Product roadmap · v1.0 · updated 18 Apr 2026

From one parcel in South Memphis to every industrial coordinate on Earth.

The pilot proved a single site. The platform has to prove a planet. Here's how Ghost Permits gets from Colossus-1 to an agent-powered accountability layer anyone can point at a pair of coordinates.

StatusPhase 1 — live pilot PlatformSentinel · Compass Horizon18 months TeamOpen · seeking contributors
North star

One pipeline. Two modes. The same satellite data answers both sides of the question.

Today Ghost Permits is a hand-curated intelligence tool. It monitors three sites we believe matter, produces a written brief, and publishes the numbers for the public record.

Tomorrow it should be an infrastructure intelligence platform. The same orbital signals that catch a facility cutting corners can tell another facility where not to cut them — clean atmospheric headroom for faster permitting, adequate grid, stable ground, suitable cooling. xAI's Memphis problem (unpermitted turbines, federal lawsuit, community backlash) was predictable from orbit before a single permit was filed.

So we split the product without splitting the pipeline. Sentinel is the watchdog — accountability for the public. Compass is the site-intelligence system — due diligence for the industry. One satellite stack serves both. The watchdog stays free because the due-diligence product pays for it.

“Sentinel says what went wrong here. Compass says what's right for what you need. Same pixels. Different question.”
Platform architecture
Two modes, one pipeline, two markets.

The same orbital signals flip direction. Sentinel asks “what's wrong here.” Compass asks “what's right for what I need.” Run the pipeline forward, you get a watchdog. Run it backward over a candidate-site grid, you get site intelligence. The data is identical; only the query changes.

Mode A · public good · free tier
Sentinel
The watchdog.

Input: a coordinate, an address, a three-word tag, or a global scan.

Output: anomaly brief · permit gap · evidence packet · continuous watch.

Users: regulators, journalists, NGOs, law firms, downwind communities.

Example query
“Is the new building at 35.0449° N / 90.0817° W emitting without a permit?”
→ 184-day permit gap, NO₂ anomaly +47%, 35 turbines visible, draft brief with citations in 60s.
Mode B · B2B · commercial tier
Compass
The due-diligence system.

Input: facility class, power requirement, cooling profile, region of interest.

Output: ranked candidate sites scored on seven orbital dimensions · comparables study · risk map.

Users: hyperscalers, data-centre developers, energy companies, siting consultants, state economic-development offices.

Example query
“Rank 200-acre parcels within 50mi of Columbus, OH for a 250MW liquid-cooled data centre.”
→ 12 candidates ranked by atmospheric headroom, grid proximity, cooling fit, subsidence risk, industrial density, community index, comparable-facility fit.
Compass scoring dimensions · what every candidate site is graded on
01
Atmospheric headroom
How far below EPA thresholds is the current airshed? More headroom = faster permit.
Sentinel-5P · 7-yr baseline
02
Grid proximity
Transmission infrastructure resolvable from orbit — substations, HV lines, transformer yards — plus OSM + utility territory overlay.
Sentinel-2 · OSM · utility data
03
Cooling suitability
Surface-temperature baseline, heat-island intensity, proximity to water, prevailing wind.
Landsat-8 TIRS · ERA5 · S2
04
Ground stability
Millimetre-scale subsidence history over 3 years. Heavy facilities on subsiding ground = expensive mistake.
Sentinel-1 SAR · InSAR
05
Industrial density
Existing emission sources in the airshed. Cumulative-impact risk drives permit timelines.
Sentinel-5P · ECHO · NPRI
06
Community sensitivity
EJScreen, demographics, cancer-registry overlay. EJ risk is permit risk.
EJScreen · Census · CDC
07
Comparable-facility fit
LFM2.5-VL finds the 10 most similar existing facilities globally and compares their operational signatures.
LFM2.5-VL · global index
Design principles

Everything we build has to pass these four tests.

