Does paying the fly less for rediscoveries push it to find more different alloys?
Withdrawn. A bug let the discount erase finds (5 fell to 1), and the run changed two things at once, so it settles nothing.
In the log: The fly, the whole brain, and a reward that refuses to pay twice
voidDate not stated in the log; it was written between the commit of 2026-09-16 14:53 and the first commit that contains it, 2026-09-16 15:53generator · fly brain0 predictions · 0 result paragraphsEXPERIMENTS.md lines 8523–8559
What E140 did and how it came out, drawn from this record and the files it names (book/assets/diagrams/exp/E140.svg).
Results
No result paragraph for this entry was found in the log.
The full record
EXPERIMENTS.md · lines 8523–8559
E140 — The fly, the whole brain, and a reward that refuses to pay twice
Every campaign this project has run returned Mo-Ta or a published refractory alloy. The cause
is in the reward: it was pure quality, the landscape has one dominant peak, and nothing ever
paid the generator to leave it.novelty.py existed but only scored finds afterwards, which
reports the problem and does not change it.
Novelty is now multiplicative inside verify(), with two live pressures:
against the literature — a composition within 0.15 of a published alloy is a
rediscovery and is paid 0.25x;
against this run — each prior qualified find within 0.12 decays the reward by 0.6x,
so the fourth visit to a neighbourhood is worth 0.13 of the first.
Multiplicative rather than subtractive on purpose: a genuinely better composition in a visited
basin still outranks a mediocre novel one. The generator is not being told to ignore quality,
only to stop being paid twice for the same discovery.
Substrate: the whole connectome, 164,506 neurons, 25,135,527 edges. Arm: the fly alone.
Class: five elements or more. Ladder: screen plus the ordering confirm rung.
Predicted:
Distinct in-class finds rise well above E136's 5.0, because the repeat decay makes a
fifth visit to the Mo-Ta neighbourhood worth a tenth of the first and the walkers are
pushed outward.
The best driving force is worse than E136's -74, because the reward now discounts the
deepest basin, which is where the best driving forces are. That is the trade being
bought, not a regression, and if quality does not fall at all the novelty pressure is
too weak to be doing anything.
Most finds still score known, because the reference list is equiatomic-only and any
non-equiatomic composition reads novel by construction — a limitation of the check, not a
result (E136).
Falsified if the find count does not move. That would mean the repetition is not coming
from the reward at all but from the walkers collapsing onto one another, and the fix is in
FlyGenerator's run-and-tumble rather than in what it is paid.
Related entries
E136 — The fly has not been used, and nothing found so far is new