Looking Glass An evidence dossier
Page 8 · a working simulator of its own limits

The Looking-Glass Engine

This does exactly what the claim describes: model futures, write an intervention, nudge the trajectory. It runs on synthetic toy dynamical systems — no real-world data feed, no forecast about anything real. Its purpose is to let you watch precisely where the method stops working, and to feel how little extra foresight enormous gains in measurement actually buy.

200-member stochastic ensemble deterministic chaotic dynamics runs entirely in your browser no data feed
Why the honesty here is structural, not a badge

The central finding of this dossier is that a disclaimer does not survive redistribution. Bill Wood’s “that book, of course, is fictional” was in the original footage at 61 seconds and was cut anyway. So this simulator does not rely on a label to make it safe. The mechanism itself is the disclaimer: every run computes the point at which its own output carries zero information and refuses to pretend past it. A stripped label leaves a lie. A stripped mechanism leaves nothing.

The removal, documented: provenance findings §5 · the removal finding

Run an ensemble

Each scenario is a synthetic coupled system with genuinely deterministic dynamics — the same equations, run 200 times from measurements that differ only within your stated precision. Nothing is random about the physics. The divergence you see is entirely the cost of not knowing the initial state exactly.

Controls

A synthetic toy dynamical system. Not a data feed, not a model of any real region, market or population.

Normalised index value at t = 0.

Per-step process noise: shocks the model does not represent.

The key control. Every extra decimal place of measurement buys the same small, fixed number of extra days — never a proportional gain.


A small push on the index, applied once.

Runs are deterministic: same controls, same seed, same output.

Predictability horizon
Spread at horizon
Nudge inside horizon
Nudge beyond horizon

The horizon is the first day on which the 5th–95th-percentile spread of the ensemble exceeds 90% of the system’s own full range. Past that line the forecast is not weak; it carries zero information — the ensemble says the index will be somewhere between its minimum and its maximum, which is what you knew before you started.

The whole lesson, in one table

What better measurement actually buys

Computed live from the scenario you have selected, holding everything else fixed. Each row is a thousand-fold improvement in measurement precision over the row above it. Every figure is the mean of six independent ensembles — a single run's horizon is itself a noisy statistic, and reporting one draw would manufacture a trend the model does not support. These runs also switch the scenario's process noise off, so measurement precision is the only source of uncertainty. Leave the volatility in and random shocks swamp the initial-condition error completely: precision then buys nothing at all, which is the harsher version of the same lesson.

Predictability horizon as a function of measurement precision
Measurement precision Horizon Gain over previous row

Improving measurement does not multiply your foresight. At best it adds a fixed handful of days each time you multiply precision by the same factor, and in stiffer scenarios it adds nothing measurable at all — the horizon is then set by the dynamics rather than by the instrument. That is what “logarithmic” means in practice, and it is why no budget and no classification level produces a device that shows the world’s trajectory. The wall is arithmetic, not secrecy.

Output

The forecast, stamped and falsifiable

Every forecast this engine emits carries an expiry date and an explicit falsifier: the observation that would prove it wrong. That is non-optional — a forecast without a falsifier is not a forecast, it is a mood.

Honest artifact · as generated

Run the ensemble to generate a forecast.

    Calibration

    Anchored to published benchmarks

    The toy model above is fiction. These numbers are not. They are what real forecasting systems achieve, and they are the reason the horizon in the simulator sits where it does.

    ~2 weeks

    Weather’s hard ceiling

    The best-funded, most data-rich, most mathematically mature predictive system humans have built degrades to unusable at roughly two weeks, because error compounds under chaotic dynamics. Human and geopolitical systems are at least as complex and vastly less instrumented.

    DTIC: “Predictive Analysis: An Unnecessary Risk in the Contemporary Operating Environment” · Lawrence Livermore CGSR

    400 days vs 80 days

    Tetlock’s superforecasters

    Philip Tetlock found “superforecasters were able to see about as accurately 400 days out as regulars were about 80 days out” — a fivefold advantage from elite structured judgement, with expert political forecasting decaying toward chance around five years.

    Edge.org, Tetlock master class

    4 of 16

    ICEWS on the Arab Spring

    DARPA’s Integrated Crisis Early Warning System, checked against the real Arab Spring, correctly flagged 4 of 16 relevant events within the same quarter and 9 of 16 within the year. The contractor had claimed “greater than 80 percent accuracy” in testing.

    Wired, 2011 · real programs §5.3

    days, not minutes

    GIDE’s own framing

    Gen. Glen VanHerck on the Global Information Dominance Experiments: “I’m talking not minutes and hours, I’m talking days”, and “the ability to see days in advance creates decision space.” Days of warning and more options — not certainty, and not a timeline.

    Computing UK · northcom.mil