JevHarness / Pokémon Lab
PYTHON + JEV · PRESET SINGLES
LEARNING FROM COMPLETE BATTLE TRACES

Evolving a better Jev flow.

Loading reflection rounds and the initial-to-selected Eval result.

Loading experiment

Reflection model: loading

Loading Train and Eval games…Connecting to the local archive.
Initial flow · Eval
Awaiting recorded games
Selected flow · Train
Observed games only
Reflection rounds
Candidate proposals

Train and Eval contain different matchups and random seeds from the same preset team pool. Train scores cover only the games sampled for each candidate; Eval guides selection.

01Watch the decisions. See the outcome.

A short battle replay with Jev's recorded probabilities, synchronized to each featured decision.

SELECTED FLOW · EVAL GAME 05
THE AGENT WRITES THE CALL. JEV JUDGES THE OPTIONS.

Key decisions → a winning finish.

Loading highlights
RECORDED BATTLELoading the featured moments…No new games or model requests are made.
Turn —

Loading the decision sequence from the recorded game.

4 featured moments

Archived gameplay and actual Jev answers. Decision previews pause for readability; their duration is separate from the recorded call latency. Omitted turns are marked between chapters.

Exact Jev arguments and results for this moment

Current decision

Initial vs evolved · full battle archive

Compare complete replays, exact calls and recorded outcomes.

Same game. Different decisions.

Compare the initial policy with the selected policy on a shared Eval game.

Loading matchup…
Awaiting results

Watch the full battle replays

The same initial matchup, played by each flow
Before / Initial policy

Initial Jev flow

Loading archived game…

Loading the archived battle…

After / Selected policy

Evolved Jev flow

Loading archived game…

Loading the archived battle…

Both policies start with identical teams, opponent policy and seeds. These are recorded outcomes; later battle states diverge as the policies choose different actions. Renderer assets require an internet connection; battle logs stay local.

Compare actions by turn
Battle turnBefore · initial policyAfter · selected policy

Actions are aligned by turn and decision phase, not by identical state. Multiple decisions in a turn are preserved.

02From candidate to flow.

Select a candidate on the left. Explore its features, Jev questions and changes on the right.

Loading candidates

Evolution tree

Actual parent → child relationships
Reading the evolution archive…
Selected policyAccepted into the poolNot accepted

A round may branch from an earlier candidate. A proposal without Eval games was not scored on Eval; it is not a zero.

Candidate

Select a tree node

TRAIN · OBSERVED GAMES
EVAL · COMMON GAME SET

Changes from parent+ Added   ↻ Changed   = Retained

    No pipeline selected.

    Node details

    Click a flow node to see its inputs and decision criteria.

    Complete pipeline specification

    03How fast does the flow decide?

    Measured execution time from archived battles. Reflection time is separate from playing a turn.

    Initial flow · decision median
    Loading recorded timings…
    Selected flow · decision P95
    95th percentile of included decision times

    End-to-end decision latency

    Median and P95 · milliseconds
    Loading archived timings…

    Inside the chosen candidate

    Follows the selection in the evolution tree
    Jev call median: —
    NodeSamplesMedianP95

    Loading recorded timings; no new games or API inference requests are made.

    Decision explorer

    Open any recorded Train or Eval game for the selected tree candidate.

    Select a game to inspect its complete decision trace.

    No game loaded.

    Executed decision flow

    Recorded inputs, outputs and timing
    Select an action.

    Runtime node

    Select a node to inspect its complete execution record.
    Complete decision · observation, memory and action
    No decision selected.
    Complete local battle protocol
    No game selected.

    Inside the evolution loop

    Recorded frontier selections and the complete inputs sent to the reflection model.

    Instance frontier selections

    Snapshots from the actual parent selection events.

    Complete reflection inputs

    View the model prompt or its full original input, with archive integrity verification.