Resquites High-Range Tests

Research

Psychometric data, norms and how they are made

Every score on this site comes with the evidence behind it. This page gathers the reliability, norming and validity data of the tests, with their sources and their limits.

.92THETIS marginal reliability (simulation, normal population)
.91THETIS retest correlation, new puzzles (simulation)
.92RUFIT internal reliability (266 valid submissions)
33THETIS puzzles per session, on average

THETIS

THETIS is scored by item response theory: each answer updates a Bayesian estimate of reasoning ability, and the test stops when its standard error reaches 4.5 IQ points (20 to 45 puzzles). From IQ 180 up it climbs one level of 10 points at a time: a level opens only after consistent success at the one below, from more than one puzzle family, and the score never passes the top of the highest level opened (engine 2.2, 3 November 2026). Its item difficulties come from its complexity model of the rules and form a bell curve, deepest from IQ 100 to 190; since bank 1.3 (15 October 2026) the puzzle families all reach the high range, so sessions keep them in balance up to about IQ 190. Bank 1.4 (16 October 2026) added an eighth family, Moorings, a family of deduction, scored since 20 October 2026. Bank 1.5 (20 October 2026) replaces the 824 Oracle's Ledger puzzles that showed a card twice under the same key card, and adds Levers and Lamps levels up to about IQ 257, with which the reporting ceiling rises to IQ 230+. A first calibration on examinees with known IQ scores (7 October 2026) was put on hold the next day: it rested on two examinees' retakes, and it raised the highest scores far too much. The difficulties will be calibrated again on first attempts of at least 10 examinees, and the aim is 30, whose IQ is known from supervised tests: people who report a supervised score in the questionnaire after the test and agree to be asked for proof, and people the author invites; a calibration is applied only if it predicts their known scores better than the current difficulties, each examinee left out in turn. Until then the norms are provisional.

A simulation of 10,000 virtual candidates across seven ability tiers (IQ 40 to 240), run with the server's own engine on the 19,345 scored puzzles of the live bank, gives the precision the engine reaches if the parameters are right: correlation between true and estimated IQ .99, average error 5.3 IQ points, marginal reliability .92 in a normal population sample of 2,000 and a retest correlation of .91 with entirely new puzzles (1,000 candidates, two sessions each).

What the simulation cannot show is validity: whether THETIS ranks people as other good tests do. That needs human data, which every first session now provides (de-identified, and only with consent for research summaries). The technical manual gives the full method and results.

THETIS simulation by ability tier (engine 2.2, bank 19,345 puzzles)
TierIQPuzzlesError (RMSE)95% range covers true IQ
Clinical40–7039.54.296%
Low average70–8533.24.992%
Average85–11531.74.993%
High average115–13031.45.093%
Gifted130–16031.65.292%
Profoundly gifted160–20033.96.690%
Beyond200–24041.810.480%
Actual error stays between 4.6 and 6.9 points from IQ 55 to 200, then widens1,600 simulated sessions at each 5-point step of true IQ (6,400 from IQ 180 to 210, where the precise range ends) · engine 2.2 on the 19,345 scored puzzles
Actual error stays between 4.6 and 6.9 points from IQ 55 to 200, then widensTwo lines over true IQ from 55 to 250. The reported standard error stays near 4.5 IQ points across the range; the actual error stays between 4.6 and 6.9 from IQ 55 to 200, then rises to 18.2 at IQ 250.Precise range: IQ 55 to 20005101520IQ pointsStopping target: SE 4.5918.2557085100115130145160175190205220235250True IQ of the simulated examinee
Table view
True IQReported SEActual error (RMSE)BiasPuzzlesStopped at 4595% range held the true IQ
554.64.6−0.540.125%96%
604.54.6−0.437.510%95%
654.54.6−0.535.45%95%
704.54.8−1.034.53%93%
754.54.7−0.933.62%94%
804.44.8−0.932.72%92%
854.55.0−0.932.22%92%
904.55.1−1.132.42%92%
954.55.0−0.932.52%91%
1004.45.1−1.031.92%92%
1054.44.9−0.930.72%92%
1104.44.9−0.830.21%93%
1154.44.7−0.831.11%94%
1204.45.0−1.231.52%92%
1254.44.9−1.031.62%93%
1304.45.3−1.431.42%90%
1354.44.9−1.331.31%94%
1404.45.0−1.531.52%92%
1454.45.1−1.731.61%92%
1504.45.2−1.831.81%90%
1554.45.4−2.031.82%91%
1604.45.4−2.032.22%91%
1654.45.4−2.032.21%90%
1704.45.6−2.333.13%90%
1754.56.2−2.533.34%89%
1804.56.2−2.533.74%89%
1854.56.3−2.634.15%88%
1904.56.5−2.734.97%89%
1954.56.5−3.036.39%87%
2004.66.9−3.238.114%86%
2054.67.5−3.439.218%84%
2104.78.5−3.940.124%83%
2154.89.7−4.441.131%81%
2204.99.9−4.442.037%82%
2255.011.5−4.843.252%84%
2305.311.9−5.344.174%80%
2355.812.1−6.044.789%79%
2406.713.9−7.844.997%80%
2457.915.0−10.644.999%87%
2509.018.2−14.844.999%89%

