wizden / moon lander
READY

SECTOR 11 / MARE TRANQUILLITATIS

Moon lander.

MISSION CONTROLSTANDING BY

Autopilot awaits launch authorization.

LANDING OBJECTIVEMoon gravity 1.62 m/s² · lunar clock x2.5 · gear limit: V < 4.5, H < 3.0, attitude < 20°Each mission moves the pad and launch point at random. In side-by-side mode both models fly the identical mission.
MODEL HAS CONTROL
PROMPT
PILOTAdd a Laya or JEV connection above
SPEED

Manual mode: W thrust · AD yaw · S coast

MISSION // CAMPAIGN

Ten landings.

00 / 10 FLOWN

    LIVE DECISION FEED

    00 CALLS
    1. Live model commands appear after launch.

    THOUGHT / TELEMETRY

    NO REQUEST
    Action
    COAST
    Phase
    APPROACH
    Terrain
    0.00°
    Provider response is redacted before display.
    TECH // SPECIFICATIONHow the lander talks to the models

    How a decision is made

    1. Flight simulationRuns the lunar physics in the browser. Every 480 ms of flight time it takes a snapshot: altitude, speeds, tilt, fuel, where the pad is.
    2. Relay serverThe page never calls a model directly. A small server checks the numbers, adds your API key, and strips keys from every response.
    3. Situation textThe snapshot is written as two short texts, one vertical and one sideways. Guided texts state the conclusion; facts-only texts state just the numbers.
    4. ModelLaya or JEV answers three yes/no questions in one call (brake? push right? push left?), each as a score from 0 to 1.
    5. ControlThe scores become one maneuver (brake, push left, push right, or coast). The craft flies it until the next answer.

    Guided vs facts only

    Both modes ask the model the same three questions and turn its answers into a control the same way. The only difference is the situation text, and that decides who does the physics.

    GuidedFacts only
    Who does the physicsThe app's flight director: stopping distance, target sideways speed, when it is time to brake.The model, from raw measurements and the briefing's rules.
    What the text saysThe conclusion: "DANGER: descending too fast … The engine must brake now", "Required sideways correction: accelerate to the RIGHT".Only facts: altitude, speeds and directions, where the pad is, gravity and engine strength, safe limits, time to the ground and to the pad. No verdicts.
    The model's jobRead the conclusion and confirm it. It makes the final call, but the reasoning is already in the text.Work out the decision itself.
    Mission briefingNot used.Optional rules sent with every decision (see below).
    JEV result100% of decisions match the plan · 32/32 landedNo briefing: 25% · 0/6 (brakes every time and hovers until the fuel runs out). With the Flight rules briefing: 14/20 landed.

    Pick a situation to see the same flight state written both ways, and the answers the flight plan expects. These are real inputs, built by the same code as every live request.

    Sends all six situations, in both modes, to each model with a saved key and shows the real scores. Facts-only requests include the current briefing.

    From answers to a control

    Braking always wins when it is needed. Otherwise the stronger sideways answer wins, and if neither side is a clear yes, the craft coasts. Braking winning also means that once a model keeps braking near the ground, the craft can no longer move sideways.

    Mission briefing (facts only)

    An optional text of up to 500 characters, added to the state as mission_briefing; each question gets the sentence "Take the mission briefing into account." It is fixed when a campaign starts and saved with its results.

    It is how you give the model decision rules. Rules written as a formula ("distance to stop is speed² ÷ 6") or as intent ("don't brake while high up") failed: the model stopped braking. Lookup tables work, because matching a number to a row is something the model can do. The "Flight rules" example uses tables plus "brake if altitude is below 50 m". The brake eases off at about 2.5 m/s and never climbs, so touchdowns are soft.

    Guided mode: flight director rules

      Flight model

        Same input for both models

        Laya and JEV receive identical state and questions. Only the transport differs:

        In a side-by-side mission, the first decision is taken from the same snapshot for both models, and the two requests are compared here.

        Timing and speed

        • A decision is requested every 480 ms of flight time.
        • Above 1× the simulation waits for each answer, then fast-forwards, so results do not depend on the speed setting. The real speed is capped by how fast the model answers (JEV takes about 0.74 s, so 4× plays at about 1.7×).
        • The screen replays the recorded flight smoothly, slightly behind the simulation, so waiting for answers never shows as stutter.

        Last request sent, per model

        LAYA 
        No Laya request yet.
        JEV 
        No JEV request yet.

        Measured results

        PromptLaya decisionsLaya landingsJEV decisionsJEV landings
        Original: raw telemetry numbers, one 4-way choice28%0/1233%—
        Guided100%31/32100%32/32
        Facts only, no briefing——25%0/6
        Facts only + "Flight rules" briefing———14/20

        Decisions: agreement with the flight plan on 120–160 labelled flight states. Landings: fixed and random missions in the headless simulator. A dash means not measured: Laya's endpoint was offline for the facts-only runs, and briefing tuning was judged on landings, not decisions. The original prompt is why guided mode exists: with raw numbers, Laya answered "push left" and JEV "push right" for almost every state.