N-TEK · Evidence
What each figure is.
Every figure on this site is one of four kinds: a published result, an N-TEK measurement, a design projection or a simulation. This page says which is which, and how each was measured.
Four kinds of figure
- PublishedIn a public paper: N-TEK’s EC-AIF preprint on Zenodo, or the paper a comparison comes from.
- MeasuredRun by N-TEK, on a public benchmark under its own protocol or in N-TEK’s own pre-registered tests. Not yet rerun by anyone else.
- ProjectedFrom N-TEK’s design analysis. NOGAI Mk. II has not been built or flown.
- SimulatedFlown in N-TEK’s own simulator, on a design-stage model of the aircraft.
NEXMIND on Gameworld 10k
| Figure | Kind | How it was measured |
|---|---|---|
| 115.6ten-game mean · 95 % interval 106.9 to 123.5 | Measured | The ten unmodified games under AXIOM’s protocol, on ten seeds never run before, each run once on a version frozen and registered beforehand. One setting for all ten games, each learned online from scratch, on one CPU core. The score, game by game · every run’s score and time. |
| 71.2 · 93.5 · 2.1AXIOM · AXIOM tuned for each game · DreamerV3 | Published | Table 1 of the AXIOM paper, run by its authors on one A100 GPU per game: about 30 minutes a game for AXIOM, about 2.5 hours for DreamerV3 (its Appendix C). N-TEK did not rerun them on its seeds; it did run AXIOM’s released code on one or two seeds per game, and those runs landed within the paper’s reported ranges on all ten games. |
| 127.6ten-game mean on development seeds | Measured | The current mind, on N-TEK’s development seeds: above what the mind behind 115.6 scored on the same seeds. Development seeds are not held out, so the published result stays 115.6. |
| 0.68 · 0.66bsuite mean · NEXMIND and DreamerV3 | Measured | DeepMind’s bsuite. Early and provisional: a hint, not a claim. |
Not built for Gameworld. It is built for deployment: the same mind flies NOGAI in simulation. It was developed by algorithm-level changes, tried on Gameworld’s development seeds, on bsuite and in flight, and kept only when they carried across environments without a regression. Nothing was tuned per game, and no change was made to chase a score.
Protocol: Gameworld 10k from its public repository at commit 85493c4; 10,000 environment steps per game, rewards summed, resets not counted, the loop checked against AXIOM’s released code. One difference: NEXMIND also receives the episode-end signal, which AXIOM’s model does not use. Runs are deterministic: the same version on the same seed gives the same reward at every step. A game took a median of 12 minutes (4 to 28), timed with several runs sharing one desktop CPU, so on a core of its own it takes no longer.
EC-AIF
| Figure | Kind | How it was measured |
|---|---|---|
| 20 of 20fresh seeds improved | Published | PPO with Achievement Distillation’s published settings on symbolic Craftax-Classic, 1 million steps, with and without EC-AIF’s intermediate goals, on paired seeds 101 to 120 fixed before the runs (registered at N-TEK commit 5e6517ff), with registered controls. The gain is measured against N-TEK’s own control, which sits below the best published PPO at that budget. The paper’s Table 1. |
| up to 8.7×for standard active inference | Published | Standard active-inference agents on DeepMind’s bsuite catch, the event layer added unchanged, ten paired runs per configuration, registered before the runs (commit d8673f4). The gain is largest for the shallowest planners: a one-step planner rises to the level a four-step planner reaches without it, and a planner that already scores high gains nothing. The paper’s Table 1. |
| 1.9×NEXMIND with the event layer, against itself without | Measured | An N-TEK internal, pre-registered test. It is not in the paper. |
The paper is a preliminary report, open access under CC BY 4.0, not yet peer reviewed. Its code is not public, and its registrations are records in N-TEK’s own repository, not third-party timestamps. EC-AIF draws on event segmentation theory and is related to event-predictive cognition (Butz and colleagues, including the THICK world models of Gumbsch et al.); it is unrelated to Environment-Centric Active Inference (Esaki et al., 2024), which shares the acronym. EC-AIF has evolved since; the paper describes its core.