The simulated examinees slip on 2 to 10% of answers, more often than the engine assumes after the first five puzzles, so the actual error runs a little above the reported one. Above the precise range the bank holds fewer well-targeted puzzles, so more sessions run to 45 puzzles and the estimate widens.

RUFIT

RUFIT, published in 2021 as the Resquites Wide-Range Test of Fluid Intelligence, has a final norm (6th revision) from 266 valid submissions, ages 16 to 78. Its candidates' earlier proctored IQ tests averaged 145 (range 111 to 174). Raw scores ran from 1 to 26 of 36, with a mean, median and mode of 15 and a standard deviation of 7. Internal reliability: .92.

The item difficulties span the whole range: from items solved by 97.7% of the sample to items solved by 6%. The norm table converts raw scores to IQ 115 (raw 1) through IQ 190 (raw 36); scores above 26 lie beyond the sample and are extrapolations.

Source: Resquites, J. C. M. (2021). Resquites Wide-Range Test of Fluid Intelligence (RWRTFI) and Resquites Universal Genius Test (RUGT): Test invitation. Telicom, 33(4), 38–41. International Society for Philosophical Enquiry.

See the RUFIT norm table
RUFIT items by percent of the norm sample answering correctly
Percent correctItems
90–100%3
70–89%5
50–69%12
30–49%5
15–29%7
under 15%4

PROTEUS and QUADRAVA-FR

Both untimed tests use provisional, model-based norms set out in the author's technical documents, and report every score with its interval. PROTEUS measures most precisely from about IQ 136 to 184 (smallest error about 3.9 points, near IQ 160) and continues to a nominal ceiling of 250. QUADRAVA-FR measures most precisely from about IQ 122 to 173 (smallest error about 2.9 points, between IQ 139 and 156), with a reporting ceiling of 170.

Since 20 October 2026 both are taken online only, untimed, with an instant certificate, so that every result comes from the same administration and the problems stay unseen. Their norms are replaced by empirical norms as first attempts accumulate.

NEREUS and THAUMAS

Both are research editions with four forms each, scored with provisional model norms (NEREUS M0.2, THAUMAS TM0.1): a model with assumed problem difficulties places raw scores 1 to 39 on the IQ scale (NEREUS IQ 103 to 239, THAUMAS IQ 104 to 238), with a model error of about 5.5 IQ points across the middle of the range. No percentile is reported, because no norming sample has been tested yet, and the four forms are built alike but not yet equated.

From 18 October 2026 every first attempt is on a form drawn at random, so that each form collects first attempts from comparable candidates. Once enough have accumulated, the forms are equated on them and the model norms replaced by empirical ones; retakes on new forms give the retest correlation and the practice gain.

