Insight: Declassified research

Star Gate remote viewing models: a technical guide

Four models from the declassified Star Gate files, rebuilt from the original pages, explained step by step and run in your browser.

Published 8 November 202611 minute read

A typed letter on a clipboard beside a pen and a film reel
504

Missions tasked to SRI and the Fort Meade unit, needing 2,865 separate viewings.

The Star Gate Archives, vol. 4, quoted in Broderick (2020), p. 358

17 of 19

Client agencies from 1973 to 1995 that came back with more tasking.

The Star Gate Archives, vol. 4, quoted in Broderick (2020), p. 358

1995

The year the programme closed, when the CIA judged its results insufficiently reliable.

Broderick (2020), p. 357

What Star Gate was

From 1972 to 1995 the United States government funded research into remote viewing, the claimed ability to describe a distant or hidden place by mind alone. The research ran at SRI International and later at SAIC. An operational unit at Fort Meade, Maryland, took tasking from intelligence agencies, with the Defense Intelligence Agency as the main funder. The effort carried several code names, among them Grill Flame and Center Lane, and was renamed Star Gate in 1991.

The documents describe more than sessions. They set out models: when results were said to be better, how a transcript was scored, which targets were thought easier to describe, and how much information a channel was thought to carry. This guide rebuilds four of those models from the original pages, so you can see exactly what each one computes and where it stops.

Read this firstThe instruments reproduce what the documents describe. Their outputs are model values. They measure nothing about the mind and are not evidence that any effect exists. The sidereal peak behind the first instrument was not found again in a later study of 3,678 trials.

How to use this guide

Each of the four sections follows the same five steps.

  1. The model. What the source says, with its document and page.
  2. The maths. The equations the engine computes, and the parts that are the engine's own choices.
  3. The instrument. Change the inputs and every output updates. Nothing you type or load leaves your browser.
  4. The worked example. The same numbers by hand, so you can check the code.
  5. The limits. What the model cannot tell you.

Sidereal window: 13.5 hours LST

The model

Spottiswoode (1997) pooled free-response trials and reported the effect size up 340% for trials within one hour of 13.5 hours local sidereal time (1,468 trials, p = 0.001), and up 450% in a second set of 1,015 trials (p = 0.05), "assuming that some unknown systematic bias is not present in the data".

Sidereal time follows the Earth's turn against the stars, not the Sun. A sidereal day lasts 23 hours 56 minutes 4 seconds, so a given sidereal time comes round about four minutes earlier each day by the clock.

The maths

JD = t / 86,400,000 + 2,440,587.5t in milliseconds since 1970, UTC

d = JD − 2,451,545.0 and T = d / 36,525days and centuries since J2000.0

GMST = 280.46061837° + 360.98564736629° × d + 0.000387933° × T² − T³ / 38,710,000°Greenwich mean sidereal time, degrees

LST = ((GMST + λ) mod 360°) / 15local sidereal time, hours; λ is longitude, east positive

gap = min(|LST − 13.5|, 24 − |LST − 13.5|)hours from the target, the short way round

open when gap ≤ 1the window

next = ((12.5 − LST) mod 24) / 1.00273790935clock hours until it opens

The sidereal time formula is equation 12.4 of Meeus, Astronomical Algorithms. The window index beside the dial is a scheduling scale (100 at 13.5 hours, 50 at the window's edge, falling to a floor of 10): the engine's convention, never a measurement.

Worked example

Meeus's example 12.b: on 10 April 1987 at 19:21 UT the mean sidereal time at Greenwich is 8 h 34 m 57.09 s. The engine returns the same value to the hundredth of a second.

How to use it

  1. Choose a place, or type a longitude. Latitude does not change sidereal time.
  2. Read the dial: the marigold arc is the window and the hand is the local sidereal time now.
  3. Read the table for the next seven windows, in the place's own time zone.
  4. Change the target or the width to explore other claims, such as the smaller peak near 7.5 hours.
Local sidereal time
--:--:--
Status
Outside
From the target
--
Window index
--A scheduling scale, not a measurement

The next seven windows
DayOpensCloses

The limits

A study of 3,678 new trials found no peak at 13.5 hours, only a small one near 7.5 hours (Spottiswoode and Ryan, 2015). Treat a window as a time to test the claim, with the failed replication in mind.

