Foresightgated spike — cheap world-model lookahead for live agents
Bet · gated
Small-batch · one session · R&D mode: expect learning, not shipping
Pitch foresight-world-model · Direction Lacuna dir 29561
Shallow simulated foresight — cheap world-model, depth ≤2, ≤4 branches — beats reactive ReAct on a real, non-sandbox GUI/web benchmark at tolerable token cost.
The session opens with a ≤20-minute go/no-go probe
Confirm a concrete non-sandbox GUI/web benchmark + agent harness we can drive today, off the shelf. The probe decides whether any cycle is spent at all.
How the table voted
Does the problem matter?Pass
Myopia is the wall for long-horizon agents (BALROG, Why Reasoning Fails to Plan); the commercially hot segment.
Is the appetite right?Pass
One session is honest because the rabbit-hole caps (coarse deltas, depth ≤2, branch ≤4) force it.
Is the solution attractive?Pass
Falsifiable — failure is a result (bet dies cheap), not a debugging spiral.
Is this the right time?Strong pass
Dir 29561 early (111 limitations / 16 findings) and accelerating; sandbox-proven, in-the-wild open. The window is now.
Is the capacity there?Flag
No visible Mila author gravity — bet rests on velocity + whitespace, not studio home turf. And the reusable-benchmark assumption is unverified → that's what the gate tests.
Done = a measured yes/no on the bet, deployed as a runnable comparison — foresight-agent vs. reactive-ReAct on the chosen benchmark, reporting two numbers: task-success delta and tokens-per-task. Not "code written." A clean no that kills the bet counts as shipped.
- Mila gravity gap on dir 29561 — a direction with no studio gravity is a weaker long-term bet even when the science is hot. Bet taken this once because the spike is cheap; revisit if a second cycle is proposed.