GALENE and EIDOTHEA

Two research editions added on 21 October 2026, also scored with provisional model norms. GALENE (model M0.2) draws each attempt's 40 problems from a bank of 400, so every attempt is a new form; its model, which allows for guessing among six choices, places raw scores 7 to 39 between IQ 72 and 248 (model error about 8 to 9 points from IQ 120 to 230), and a perfect score at 258. EIDOTHEA (model EM0.2), with one fixed form, places raw scores 0 to 40 between IQ 81 and 202, with a model error of about 4.5 points across the middle of its range. Every answer key of both tests was proven by exhaustive search over every structure their rules allow, and confirmed by an independent check. Each GALENE problem's success rate in first attempts is recorded, so that the bank can be calibrated item by item.

AMPHITRITE and DYNAMENE

Two adaptive research editions added on 27 October 2026, with no human calibration yet (N = 0). Each attempt opens with one practice problem per family, explained after it is answered and never scored; then the scored problems are chosen one at a time from the answers so far, never a problem the person has been shown before, nor a variant built on the same rule. AMPHITRITE (model AM-M01, a four-parameter logistic model with assumed difficulties over eight construction levels in each of its twelve families) chooses the most informative problem among the families met least, for 40 problems: its rule would end an attempt from the 24th problem once the standard error is at most 4.5 points, but at the model's difficulties 40 problems leave an error of 5 to 6 points in practice (4.5 at the very best), so every attempt runs to the 40th. In a simulation of its engine on its bank (1,500 attempts at each of 27 levels from IQ 55 to 250), its error was about 4.8 to 7.0 IQ points from IQ 55 to 210, and about 9 points from IQ 220, toward its research range of 225+. DYNAMENE follows a staircase over five construction levels for 40 problems, balanced across its six families, and is scored with the difficulties fitted in its simulation study S01 (at its rebuilt levels 4 and 5, their construction targets); in the same simulation its error was about 6.3 to 7.5 IQ points from IQ 70 to 250. Both report the model's estimate with its 95% interval and no percentile. Every answer of both banks was confirmed by an independent solver (AMPHITRITE's 28,160 problems and DYNAMENE's 2,400, retired ones included), and the practice explanations come from a second independent solver whose answers matched every key. First attempts are recorded problem by problem, so that each bank can be calibrated on real responses and its provisional model replaced by empirical norms.

On 28 October 2026 the top of both banks was rebuilt (AMPHITRITE A03, DYNAMENE D0.2). A review had found that their highest construction levels grew in size more than in depth, and that their wrong choices became no harder to rule out: at AMPHITRITE's level 8 the wrong choice closest to the answer still broke two to three conditions on average, and in every role-correspondence problem at levels 7 and 8 at least one wrong choice could be ruled out by counting arrows. AMPHITRITE's levels 7 and 8 in all twelve families, and every level of Recoded worlds and Decisive evidence, now need a step that following the clues one at a time cannot complete: testing a hypothesis, combining conditions that settle nothing alone, or following a structure several steps deep. Their wrong choices are near misses: each breaks only one condition, follows from a single slip, or agrees with the right answer on everything a quick check can see (in Structural equivalence, no command sequence shorter than four commands, or five at level 8, tells a wrong world from the reference). DYNAMENE's levels 4 and 5 have larger worlds, more fits, routes and occurrences to keep in view, rule patterns that differ by a single arrow, and near-miss choices throughout. Two independent solvers confirmed the one answer of every new problem (5,760 for AMPHITRITE, 960 for DYNAMENE). The problems they replace are retired, never given again, and kept, so that every earlier result is scored exactly as it was. The IQ values of the levels remain the model's assumptions until first attempts are analyzed: the rebuild makes the top levels harder by construction; it does not yet measure how hard they are.