Transcript scrub: shape before names

The model

The Grill Flame report found that non-analytical data (shape, form, configuration, patterns) "are generally of higher quality than analytical data (name, function, labeling)" (Table II, p. 12). Viewers' naming and guessing came to be called analytical overlay. Results were rated from 0 to 3 (Table 1, p. 11), and concepts were scored against a target list with relevance from trivial (1) to key (5) (pp. 33 to 34).

The 1991 SAIC plan defines accuracy as the share of the target described correctly, reliability as the share of the response that is correct, and the figure of merit as their product. It warns that "an absolute quantitative evaluation of an individual figure of merit is not yet possible" (pp. 38 to 39).

The maths

F = overlay terms foundnaming, function, hedging

P = sensory primitives foundshape, texture, dimension

S = max(0, 100 − 15 × |F|)sanitisation, %

T = {(e_i, r_i)}, r_i from 1 to 5target elements and their relevance

M = elements of T in the textwhole words only

A = Σ r_i over M / Σ r_i over Taccuracy

R = |M| / |P ∪ F ∪ M|reliability, never above 1

FM = A × Rfigure of merit

band 0 to 3FM above 0.05, 0.25 and 0.45

The word lists, the 15-point penalty and the band edges are the engine's own. Grill Flame's analysts judged each concept by hand; here a whole-word match counts as correct and anything else as missing.

Worked example

The sample transcript holds six overlay terms (base, facility, looks like, missile, seems to be, silo) and six sensory primitives (45 meters, cold, cylindrical, gray, metallic, rock). Four of the five target elements appear: cylindrical (5), metallic (4), underground (4) and gray (2). Concrete (3) does not.

A = (5 + 4 + 4 + 2) / (5 + 4 + 4 + 3 + 2) = 15 / 18 = 0.833accuracy

R = 4 / 13 = 0.30813 distinct primitives: six sensory, six overlay, underground

FM = 0.833 × 0.308 = 0.256band 2, "A hit"

S = 100 − 6 × 15 = 10sanitisation, %

How to use it

  1. Paste a transcript, or keep the sample.
  2. List the target elements, one per line, each with a relevance from 1 to 5.
  3. Read the marks: struck words are overlay, underlined words are kept, marigold words match the target.
  4. Read the scores, then restore the sample to start again.

Marked transcript

Overlay struckSensory keptTarget matched

Sanitisation
--
Accuracy
--
Reliability
--
Figure of merit
--
Rating, 0 to 3
--
Words
--

Overlay struck

    Sensory kept

      Target elements missed

        The limits

        The engine sees only words and cannot read meaning. A description can match a target by chance, and a careful viewer who avoids the listed words scores well whatever they describe. The SAIC plan's answer was to score a response against a whole pool of possible targets and rank the true one (p. 39), which a single list cannot do.

        Target entropy: change you can measure

        The model

        The 1991 SAIC plan calls the change of entropy "the most promising target aspect" for predicting how well a target is described (p. 28), and writes it as the spatial change plus the temporal change (Equation 1, p. 40). Early calculations for a few video targets showed "a crude correlation with AC quality", in line with a reported preference for moving video over still photographs (pp. 28 to 29).

        The maths

        The plan's Equation 1:

        ΔS = ∫ ∇S · dr + ∫ (∂S/∂t) dtspatial change plus temporal change

        Measured from a picture, the engine makes it discrete. The picture is turned to grey, scaled to at most 256 pixels on its long side and cut into 16-pixel tiles.

        grey = 0.299 R + 0.587 G + 0.114 BITU-R BT.601 luma

        H_k = − Σ p_g log₂ p_gShannon entropy of tile k, 0 to 8 bits

        S_xy = mean |H_a − H_b|over tiles sharing an edge: the spatial term, bits

        S_t = mean |H_k(after) − H_k(before)|over all tiles: the temporal term, bits

        ΔS = S_xy + S_ttotal change, bits

        Judged instead of measured, the engine keeps the simpler scale it was first given: your 0 to 10 judgements, a lower weight for still targets and an index from 0 to 100.