NOGAI Mk. II
| Figure | Kind | How it was measured |
|---|---|---|
| 330 km/hmaximum continuous speed | Projected | The conservative floor of a performance closure that solves mass, fuel, power and speed together: at take-off mass, sea level, standard day, the engine at its continuous rating. A speed it holds, not a burst. |
| 400 kgclass | Projected | Mass from the CAD model’s own volumes, kept as a register with ranges. |
| 2.5 to 3.5 hendurance | Projected | From the same closure, with fuel from the net tank volumes. |
| about 18 minto an area 100 km away, from take-off | Projected | From the projected speed, with take-off. |
| 2 aircraftkeep one over an area 100 km out at all times | Projected | Each NOGAI is away about 70 minutes per cycle (18 out, 18 back, and 30 to 40 to refuel, an assumption) and stays over the area for about 90 to 100 minutes per sortie (projected, with a 20 to 30 minute fuel reserve kept), so two take turns and one is always overhead. |
How we know. NOGAI Mk. II exists as a complete engineering model, not a render. Its airframe is designed in exact solid CAD down to the internal structure, with every moving part checked through its full travel. Its drag comes from CFD of the whole aircraft, with the rotor’s downwash and the engine’s exhaust in the flow, and an independent audit adds what the CFD leaves out at a high value. Its rotor is modelled from published section data, with compressibility, retreating-blade stall and the rotor’s dynamic stability checked; its propellers by blade-element design; its engine from the maker’s published ratings. Key structural parts have finite-element checks, and the landing gear a drop analysis with FAA Part 27 as the reference load cases (no compliance claimed). The flight-dynamics model NEXMIND flies is built from the same data, and that data reproduces the closure’s power within about half a percent.
Tested on known answers first. Before N-TEK trusted its CFD, the same recipe ran on three published benchmarks with known answers, to measure its bias: a flat plate, a NACA 0012 wing section and the DARPA SUBOFF hull. Its totals run high, not low: on the SUBOFF hull, within 3.4 % of the towing-tank value. So if anything, NOGAI’s drag figures err high. A simple power model, used as a cross-check and run on the published drag of a 300 kg-class compound unmanned helicopter that has flown, brackets its measured flight power: 76 to 81 kW against 78.3 kW.
Why 330 km/h. After a major earthquake the roads into the worst-hit districts are blocked, and the first look at each one decides when rescue can start there. Bases have to sit outside the damage zone, which puts the city far away, and speed closes that gap. At 330 km/h a district 100 km out gets its first look about 18 minutes after take-off; at 200 km/h it would take about 30. In the same 18 minutes one base reaches almost three times the area.
Two variants. NOGAI Mk. II-E is the endurance-first variant in design: a piston engine in place of the turbine, keeping the compound layout and its pushers. It is built to stay over an area for many hours, where Mk. II-S, the aircraft on this page, is built to get there first.
NOGAI Mk. II is designed end to end by N-TEK. It has not been built or flown: every figure here is a projection or a simulation. Next: CFD that resolves the rotor, a wind tunnel, then hardware.
Missions: an earlier version, outdated
| Figure | Kind | How it was measured |
|---|---|---|
| 755 of 760detectable survivors located · 64 fresh districts | Simulated | Districts the simulation had never flown, with no false alarms. |
| 64 of 64missions in wind, all completed | Simulated | Fresh districts 100 km out all round the base, steady winds up to 54 km/h from any direction; every landing within 1.5 m of the pad’s centre. |
| 48 of 48further missions, in wind and turbulence | Simulated | Sixteen each in steady wind, light turbulence and moderate turbulence: 557 of 558 detectable survivors found, every landing within 2.2 m. |
| 60 of 60missions with one to six aircraft | Simulated | Every detectable survivor located and on the rescue list at the end; every aircraft back on its pad. |
| 8 of 8paired sorties where CFRM learned mid-flight | Simulated | Eight districts, each flown twice with the same recurring sensing fault, with CFRM and without; the story shows one of them. |
Outdated. These results come from an earlier version of the system and are outdated. Detectable: a survivor the simulated scenario gives the aircraft something to sense, fixed before the flight; whether the aircraft senses them depends on where it flies and on conditions. An earlier build also flew in Gazebo, an established third-party simulator, to check NEXMIND outside N-TEK’s own; the aircraft model there is N-TEK’s, and none of these figures come from it.
To check a figure
Reviewing the Gameworld result for research or a partnership? By arrangement, N-TEK can run ten seeds of your choosing on the same frozen version of NEXMIND, under AXIOM’s protocol, and publish every run’s score here: write to contact@n-tek.tr. Before each new result, N-TEK will publish a fingerprint of its registration here, so anyone can check later that the plan came first. N-TEK’s designs and code stay in-house; the figures, and how they were measured, are here.