PRONOE and PSAMATHE

Two adaptive research editions added on 29 October 2026, with no human calibration yet (N = 0), each from its author's package: PRONOE PN-P01 (numerical dynamics: 896 scored problems in four families of seven construction levels) and PSAMATHE P01-E01 (relational synthesis: 768 scored problems in four families of four bands). Each attempt opens with its package's four public practice problems, the same at every attempt and shown on the test's page, each explained after it is answered; then 40 scored problems, ten per family: the family met least, in turn, and in it one of the fresh problems most informative at the current estimate (the Fisher information of a Rasch item, a random choice among those within 98% of the best, twelve at most), never a problem the person has been shown before. PRONOE is scored by its package's model PN-A01 (Rasch items with assumed difficulties b = −2 to 4 over its seven levels, a standard normal prior, the posterior mean on a grid from −12 to 12); PSAMATHE by PS-A01, the difficulties of its package's model PS-M02 (b = −2, 0, 2 and 4 by band) with the same prior and the same choice of problems, while the package itself gives PSAMATHE as a fixed, balanced form. In a simulation of both engines on their banks (1,500 attempts at each of 20 levels from IQ 55 to 197.5), the error was about 4.5 to 6.7 IQ points (PRONOE) and 4.4 to 6.5 (PSAMATHE) from IQ 70 to 145. Above that, the prior holds the estimates back, by about 7 to 8 points at IQ 160 and 10 to 11 at IQ 175, so a high result is a cautious one; forty right answers place the estimate at IQ 182, and both scales are reported to IQ 180+. Every PRONOE forecast is the same for every set of rules its evidence allows, under an exhaustive check of its stated grammar, and every PSAMATHE key lists all the networks its board allows that meet the task, from an exhaustive check of every network. The answer pattern is checked as on THETIS, against simulated honest sessions (0.8% of honest PRONOE attempts and 1.1% of PSAMATHE attempts are flagged).

Their norms, as those of every adaptive test, are re-estimated from first attempts only, anchored on examinees whose IQ is known from a supervised test: anchor cases the author has checked, and the supervised score an examinee may give at registration (a psychologist's Wechsler or Stanford-Binet scale, Raven's or Cattell's test under supervision, or a high-IQ society's supervised admission test), only from attempts with a clean integrity record. A family's difficulty shift is estimated with each anchor's ability held at the known IQ, and it is applied only when it rests on at least ten anchors and, checked on each anchor it was not fitted on, comes closer to their known IQs than the scale in force.

PANOPEA

An adaptive research edition of figure matrices added in November 2026, with no human calibration yet (N = 0): bank PA3.0 (11 November 2026), 5,129 scored problems in eight families of rules over eighteen construction levels, every one drawn by code from a formal grammar whose five figures all have a direction, so that a turn is always visible. Each problem's eight answers are every combination of its three deciding facts, each right or as the most natural mistake about its rule gives it (a figure followed by its place instead of its shape, two sets put together instead of canceled, a step kept where it doubles), so all three must be found; above level 3 none is a value that stays the same along the row; and a problem's level is computed from what its three facts take to find and the rest of its panels to see through, about ten IQ points a level, from IQ 59 at level 1 to 225 at level 18. Levels 13 to 18 came with PA3.0: four figures round the panel, two or three of them turning by doubling steps of their own; three kinds of figure on the lattice, each moved and then joined, taken away or canceled; and at levels 17 and 18 a rule the lower levels do not have, in which each place of the panel turns whoever arrives in it by its own amount, so that a figure must be followed by what it is while its turn is read from where it lands. From level 15 up only two families, hidden partners and rules on rules, reach the levels. An independent solver, a separate program that sees only the panels, searched every rule of its grammar that fits each problem's rows and columns, under every way of following the figures (by place, by shape, by fill), and a problem was kept only when every rule that fits points to its one answer; a second check, written apart from it, confirmed all of them. No way of counting what the panels show finds the answer more often than chance. The earlier banks (PA1.0, whose problems proved easier than their levels assumed, PA2.0 and PA2.1) are retired, never given again and still scored. Each attempt opens with two practice problems, explained after they are answered and never scored; then 40 scored problems: two at the middle levels from two families, then each the most informative at the current estimate (the Fisher information of a three-parameter item; a random choice among the problems within 95% of their family's best, in the families within 90% of the best), at most eight from one family unless the ladder needs more of its open level, never a problem the person has been shown, nor one made from it by renaming its shapes and fills or turning the whole matrix. Every scored problem has its level's clock, from 70 seconds to four and a half minutes, kept by the server. The model PA-M01 gives every problem its level's assumed difficulty (b = −2.75 to 8.3 logits), a discrimination of 1.5 and a guessing chance of 1 in 8; the estimate is the posterior mean with a Jeffreys prior, the square root of the information of the attempt's own problems, which the choice of problems follows too.

