◉ Unified Enactive Agent Living World · v2.13.11 · Enactive AI v0.4
Two situated agents with independent predictive memories, live multiscale hypotheses, negotiated construction, and an independent interaction observatory. Sessions begin paused.
Click to place the selected resource. Shift-click or right-click always places a hazard; Alt-click places a fixed wall. Ctrl/Cmd-click inspects a drawing episode. For interaction-only drawing, disable both resource types and set Drift to 0%. Joint projects require two-way consent.
Collisions reduce viability; each hazard has a 28-step cooldown per agent.
—Selected agent viability
—Selected agent nourishment
—Selected agent hazard contacts
—World steps
Viability over time
The world and agent evolve while this tab remains active.
Observed action → consequences
Sense-making curves · traditional and composite
Click a point on the curve to inspect language, coordination, foresight and before/after evidence. The inspector does not establish causality.
User = Aster as a simulated user-role proxy, not a human participant. Traditional code: fluid clamped action −1, pause 0, exploratory unclamping +1. Experimental composite uses evidenced action/cognitive exploration and relational reorganization, modulated by breadth and depth. Traditional v2.2 scoring is preserved. Curves accumulate signed codes; view Instantaneous to inspect momentary changes. Combined captures dyadic coupling beyond a simple average. The fourth trace is an exploratory, phase-matched, time-shifted partner excess-alignment measure; it is independent of the combined composite formula, but does not demonstrate causality or emergence. Rolling-window mode shows mean signed code, not accumulated total.
Experiment / Benchmark · controlled world runs
Ready
Fresh learning and environment per run. Select agent count, hazard condition, maze, and a simulation-step duration. Results remain only in this page's live memory across successive runs (not localStorage): a refresh or Clear History removes them.
No benchmark running · ordinary free-play remains available.
0Runs in memory
—Food / agent / 1k steps
—Damage / agent / 1k steps
—Mean prediction MAE
Run
Conditions
Steps
Food / agent / 1k
Damage / agent / 1k
Resets
Final viability
Maze gates / agent
Prediction MAE
Joint outcomes
Run a benchmark to populate comparable results.
Research evaluation · learning and transfer
No trained models
Architecture ablations change which mechanisms can influence behavior. Frozen tests use only a model learned in Train mode on earlier environments. Set different seed ranges for training and test. Perturbations occur halfway through a run. A partner reset is an unfamiliar-partner proxy, not a human generalization test.
—Prospective +H hazard MAE (pixels)
—Position persistence MAE
—Expression reply rate
—Post-shift prediction recovery
Start a research run to see task progress and strictly prospective model forecasts.
Matched-seed contrasts · active session
Compares independent and communicating two-agent runs with identical protocol, architecture and seed. Differences represent task-score contrasts; CIs summarize replicate variability, not causal identification. Unequal world trajectories and shared environmental randomness remain limitations.
Matched comparison
Pairs
Mean Δ task score
Approx. 95% CI
Mean Δ food / agent / 1k
Run matched independent and social dyads to compare.
Protocol and validity
Primary endpoints: per-agent food, observed damage, resets and forecast error; task score is a secondary declared composite. Predictions are issued from each agent's remembered observations before the future outcome; environmental truth is used only by the scorer. Nonvisible objects cannot be forecast. Skipped or vanished targets reduce forecast coverage. Training knowledge is restricted to action associations, predictive tendencies and lexical associations. Tests freeze learned action/prediction/lexical/social associations while still updating the transient sensed world state and regulation mode. Generated expression uptake and exploratory Fun Score are behavioral proxies, not proof of semantics or experienced fun. Multiscale predictions are strictly prospective and evaluated against persistence; organization uses A→B→A action/message alternation compared with phase-matched shift controls (not causal identification). Existing UEA movement heuristics still use globally nearest resource/danger distances, unlike the finite-range imagination module; these benchmarks therefore do not yet establish partial-observation control.
Maze gates (0–4) count the number of successive barrier x-thresholds crossed by each agent, averaged across active agents; they are a coarse navigation proxy, not proof of maze completion. Metrics are normalized per active agent and 1,000 simulation steps. A solo run has no dyadic interaction, negotiation, or emergent partner signal; joint results are therefore marked N/A. Seeded trials reproduce initial randomized draws, but they are descriptive simulations, not controlled causal estimates: agent count alters subsequent random-number consumption and trajectories. Run replicates to assess variability.
Enactive dynamics · organization, intelligence & fun
Prospective evaluation
Organized behavior = temporally alternating action/message adjustments (A→B→A), beyond parallel joint movement and an environmentally phase-matched shift control. Adaptive intelligence = prospective local/regional/global prediction skill against persistence. Fun = optional, low-urgency social exploration, not felt enjoyment.
Adaptive intelligence—LRG weighted skill against persistence · 50 = parity
Fun score—Low-urgency, non-task social exploration proxy
Plot uses recent, prospectively scored observations. The vertical axis may zoom to show small differences; 50 is forecast parity with persistence.
Future horizons: Local +5 / Regional +20 / Global +80. Forecasts must mature before scoring.
Waiting for a sufficient number of action adjustments and environmental phase-matched controls.
Off-task social onset forecast · next +60 steps: awaiting observations.
Play is available only when viability and immediate environmental safety permit. Exploration may compete with task progress.
Experimental indices, not validated psychological constructs. A→B→A alternation alone does not establish causal influence; matched shifts reduce but cannot eliminate common-input confounds. Predictions are issued before outcomes and separately scored at +5/+20/+80 steps; a negative gain means persistence won. Fun labels observable voluntary behavior rather than subjective experience.
Emergent language · lexical history & interaction evidence
Word meanings are sender-grounded hypotheses or receiver inferences. Subsequent coordination and co-creative changes are temporal associations, not causal effects.