01
Reproducible, always.
Every claim traces to free, public data. A competent student should reproduce any finding in an afternoon.
02
Legible to non-experts.
If a community member, journalist or lawyer can't understand the output in 60 seconds, the tool has failed. Jargon is a regression.
03
Agentic, not autonomous.
Agents gather, cross-check, and draft. Humans decide what gets published and what gets flagged. The verdict is never the model's.
04
Free to the public, funded by power.
Anyone downwind can use it at zero cost. Regulators and law firms pay for the API, the audit log, and the watchlist depth.
Phased rollout
Five phases. Roughly eighteen months.
Timelines are targets, not promises · reprioritised every quarter
Phase 0 · shipped
The Memphis pilot
Q1 2026
Phase 1 · in flight
Any coordinate, any layer
Q2 2026
Phase 2 · next
Watchlists & alerts
Q3 2026
Phase 3
The investigation agent
Q4 2026
Phase 4
The accountability API
H1 2027
00
The Memphis pilot — proof that the pipe works.
● Shipped Q1 2026 Three sites · one brief per site
v1.0 · public

What we proved

The three independent signals — atmospheric chemistry, optical counting, permit paperwork — can be fused into an evidence score that holds up to scrutiny. A 184-day gap is legible. Boxtown knows what the sky already knew.

The Sentinel console exists; the methodology is public; the press kit is written. Ghost Permits works as a finished artefact about a specific case. That's the foundation everything else depends on.

Definition of done

Evidence briefs published for all three Colossus-adjacent sites with reproducible pipeline, open-source code, and cited data provenance.

Shipped

  • Core
    Sentinel-5P + Sentinel-2 fusion. NO₂/HCHO/SO₂ anomaly detection + LFM2.5-VL equipment counting.
  • Core
    Permit cross-reference. TDEC/MDEQ/EPA-ECHO filings joined to coordinates.
  • UX
    Live GIBS map on Sentinel. Six real-time layers: true-color, thermal, aerosol, land temp, NO₂, snow/cloud.
  • UX
    Methodology + Press kit pages. Full reproducible writeup, fact sheet, copy-ready boilerplate.
01
Any coordinate, any layer — the universal console.
● In flight Q2 2026 Sentinel public beta · coordinate-first
Target · Jun 2026

The shift

Stop hard-coding sites. The Sentinel page accepts any input that resolves to a coordinate: lat/lon, address, what3words code, a Google Maps URL, a dragged pin. The full fusion brief runs on whatever the user points at.

Each investigated coordinate gets a memorable three-word tag — the product's what3words-style handle — so a reporter can share /site/ember-kiln-riverbank instead of a raw decimal pair. The tag is the unit of conversation; the coordinate is just what it resolves to.

Compass · concurrentPrivate alpha of the Compass console opens to three design partners at end of Q2. The same pipeline runs backward — enter a region + facility spec, receive a ranked shortlist. No public URL yet; invitation-only.

Success =

A user who has never seen Ghost Permits can paste any lat/lon, get a provisional brief in < 60 seconds, and share the result as a single URL.

Building now

  • UX
    Coordinate search bar. Lat/lon, address, what3words, Google Maps URL, drag-drop KML. Fallback geocoder: Nominatim + Mapbox.
  • UX
    Three-word site tags. Deterministic slug generator (adjective-noun-noun) seeded on coordinates. Readable, shareable, uncollidable.
  • Core
    Dynamic baseline. Pipeline auto-computes the 5-year pre-op baseline for any parcel, not just the three pre-curated ones.
  • Core
    On-demand VLM inference. LFM2.5-VL runs per coordinate; result cached for 7 days. Rate-limited for anonymous users.
  • Data
    Permit lookup by jurisdiction. Reverse-geocode → state/county → correct permit registry. Covers all 50 US states by end of phase.
  • UX
    Provisional verdict card. Score + caveat: "analysis in progress, permit trail pending human review" until a human confirms.
02
Watchlists, alerts, and the inbox the atmosphere writes to.
● Next Q3 2026 Persistent monitoring
Target · Sep 2026

The shift

One-off briefs are a feature; continuous watching is the product. Every dossier gets a “Track this” button. Once tracked, the pipeline reruns daily and the user is notified the moment the anomaly crosses a threshold, a permit gets filed, a news story drops, or a heat signature appears.

Alerts meet users where they already are — not in another tab. Email, SMS, WhatsApp, Telegram, Signal, Slack, Discord webhook, RSS, ICS calendar, or a shareable status page with its own three-word URL.

Compass · alphaFirst three Compass design partners go live. They run region scans on real site-selection projects and give us feature telemetry in return. No public marketing yet — we learn what works before we name a price.

Success =

10,000 tracked sites across all delivery channels, median alert-to-delivery under 90 seconds, < 2% false-positive rate on threshold crossings.