In a simulation of the engine on the bank PA3.0 (2,000 attempts of a normal population): marginal reliability .94 (the squared correlation of the estimate with the true level .94), a retest correlation of .94 on new problems, and no problem given to more than 7% of the attempts near its level. In the same simulation, 300 attempts at each level from IQ 55 to 245, the error was 3.6 to 4.2 IQ points from IQ 65 to 175 and 4.7 to 6.5 from 185 to 215, with the estimate within about a point of the true level over that range. The answer pattern is checked as on the other adaptive tests, against simulated honest sessions (about 1.5% of honest attempts are flagged). Every figure here rests on the model's assumed difficulties and discrimination, and above about IQ 190 on an extrapolation beyond any population: the norms from first attempts will show how far the problems behave as assumed.

A high score must rest on consistent success at the levels below it (the ladder, since 5 November 2026). The levels a score rests on open one at a time, as THETIS's do above IQ 180: since 11 November 2026 a score of 147 or more needs level 10 solved consistently (at least three right answers there, from at least two families, with at least two in three of its problems right), and 160 or more level 11, 180 or more level 13, 190 or more level 14, 205 or more level 16, 220 or more level 17 and 225+ level 18, each very consistently (four right answers, at least three in four right). Until then no problem above the open level is given and the score is held just below the next step. The rule followed an attempt on the first bank that reached level 12 after four right answers at levels 5 to 10 and took 18 of its 24 problems there, so that its IQ 160+ rested on no level between. In the simulation it leaves the accuracy as it was and makes the top cautious: at a true IQ of 225, 19% of attempts reach 225+, at 235 76% and at 245 96%; from IQ 145 up it holds 3% to 19% of honest results just below a threshold until the level under it is cleared. It cannot make an easy level hard: only the bank's construction, and the norms from first attempts, do that.

On the evening of 5 November 2026 AMPHITRITE, DYNAMENE, PRONOE and PSAMATHE got their ladders too, each by one rule from its own levels: a threshold (then 135, 145, 150 and 160, then the societies' criteria of 180, 190, 205 and 220, up to the test's top; since 11 November 2026, the societies' criteria alone, 147, 160, 180, 190, 205 and 220) rests on the highest level whose problems lie at or below it, solved consistently, and from 160 up very consistently (four right answers, at least three in four right). So AMPHITRITE's 150+ rests on its level 6 (IQ 150), its 190+ on level 7 (IQ 185) and its 225+ on level 8 (IQ 225); DYNAMENE's 160+ on its level 3 (IQ 160), its 205+ on level 4 (IQ 205) and its 250+ on level 5 (IQ 250); PRONOE's 145+ on its level 6 (IQ 145) and its 160+ on level 7 (IQ 160); PSAMATHE's 160+ on its top band (IQ 160). In simulations of their engines (2,000 attempts of a normal population and 400 at each level) the ladder left their accuracy as it was: the correlation of the result with the true level .945 for AMPHITRITE and PRONOE, .944 for PSAMATHE and .911 for DYNAMENE (.944, .945, .944 and .905 without it). It makes the top more cautious: at a true IQ of 160, 55% of AMPHITRITE's attempts reach 160+ (55% without it), 38% of DYNAMENE's (52%), 5% of PRONOE's (8%) and 3% of PSAMATHE's (8%), whose prior already holds high estimates back; and it holds up to 7% (AMPHITRITE), 14% (PSAMATHE), 17% (PRONOE) and 18% (DYNAMENE) of honest results just below a threshold until the level under it is cleared. What it adds is a guarantee, that no result above a threshold stands without consistent success at the level below it. It cannot make an easy level hard: only the norms from first attempts can find such a level.