        ΔS′ = s + w × tw = 1.5 for a moving target, 0.25 for a still one

        V = 100 × (1 − e^(−0.35 × ΔS′))index from 0 to 100

        Worked examples

        A flat grey tile carries 0 bits. A tile split evenly between black and white carries exactly 1 bit. A tile that uses each of the 256 greys once carries 8 bits, the most a tile can hold. Judged, a spatial change of 4 and a change over time of 3 for a moving target give ΔS′ = 4 + 1.5 × 3 = 8.5 and an index of 94.9.

        How to use it

        1. Load a picture for the spatial term, and a second frame of the same scene for the temporal term. Or use the sample frames.
        2. Read the dot grid: each dot is a tile, sized by its entropy.
        3. To compare with the plan's own scale, set the judgement sliders below it.

        Measure a picture

        Pictures are read in your browser and never sent anywhere.

        Whole picture
        --
        Mean tile
        --
        Spatial term
        --
        Temporal term
        --
        Total change, ΔS
        --
        The picture, in grey
        Tile entropy, 0 to 8 bits

        Judge a target

        Judged change, ΔS′
        --
        Index, 0 to 100
        --

        The limits

        Entropy measures the information in the picture, nothing about anyone's perception of it. The plan called its correlation crude and based it on a few video targets, and a still photograph has no temporal term at all.

        Channel: bits and beats

        The model

        A 1975 Army medical intelligence survey for DIA reports Soviet experiments by Kogan in which "the rate of telepathic information transmission varied between 0.005 and 0.1 bit/sec", falling with distance "from 0.1 bit/sec for a distance of several meters to 0.001 bit/sec for a distance of 4,000 km" (p. 26). The same page notes that only "qualitative images eliciting some kind of sensation (shape, color, hardness)" were perceived, the point Grill Flame's Table II makes eight years later.

        The 1983 Army assessment of the Monroe Institute's Gateway training explains Hemi-Sync: each ear hears a different tone and the brain hears the difference, the beat (p. 8). It cites Bentov's model of the body resonating at 6.8 to 7.5 Hz, and the ionospheric cavity at about 7 to 7.5 Hz with a wavelength of about 40,000 km (p. 9). A 1984 science panel, appended to the 1993 Star Gate overview, recorded an estimate of "about one bit per thousand" above noise (pp. 24 to 25).

        The maths

        k = log(0.1 / 0.001) / log(4,000 km / 0.01 km) = 0.3570the slope on log scales

        R(d) = 0.1 × (d / 0.01 km)^(−k)bit/s, held between 0.001 and 0.1

        byte = 8 / R(d)seconds per byte

        f_left = f_carrier − beat / 2left ear

        f_right = f_carrier + beat / 2right ear

        The straight line between Kogan's two points on log scales, and reading "several meters" as 10 metres, are the engine's assumptions.

        Worked example

        With k = 0.3570, a distance of 600 km gives R = 0.1 × 60,000^(−0.3570) = 0.00197 bit/s, so one byte takes 8 / 0.00197, about 4,064 seconds or a little over an hour. A 200 Hz carrier with a 7.15 Hz beat plays 196.425 Hz in the left ear and 203.575 Hz in the right.

        How to use it

        1. Slide the distance from 10 metres to 20,000 km.
        2. Set the carrier and the beat. The default beat sits in the middle of Bentov's range.
        3. Play the pair through headphones at low volume to hear what a binaural beat is.

        Headphones, low volume. Stop if it is uncomfortable. The tones are made in your browser.

