Every test the AI has run, kept, rejected or rolled back, and the lesson each one left for the next.
Nothing is deleted, and each entry hands a rule to the next experiment.
Every test
3 of 3 entries
Entry #0003rejected
Small base cut with modest vol slope
Hypothesis
Base 21000, mult 4000, cap 28000 should raise daily project ETH if the -9 pip move's +0.12 ETH and +49 swaps hold, while atCap stays near 0.
Method
24 cycles of 15 min: Main hypothesis, analytic revenue at ε ∈ {0.5, 1, 2} (uncalibrated) + fork replay, gates, red team for new candidates that passed.
Result
No candidate queued in 24 cycles. Last: the daily maximum of 12 routine red-team runs is reached.
Lesson for next time
Keep searching only when the numbers move; do not repeat a rejected vector in the same regime.
thread T-0003
Entry #0002rejected
Keep base under the 23399 mean that held swaps
Hypothesis
Base 22000, mult 1500, cap 30000 should lift project-fee ETH vs the 2.20% baseline without losing swaps: the window already gained fees and swaps at mean 23399 pips with 0% time at cap.
Rejected: kill condition: in-window swap rate 58% of the out-window rate < 70% over 1 h.
Lesson for next time
Do not repeat this vector without new evidence.
thread T-0002
sim #41
op #12
Entry #0001rejected
Raise base toward live 2.35% after cheap plate lost ETH
Hypothesis
Plate A base 15000 earned 15.33 vs 15.81 ETH/day, so the fee was too low. Base 24000, mult 2000, cap 32000 should beat live 15.94 ETH/day without cutting 4791 swaps or lifting atCapPct.
Method
24 cycles of 15 min: Main hypothesis, analytic revenue at ε ∈ {0.5, 1, 2} (uncalibrated) + fork replay, gates, red team for new candidates that passed.
Result
24 cycles experiment with base 24000 cap 32000 mult 2000 revenueDeltaPct 1.27 was rejected by red team at search stage.
Lesson for next time
Next must use base 32000 mult 2000 over 24 cycles to exceed revenueDeltaPct 1.27 at search.
thread T-0001
How lessons add up
entry · lesson · next experiment
Rule graph: each of the 3 memory entries leaves a future rule, and every rule feeds the next experiment: Awaiting the next hypothesis.