 0, "remaining_today": 100}

The core call. Returns a decision plus the full reasoning behind it — no execution, no side effects.

{"model_name": "llama-3-70b", "gpu_required": "H100",
 "input_size_mb": 512, "estimated_compute_seconds": 30,
 "job_type": "inference", "priority": "high", "max_latency_ms": 200}

{
  "job_id": "d26e8183-da6",
  "recommended_target": "terrestrial_aws",
  "target_name": "terrestrial_aws",
  "preference_score": 74.28,
  "confidence": null,
  "confidence_status": "uncalibrated",
  "confidence_basis": {
    "status": "uncalibrated",
    "measured_base_rate": null,
    "base_rate_source": "graded_jobs",
    "base_rate_method": "One-sided 95% Wilson lower bound on the deadline hit rate over real graded jobs (internal synthetic load excluded), gated at n >= 50.",
    "sample_size": 0,
    "min_sample_size": 50,
    "degradation_factors": { "element_age": {"factor": 1.0, "why": "…"}, "…": {} },
    "degradation_product": 0.9
  },
  "estimated_latency_ms": 261.0,
  "estimated_cost_usd": 0.0346,
  "orbital_option": {"asset": "OrbitCloud Alpha", "score": 61.03, "latency_ms": 3316.8,
                     "cost_usd": 0.0995, "capacity_class": "demo",
                     "reliability": null, "reliability_status": "unmeasured"},
  "terrestrial_option": {"provider": "terrestrial_aws", "score": 74.28,
                         "latency_ms": 261.0, "cost_usd": 0.0346},
  "reasoning": ["Terrestrial (terrestrial_aws) selected — score 74.28", "…"],
  "ground_station": "Mumbai, India",
  "next_orbital_window_utc": "2026-07-30T15:05:55Z",
  "timestamp": "2026-07-30T15:07:28Z"
}

Read preference_score and confidence as different things. preference_score is the engine's weighted ranking of the chosen option, 0–100. confidence is a probability that the prediction holds, and it is null until a measured base rate exists — the platform publishes no confidence it cannot measure. capacity_class tells you whether an orbital option is a demo asset, a modeled expander asset, or federated operator capacity. All of it is derived on /methodology.

Cost only, across three clouds and the modeled orbital rate, plus the price at which orbit wins for this workload. Same request schema as /route.

{
  "orbital_cost_usd": 0.0995, "aws_cost_usd": 0.0346,
  "gcp_cost_usd": 0.0332, "azure_cost_usd": 0.0358,
  "savings_pct": -200.0, "orbital_premium_pct": 200.0,
  "basis": "modeled",
  "crossover": "Orbital pricing is modeled at 3.0x the cheapest terrestrial cloud rate ahead of the 2027 capacity wave — orbital reaches price parity for this workload when its $/GPU-hr crosses $3.98 (H100).",
  "recommendation": "terrestrial"
}

basis: "modeled" is the whole disclosure: the three cloud numbers are published list prices, the orbital number is a modeled multiple. The same math runs client-side, with the rate card visible, on /calculator.

Same request as /route, plus an optional deadline_utc. If the deadline cannot be met in orbit, the job is re-homed to terrestrial cloud and remediation records why.

{"model_name": "llama-3-70b", "gpu_required": "H100", "input_size_mb": 512,
 "estimated_compute_seconds": 30, "deadline_utc": "2026-08-01T18:00:00Z"}

{
  "job_id": "22cbe03b-4a0", "status": "queued",
  "target": "terrestrial_aws", "target_name": "terrestrial_aws", "asset_id": "",
  "gpu_type": "H100", "model_name": "llama-3-70b", "priority": "medium",
  "station_name": "McMurdo, Antarctica",
  "scheduled_start": null, "scheduled_end": null,
  "deadline_utc": "2026-08-01T18:00:00+00:00",
  "predicted_latency_ms": 30261.0, "predicted_cost_usd": 0.0346,
  "actual_latency_ms": null, "actual_cost_usd": null,
  "submitted_at": "2026-07-30T15:08:48Z", "started_at": null, "completed_at": null,
  "failure_reason": "", "remediation": ""
}

predicted_latency_ms is end-to-end — submit to complete, including pass-window wait — not link RTT.

Your jobs, newest first. ?limit= defaults to 50, max 200. Returns [] when you have none.

One JobRecord, including predicted-vs-actual once graded. 404 if it does not exist or is not yours — the two are indistinguishable on purpose.

Cancel while queued or scheduled; returns the record with status: "cancelled". 409 once it is running or terminal.

queued -> scheduled -> running -> completed
                              \-> failed
queued | scheduled -> cancelled

The scoreboard behind /sla: predicted vs actual across every graded job, misses included, plus scheduler counters. synthetic_jobs_graded is separated from jobs_measured so internal load can never inflate the customer-facing record.