Feature set

  • UX
    "Track this" watchlist. Pin any coordinate to a personal dashboard. Free tier: 5 sites. Supporter: unlimited.
  • UX
    Threshold editor. Per-site rules: NO₂ > baseline + X%, new permit filed, news mention, unit count change.
  • Agent
    Multi-channel delivery. Email, SMS (Twilio), WhatsApp (Meta Business), Telegram bot, Signal, Slack/Discord webhook, RSS, ICS.
  • Data
    Scheduled rerun engine. Daily pipeline invocation per tracked site. Celery/RQ + Redis + S3-backed cache. Cost-capped per user.
  • UX
    Public watch pages. Shareable read-only URL per tracked site: reporters link it in stories, communities bookmark it.
  • Biz
    Identity + quota. Email-magic-link auth, no passwords. Free tier + supporter tier + org tier.
  • Compass
    Candidate-site grid scan. Compass alpha: tessellate a region with 500m cells, run the full 7-dimension score on each, return top N. Region-size quota by tier.
  • Compass
    Grid-infrastructure agent. Substation + HV-line detection from Sentinel-2 × OSM overlay. Estimates available capacity before a utility load study is commissioned.
  • Compass
    Subsidence agent. Sentinel-1 InSAR pipeline: 3-yr mm-scale ground-deformation history, flagged per parcel.
03
The investigation agent — a reporter that never sleeps.
● Future Q4 2026 Agentic research
Target · Dec 2026

The shift

A pin on a map doesn't tell you who owns the land, what's been filed with whom, what the local paper has written, or whether it's already in court. A human reporter spends days on that. An orchestrated agent swarm can do it in minutes — and cite every source.

We run a fleet of narrow agents (see next section), each owning one sub-question. A planner decomposes the user's input into tasks, dispatches, collects, deduplicates, and drafts. Nothing is published without a human pass — but the draft is ready when the human arrives.

This is where AskNews, Perplexity Sonar, and open-web retrieval enter as first-class data streams alongside the satellite feeds.

Compass · GACompass exits invite-only and opens as a paid product. Three design-partner contracts convert; two more hyperscalers and one state economic-development office sign. The money that funds the public tier starts landing.

Success =

A human investigator using Ghost Permits produces a publishable draft brief 5× faster than they do today, with every claim footnoted to a primary source.

Agent capabilities

  • Agent
    Planner / orchestrator. LFM2.5 decomposes "investigate this coordinate" into sub-tasks, routes to specialists.
  • Agent
    AskNews integration. Real-time news index: every press mention of the coordinate, facility name, or operator in the last 5 years with sentiment + entity extraction.
  • Agent
    Court-record agent. CourtListener / PACER / state court crawler. Surfaces active suits, consent decrees, settlements.
  • Agent
    Ownership agent. Parcel → corporate filing → UBO chain. SEC EDGAR + OpenCorporates + local assessor APIs.
  • Agent
    Citation verifier. Fact-check agent re-reads every claim in the draft, flags any statement without a reachable source URL.
  • UX
    Human-in-the-loop console. Drafts arrive as markdown with accept/reject on each claim. The model never publishes; the editor does.
  • Compass
    Comparables agent. LFM2.5-VL finds the 10 most similar operating facilities globally and extracts their cooling / yard-layout / grid-interface signatures from orbit. Benchmark-in-a-pixel.
  • Compass
    Cooling-fit agent. Landsat-8 TIRS + ERA5 reanalysis: surface-temperature profile, heat-island, prevailing wind, water-body distance for liquid/evap cooling.
  • Compass
    Community-sensitivity agent. Reused from the public side. EJScreen, demographics, cancer registries. Whether a Sentinel brief or a Compass ranking, the question is the same: who lives downwind?
04
The accountability API — Ghost Permits as infrastructure.
● Future H1 2027 Platform
Target · Jun 2027

The shift

Once the pipeline produces scored briefs for any coordinate and the agent fleet enriches them, Ghost Permits becomes a piece of critical civic infrastructure — and infrastructure needs an API, a contract, and a business model.

Newsrooms license a streaming endpoint: new anomalies in their beat, delivered as webhook events. Regulators run it in reverse: what permits are missing where combustion is already happening. Law firms pay for depositions-ready evidence packets with chain-of-custody metadata.

We stay free for the public. Power-users pay for the power.

Success =

Three anchor customers (one newsroom, one regulator, one law firm) on paid contracts; public product still operating at zero cost to the end user.

