Private / unlisted project page

LookingGlass: a generative model of future world events.

LookingGlass is a time-aware AI system that learns from the historical global news stream to forecast and simulate future political, financial, conflict, infrastructure, and security events with explicit uncertainty, evidence trails, and later scoring.

Illustrative visor query

Enter a date window: July 12 to July 20.

Output A / 40% / biological-security incident / fictional location

July 12, 06:40 local: the first weak signal is not a headline. It is a one-paragraph notice from Port Meridian General asking non-emergency patients to use two inland clinics. By noon, the waterfront school district cancels outdoor activities with no explanation. The mayor says there is “no confirmed public danger,” but the phrase appears across three agencies with identical punctuation, suggesting a coordinated message rather than routine caution.

July 14: a regional broadcaster reports that three harbor workers were transferred out of county under sealed medical protocols. The report is deleted, then mirrored by anonymous accounts. July 15: satellite vans gather outside the old customs building; police close two roads; officials deny a chemical spill, which makes the absence of a normal industrial explanation more conspicuous. July 17: national health and security personnel arrive separately, then brief together. LookingGlass forecasts the likely public headline as: “Federal agencies investigate suspected biological-security incident near Port Meridian.” The model does not infer method, actor, or capability. It predicts the story-shape: confusion, perimeter, denial, interagency arrival, then controlled disclosure.

Output B / 64% / coordinated port disruption / medium evidence

July 12: Dockworkers Local 88 rejects a narrow wage offer in Grayhaven, but the model flags a stranger signal: two shipping insurers quietly add delay surcharges before any strike vote. July 13: a refrigerated cargo operator diverts three vessels to North Sable even though Grayhaven remains officially open. July 14: warehouse managers begin texting night-shift workers not to report until called.

July 16: the visible story becomes labor, but the forecast says the deeper storyline is synchronization. Crane crews in Grayhaven, customs clerks in North Sable, and inland truckers all slow at once without any single group declaring a full stop. Produce spoils in outer lots. A medical supplier asks for emergency routing. By July 19, national officials call it “contained,” but grocery chains begin rationing two categories of imported goods. LookingGlass forecasts a negotiated pause by July 20, not because the dispute is solved, but because the disruption becomes visible enough to threaten broader market confidence.

Output C / 27% / cyber escalation / weak-to-medium evidence

July 12: a sanctions speech in West Ardent uses the phrase “systems-level consequences,” which the model links to prior escalation cycles. July 13: transit-card payments fail for nineteen minutes in the capital. July 14: a water authority posts that “all safety functions remain independent,” a line LookingGlass marks as abnormal because no one had publicly asked about safety functions.

July 16: a private grid operator says an outage was caused by “vendor authentication drift,” then removes the phrase. July 17: West Ardent’s foreign ministry warns of “responses below the threshold of armed conflict.” The forecast branches from there. In the low-severity branch, agencies bury the incident as routine infrastructure noise. In the darker branch, a second outage lands during a televised cabinet meeting, insurance markets reprice regional cyber exposure, and two allied governments issue synchronized travel advisories. LookingGlass assigns only 27%, but marks the narrative as dangerous because each denial increases the cost of admitting coordination later.

These are example outputs, not active forecasts. The point is the format: date range, event family, probability, location scope, evidence strength, possible storyline, branches, and later scoring.

Development timeline

Current phase: data / research

Signal collection
Data / research
Training
Forecast trials
Deployment review

LookingGlass is not in model training, deployment, or operational use. The present work is dataset construction, source review, event taxonomy design, and research validation.

militaryintelligencewarfaresocietal determinismapocalyptic riskcivil unrestfinancial shockinfrastructure failuresecurity escalationtruth-seeking

Project overview

Most AI systems can explain what has already happened. LookingGlass asks a harder, more accountable question: given everything publicly known at a specific time, what is likely to happen next?

The prototype ingests time-stamped global reporting and structured public signals, builds a representation of ongoing situations, and rolls that state forward into detailed, falsifiable future trajectories.

High-consequence technology

If the core mechanism works, LookingGlass should be treated as a potentially class-5 threat-level technology: a system that could materially change how institutions anticipate geopolitical, military, financial, infrastructure, and security events.

The potential applications include defense, military planning, intelligence analysis, infrastructure risk, market monitoring, diplomacy, and emergency preparedness. That same power is why the project must be designed around evidence, calibration, audit trails, uncertainty, and refusal to claim certainty.

Name and myth

The name intentionally echoes the old Looking Glass / Chronovisor mythology: the idea of a machine that could look across time. The project does not treat that mythology as true; it uses it as a cultural reference for a modern, empirical version of the same desire.

The relevant legend is the Vatican-adjacent Chronovisor story, generally regarded as pseudoscientific. LookingGlass is the opposite kind of claim: no mysticism, no hidden certainty, only timestamped public evidence and forecasts that can be scored later.

Architecture

1Time-stamped public reporting
2Structured event ledger
3Situation representation
4Forecast models
5Scenario generator
6Evidence and scoring layer

Forecasts are generated from structured model outputs first, then turned into human-readable scenario briefs with cited evidence and cutoff dates.

Truth-seeking alignment

LookingGlass treats future claims as accountable objects. Every forecast has a timestamp, evidence bundle, uncertainty estimate, and eventual score. The goal is to make AI-assisted reasoning about the future less like persuasive autocomplete and more like testable empirical forecasting.

What will be measured

  • Calibration over 1-, 7-, and 30-day forecast windows
  • Improvement over simple historical and persistence baselines
  • Evidence quality and cutoff-time integrity
  • Uncertainty exposure when the model is wrong or underspecified
  • Usefulness of generated scenario briefs for human review

What LookingGlass will not claim

LookingGlass will not claim deterministic prediction, omniscience, or certainty about rare shocks. It will not claim to know hidden intent, classify people as future threats, or replace human judgment in high-stakes decisions.

The prototype measures whether time-safe public signals can improve calibrated forecasts over baselines while making its uncertainty and failures visible.