{"jobs_measured": 9, "avg_latency_error_pct": 115.24,
 "deadline_hit_rate_pct": 88.9, "synthetic_jobs_graded": 19838,
 "scheduler": {"ticks": 109, "jobs_scheduled": 0, "jobs_dispatched": 13,
               "jobs_completed": 15, "jobs_offloaded": 0, "jobs_failed": 0,
               "last_tick_utc": "2026-07-30T15:09:13Z"}}

Reading from 2026-07-30. Nine graded jobs is a small number and it is published as one — below 50 the platform calls itself uncalibrated everywhere it matters.

SGP4 positions for every tracked asset, each with its capacity_class and the age of the orbital elements it was propagated from. This is a large payload; for a compact form of the same state use /api/v1/viz/snapshot.

Upcoming ground-station passes for one asset, gated by each station's own elevation mask.

{"asset_id": "OC-ALPHA", "passes": [
  {"pass_start": "2026-07-30T15:22:54Z", "pass_end": "2026-07-30T15:29:37Z",
   "duration_s": 403, "max_elevation_deg": 54.4,
   "station_id": "gs-dongara", "station_name": "Dongara, Western Australia",
   "minutes_until": 5, "in_progress": false,
   "truncated_at_window": false, "start_clipped": false}]}

in_progress and start_clipped exist because a pass already underway is a different fact from one that has not begun — the API says which.

All 37 modeled sites with position, link type, modeled bandwidth, elevation mask, and current weather-gated link recommendation.

{"count": 37, "stations": [
  {"station_id": "gs-svalbard", "name": "Svalbard (SvalSat)",
   "lat": 78.2307, "lon": 15.3896, "link_type": "both",
   "bandwidth_gbps": 4.0, "elevation_mask_deg": 5.0,
   "network": "KSAT", "source": "modeled",
   "weather": {"cloud_cover_pct": 62.0, "optical_viable": false,
               "rf_available": true, "recommended_link": "rf",
               "source": "simulated", "source_kind": "simulated"}}]}

What the network can actually do: contact duty cycle and worst gap for reference orbits, simulated against those exact 37 stations and their masks.

{"stations": 37, "stations_modeled": 37, "aggregate_bandwidth_gbps": 148.5,
 "contact_duty_cycle": [
   {"orbit_class": "leo_550_53", "label": "550 km / 53.0 deg (Starlink-like)",
    "duty_cycle_pct": 44.7, "worst_gap_min": 40.5, "contact_intervals": 79,
    "stations_considered": 37, "window_hours": 24, "step_seconds": 30},
   {"orbit_class": "sso_550_97", "duty_cycle_pct": 58.2, "worst_gap_min": 30.5},
   {"orbit_class": "leo_420_51", "duty_cycle_pct": 32.7, "worst_gap_min": 47.0}]}

A duty cycle in the forties with a worst gap around 40 minutes — the reading above, taken 2026-07-30 — is the honest shape of a 37-site network. A latitude-band count would have read better and meant less. Curl it for the current values.

The compact whole-system state behind the globe: satellites as positional arrays [lat, lon, alt_km, class], stations, modeled uplinks, and the subsolar point. Roughly an order of magnitude smaller than /satellites/positions — prefer it for bulk reads.

{"sats": [[-7.41, 93.16, 432.3, "t"], …],
 "legend": {"s": "starlink", "o": "oneweb", "t": "crewed", "x": "other"},
 "stations": [{"id": "gs-svalbard", "name": "Svalbard (SvalSat)", "lat": 78.2307,
               "lon": 15.3896, "bandwidth_gbps": 4.0, "source": "modeled", …}],
 "links": [{"name": "STARLINK-1807", "sat": [35.61, 139.44, 463.6], "si": 8,
            "elev": 87.1, "type": "optical", "gbps": 10.0}, …],
 "links_note": "Modeled line-of-sight uplinks, LEO only (<= 2000 km), ranked by bandwidth x sin(elevation) — usable geometry, not elevation angle alone.",
 "sun": [18.397, -48.545],
 "stats": {"tracking": 11690, "active_links": 7113, "pass_windows": 118499,
           "fleet_size": 11692, "last_tle_refresh": "2026-07-30T14:48:59Z"}}

Register once. The returned or_op_ key is shown exactly once and is not recoverable.

{"name": "Acme Orbital", "contact_email": "ops@acme.example", "jurisdiction": "US"}

Advertise a node. It is directory-listed immediately; it becomes routable only when OrbitRoute can propagate its orbit (a resolvable norad_id, or TLE lines under 14 days old) and has attested that you operate that spacecraft. The 201 tells you which condition is outstanding.

{"node_id": "acme-tile-1", "norad_id": 25544, "gpu_type": "H100",
 "gpu_count": 8, "power_budget_w": 12000, "price_usd_per_gpu_hr": 1.99}

Additional advertised physics — battery_fraction, thermal_margin_c, storage_gb, isl_neighbors, jurisdiction, attestation — is defined in the full reference under CapabilityAdvertisement. Fields the scorer does not yet read are stored and served back, not silently dropped.

The public directory: every advertised node with its live position, illumination and verification state. Returns [] today — no operator has registered yet, and an empty directory is published as empty.

Your own nodes, including anything not yet routable and the reason.