arc-witness replay

0 / 0 speed

Levels: filter action history:

Game CRT

↑ ← ↓ →

⟲ ✓

Now

Action history ◆ ORAI = LLM call · dim = plan exec

Plan

Memory

Active rules

running state at this step

● verified ● proposed ● transferred ● disputed

Compacted (last LLM-fed snapshot)

Last change

Memory growth

Reasoning trace

What the LLM saw (semantic ASCII board)

Rules fed to ORAI (## Current Knowledge)

Observations (recent action effects)

LLM response (meta = thinking, add/del = rule churn, plan = next actions)

What is the agent doing right now?

The agent runs the ORAI reflect loop (see /changelog.html on the runs page for harness conventions). The current stage shows in the Now panel and is color-banded in the timeline above:

reconnaissance SCOUT probes each action a few times from different states to learn the action grammar (no LLM calls) and builds the transition table (world model). The semantic-ASCII board itself is ground-truth perception from the env (env_ground_truth), not inferred from these probes. ~10% of the step budget.
planning The bulk of the run. The agent's ORAI loop (Observe–Reflect–Act–Inject) reflects — reads the current board, the recent action effects, and its rule set — then updates its rules and proposes a short action plan (which may include 0 = RESET) that the Worker executes step by step. The ◆ ORAI rows in the action history are the reflection (LLM) calls; the plain rows between them are that plan being carried out. The recent action effects (the "Recent actions" line) come from the env ground-truth cursor path (_path).
reset RESET is an agent-chosen action 0 inside the ORAI plan (⟲ in the plan pane / D-pad) — the board returns to the level start, but nothing else starts over: same level, same segment on the timeline, same step budget, step counting uninterrupted (a reset salvages a doomed attempt, it never refunds its cost).