        Kogan rate
        --
        Time per byte
        --
        Left ear
        --
        Right ear
        --

        Body resonance (Bentov)
        6.8 to 7.5 Hz
        Ionospheric cavity
        About 7 to 7.5 Hz, wavelength about 40,000 km
        Science panel, 1984
        About one bit per thousand above noise
        Monroe, as quoted
        Beta signals "around 2877.3 CPS", where Beta is normally 30 to 40 CPS

        The limits

        Kogan's figures are reported results from Soviet experiments, summarised second hand; the curve between them is the engine's assumption. The tones show what a binaural beat is. The resonance figures are the authors' models, and the Monroe figure, quoted from a magazine on p. 25 of the Gateway assessment, is on a page missing from the CIA's online copy.

        How the engine is checked

        • Sidereal time against Meeus's worked example 12.b.
        • A second, independent implementation of every model in Python. The browser engine matches it on 151 cases: times, places, transcripts, judged entropies and distances.
        • Entropy against constructed pictures with exact answers: 0, 1 and 8 bits per tile, and spatial and temporal terms of known size.
        • This page. Every worked example above is recomputed by the test suite on each build, so the numbers here cannot drift from the code.

        Sources

        Government documents are in the public domain and in the CIA's online reading room. Page numbers are the PDF's own.

        DocumentUsed for
        Soviet and Czechoslovakian Parapsychology Research, US Army Medical Intelligence and Information Agency for DIA, September 1975, CIA-RDP96-00792R000600350001-3Kogan's rates, p. 26
        Analysis and Assessment of Gateway Process, US Army INSCOM, 9 June 1983, CIA-RDP96-00788R001700210016-5Beats, p. 8; Bentov, p. 9; Monroe, p. 25 (released separately)
        The Gateway Program, Monroe Institute, undated, CIA-RDP96-00788R001700270006-0The institute's account of binaural beats
        Grill Flame Project Report, DIA, 19 October 1983, CIA-RDP96-00788R001000410001-6Rating scale, p. 11; Table II, p. 12; concept analysis, pp. 33 to 34
        A Comprehensive Research Plan for Anomalous Mental Phenomena, May and Luke, SAIC for DIA, 7 June 1991, CIA-RDP96-00789R003100080001-9Entropy, pp. 28 to 29 and 40; figure of merit, pp. 38 to 39
        Star Gate Project: An Overview, DIA, 30 April 1993, CIA-RDP96-00789R002800180001-2The 1984 science panel, pp. 24 to 25
        Meditation investigation, Science News, 15 December 1973, CIA-RDP79-00999A000200010091-5Context: a magazine issue the CIA filed
        Soviets disband Krishna group, Philadelphia Inquirer, June 1982, CIA-RDP90-00806R000100460001-9Context: a press clipping the CIA filed
        Essay review: The Star Gate Archives, volumes 1 to 4, Broderick, Journal of Scientific Exploration 34(2), 2020, pp. 355 to 363 (CC BY-NC)Programme figures, pp. 357 to 358
        Apparent association between effect size in free response anomalous cognition experiments and local sidereal time, Spottiswoode, Journal of Scientific Exploration 11(2), 1997, pp. 109 to 122The 13.5 hour peak
        Spottiswoode and Ryan, in May and Marwaha (eds), Extrasensory Perception: Support, Skepticism and Science, vol. 2, Praeger, 2015The failed replication
        Meeus, Astronomical Algorithms, 2nd edition, Willmann-Bell, 1998Sidereal time, eq. 12.4 and example 12.b

        FAQ

        Quick answers

        Anything else? Email contact@sovereignsystemslabs.com.

        Does this show that remote viewing works?

        No. The instruments reproduce the models in the documents, and their outputs are model values. The programme closed in 1995, when the CIA judged its results insufficiently reliable.

        Why 13.5 hours sidereal time?

        It is the peak Spottiswoode reported in 1997 from 1,468 trials. A later study of 3,678 trials did not find it, so the target is editable and both results can be explored.

        Is anything I type or load sent anywhere?

        No. The instruments run in your browser. Transcripts, target lists and pictures stay on your device.

        Where can I read the original documents?

        Every government document is in the CIA's online reading room, linked in the sources table. Page numbers on this page are the PDF's own.