Platform surface

  • Biz
    Streaming anomaly API. Webhook events: "new anomaly", "permit filed", "operator change". JSON + Protobuf. SLAs per tier.
  • Biz
    Bulk retrospective. Every industrial coordinate in a region, scored. Batch export in GeoParquet / GeoJSON.
  • Biz
    Evidence packets. Signed PDF + raw tiles + chain-of-custody hash. Deposition-grade. Priced per packet.
  • Data
    On-orbit inference partnership. LFM2.5-VL deployed to DPhi Space payloads for low-latency edge detection. Direct-to-ground alerts.
  • Agent
    Custom investigation workflows. Org-specific agents: “check every coal plant in West Virginia monthly”, “flag any data-centre permit in the Midwest”.
  • Compass
    Site-intelligence API. Programmatic access to the 7-dimension scoring. Siting consultants embed it; state econ-dev offices license it per parcel.
  • Compass
    Pre-permit feasibility report. Deposition-grade counterpart to Sentinel's evidence packet: the document a permitting authority wants on its desk, answered before the application is filed.
  • Biz
    Open-data tier stays free. Contractually. Non-negotiable. Every commercial tier subsidises the public tier's compute budget.
Agent architecture

Twelve narrow agents, one planner, zero autonomy.

We don't trust a single model to run the investigation — we trust a fleet of narrow ones, each testable, each replaceable, each cited. The planner routes, the specialists fetch, the verifier double-checks, and the human signs. Three agents (Grid, Subsidence, Cooling, Comparables) serve Compass; the rest serve both modes.

01 · Planner
Orchestrator
Decomposes the user request into a task DAG, budgets tokens + wall time, dispatches to specialists, assembles the draft. Uses LFM2.5 locally.
Owns: routing · cost · timeouts
02 · Vision
Scene reader
LFM2.5-VL-450M on Sentinel-2 / Planet / NAIP tiles. Counts equipment, reads construction phase, flags heat signatures.
Owns: what the ground looks like
03 · Atmospheric
Column analyst
Sentinel-5P TROPOMI + MODIS + OMI. Computes baselines, detects anomalies, cross-correlates NO₂ / HCHO / SO₂ / CO.
Owns: what's in the air
04 · Permits
Filing bloodhound
State + federal permit crawlers. EPA ECHO, state agencies, local air quality boards. Joins filings to parcels by fuzzy address match.
Owns: what was (and wasn't) filed
05 · News
Press scanner
AskNews + Perplexity Sonar. Every mention of the coordinate, the facility name, or the operator across 5 years of licensed news, deduplicated, entity-tagged.
Owns: what's been written
06 · Court
Docket watcher
CourtListener, PACER, state e-filing portals. Active suits, consent decrees, settlements, violations. Flags operators with pattern of priors.
Owns: what's in litigation
07 · Ownership
UBO unwinder
Parcel → corporate shell → ultimate beneficial owner. OpenCorporates, SEC EDGAR, state business registries, assessor APIs.
Owns: who profits
08 · Community
Demographics analyst
CDC EJScreen, EPA EJ Index, Census, cancer registries. Who lives downwind, what health burden they already carry, what exposure stacks.
Owns: who pays the cost
09 · Verifier
Citation cop
Reads the assembled draft. Every claim must trace to a specialist's output with a reachable source URL. Hallucinations get struck. Uncited sentences get flagged for human edit.
Owns: what gets published
10 · Grid Compass
Infrastructure reader
Detects substations, HV lines, transformer yards from Sentinel-2. Overlays OSM + utility service-territory data. Estimates available grid capacity before a costly load study.
Owns: whether the grid is ready
11 · Subsidence Compass
Ground-truth sentinel
Sentinel-1 InSAR pipeline. Three-year, millimetre-scale ground-deformation history per parcel. Flags subsiding or unstable ground before foundations go in.
Owns: whether the ground holds
12 · Comparables Compass
Benchmark matcher
LFM2.5-VL finds the 10 most similar operating facilities globally by climate, grid context, industrial density. Extracts their cooling + yard-layout signatures from orbit as a benchmark.
Owns: what “good” looks like
Known risks · mitigation

Being right in public is harder than being right.