NEMERTES

An adaptive research edition added in November 2026, with no human calibration yet (N = 0): the decisive experiment. Bank NM1.2 (14 November 2026; NM1.1 with its levels 15 and 16), 1,016 scored problems and four practice problems over sixteen construction levels (IQ 70 to 225), in four families, the four kinds of machine (change in place, disappear, move, copy). Every problem shows two machines whose rules form a minimal pair, a few examples of each and a start board on which both agree; the examinee constructs the one change that makes them differ (at levels 15 and 16, two changes at once, on two cells, where no single change does), and every change that does so is right. Each problem is built by code from a formal grammar of 41,200 rules, and an independent solver, written separately and seeing only what the problem shows, proved for every problem that every rule of the grammar consistent with each machine's examples gives the same right changes, and that they are the key's. The difficulty model NM-D04 places each problem within five points of its level's IQ from how it is built: the rules, how hidden their difference is, how precise the change must be, how many steps ahead the machines must be followed, and the size of the board. Scores come from the model NM-D05: each problem at its level's assumed difficulty, with its own guessing floor (the share of all possible changes that are right: about a third at level 1, under one in a hundred from level 8), with a Jeffreys prior; the four families take turns, ten problems each, each the most informative fresh problem at the current estimate, and the ladder holds a high score until the levels below it are solved consistently.

In a simulation of the engine on the bank NM1.2 (4,000 pairs of attempts of a normal population, whose examinees guess as often as chance allows): marginal reliability 0.85, a retest correlation of 0.86 on new problems, and an error of about 5 to 7 IQ points from IQ 70 to 210, with no bias beyond about a point and a half from IQ 80 to 200. The top is cautious: estimates are held back by about 3 points at IQ 210 and 5 at IQ 220 (error about 9), and no simulated attempt, even at a true IQ of 230, reached 225+. The scale is reported up to IQ 225+, the level of the hardest problems released (levels 15 and 16, forty problems each).

Validity studies in progress

  • Concordance. Candidates may report earlier proctored scores (WAIS, Stanford–Binet, RAPM and others); these are used to link the scales. The adaptive tests ask for a supervised score in a form of its own, and re-norm on it from first attempts.
  • Retest. THETIS retakes (never with a puzzle seen before), NEREUS and THAUMAS retakes (each on a form never seen) and GALENE retakes (each on problems never shown) give retest data on new problems: the test–retest correlation and the practice gain, by the time between sessions.
  • Calibration. The fit of every puzzle family and level is estimated from first attempts, with each examinee's ability held fixed, and reviewed as data come in (snapshots at 30, 50, 100, 200, 500 and 1,000 first attempts). Once at least 10 examinees with supervised scores have taken THETIS for the first time, item difficulties are calibrated on them; after about 500 first sessions, they are re-estimated from all human answers and the provisional parameters replaced.
  • Fairness. Differential item functioning by first language and by device (phone, tablet or computer) is computed from first attempts by the Mantel–Haenszel procedure; by sex as questionnaire data allow.

Take part

Every first attempt helps the norms. If you consent, your de-identified result may appear in published research summaries; names, emails and certificate IDs never do. Researchers interested in the data may write to jresquites@mensaphilippines.org.

Have you taken a professionally given IQ test, such as the WAIS or the SB5? Your first attempt at THETIS or a research edition can anchor the calibration of its problems, for half the fee (PANOPEA at its fee): help norm the tests.