Severity: high
The Sentinel ↔ Compass wall.
Watchdog and site-intelligence under one roof is a conflict of interest unless the wall is visible, enforced, and audited. A Sentinel target could be a Compass customer tomorrow.
MitigationHard organisational separation — no shared account managers, no Compass data leaks into Sentinel prioritisation, no Sentinel findings quietly softened for commercial reasons. Public conflict-disclosure page. Independent ethics board reviews every new Compass contract. Published exclusion list of operators we will never sell to.
Severity: high
Defamation / legal exposure from automated briefs.
An agent-drafted claim about a named operator, unreviewed, is a liability event waiting to happen.
MitigationNothing autonomous publishes. Every agent-produced claim sits in a human-review queue with the supporting citation. Corrections page, correction SLA, editorial insurance from Phase 3.
Severity: high
Model hallucination in the agent fleet.
VLM miscounts turbines; news agent fabricates a source; planner invents a court case. Each is a credibility extinction event.
MitigationCitation-Verifier agent (agent 09) re-reads every draft and strikes uncited claims. All source URLs must HTTP-200 at publish time. Ranges, not point estimates, on anything counted from pixels.
Severity: medium
Satellite revisit latency vs. user expectation.
Users expect "live"; Sentinel-2 revisits every 2–5 days, Sentinel-5P daily. Expectations gap = trust gap.
MitigationShow actual tile-capture timestamp prominently on every layer. Animate refresh. Never use the word “live” without qualifier.
Severity: medium
Cost of running the pipeline on 10,000 tracked sites.
Compute + storage + VLM inference at scale. Public tier promises cost-free; infrastructure is not.
MitigationAggressive caching (7-day TTL for scored briefs, 24h for tile composites), scheduled batch reruns, commercial tiers cross-subsidise. Cost-per-site target < $0.04/day at 10k sites.
Severity: medium
Notification delivery ecosystem fragility.
WhatsApp API quota changes, SMS costs, email deliverability. Each channel is a separate integration.
MitigationPluggable delivery adapter. Default to email. Degrade gracefully. Channel availability published on status page.
Severity: low
Competitor clones the methodology.
Our recipe is open. Anyone can rebuild it.
MitigationGood. That's the goal. Our moat isn't the pipeline — it's the watchlist, the trust, the agent fleet, and the archive. Let a thousand Ghost Permits bloom.
Success metrics · 18-month targets

The numbers we'd hold ourselves to, in public.

Sites tracked
10,000
Global. Across all user watchlists by end of Phase 2.
Brief → publication
< 60s
P50. Provisional brief for any coordinate, any user, any time.
Alert latency
< 90s
Threshold → inbox. Any channel, any continent.
Citation coverage
100%
Every published claim. Traces to a reachable primary source.
Correction SLA
< 72h
From operator dispute to visible annotation. Never removal.
Public-tier cost
$0
Always. Funded by commercial tiers, not users.
False-positive rate
< 2%
On threshold-crossing alerts. Measured on 90-day retro audit.
Anchor customers
3
One newsroom, one regulator, one law firm on paid contracts.
Compass design partners
5
By Q4 2026. Hyperscaler, data-centre developer, energy utility, siting consultancy, state econ-dev office.
Compass scan latency
< 4h
Region scan → ranked shortlist. 50×50mi region, 500m cells, all 7 dimensions.
Commercial → public ratio
≥ 3:1
Compass revenue funds 3× the public-tier compute budget. If it dips below, we pause Compass growth — not the public tier.
Open questions · help us answer these

The hard parts we haven't solved yet.

Q01
Where exactly do we draw the Sentinel / Compass wall? Separate legal entities, shared back-end but separate front-ends, or just an ethics policy? Each option trades integrity against leverage.
Owner · Ethics board
Q02
Should tracked-site history be public by default? Open archives build trust but can expose early-stage investigations. Private-until-published is safer but slower.
Owner · Product
Q03
How do we handle a false positive in production? Scrub, annotate, or keep with correction? Each has legal + trust implications.
Owner · Legal
Q04
What's the minimum agent-review depth before publication? A single human pass is cheap but thin. Two reviewers is defensible. Three is slow.
Owner · Editorial
Q05
Do we federate or centralise? One global index vs. regional instances run by local orgs. Federation is resilient but harder to keep consistent.
Owner · Eng
Q06
What's the threshold for auto-notifying downstream regulators? A high-confidence finding could be emailed to the relevant agency automatically. That is either helpful or reckless, depending on who you ask.
Owner · Policy
Q07
How do we price the evidence-packet tier? Per-packet, per-site-subscription, per-org-seat. Each optimises different behaviour.
Owner · Biz
Q08
Which commercial partners are we willing to work with? Clear no's: any operator we monitor, fossil-fuel trade groups. Clear yes's: newsrooms, public-interest law. The grey middle is where we'll get tested.
Owner · Ethics board
What we need · next 90 days

This roadmap only ships with the people building it.

Engineers for the agent fleet, data partners for permit feeds, an editorial lead for the review queue, a legal advisor for the correction policy, and anyone who's been downwind and wants to be early.