Experiments · E209
Not yet. It was never run; like the five-element test, it waits on an energy-model refit that passes its tests.
In the log: local competition
book/assets/diagrams/exp/E209.svg).E209 — local competition. FORAGER_LOCAL_COMPETITION=1, elitist + fly. Bars:
fly depth ≤ −72 (from −66.8, i.e. reaching elitist's level) and finds ≥ 260
(within one seed-sd of 284). Either alone is a partial pass; failing both falsifies the
mechanism as tuned here, not the idea — the niche grid (12 × 10 over VEC 4–11 and δ 0–13)
is MapElites' and has never been calibrated for this reward, and a partial pass would make
the grid the next thing to vary. A rise in finds with no depth gain would mean the
mechanism is acting as a novelty bonus rather than a quality-diversity one.
Queue: both wait on E207 (the hysteresis test), which holds the CPU. E207 first because it decides whether rung 1's protocol or the model is at fault, and that decision outranks either of these.
Scorecard defect, found by review and fixed (2026-09-20 21:34). Yesterday's like-for-like fix
carried a bug of exactly the kind it existed to prevent: FC17's entry was annotated
observable="cv_peak" while its own protocol on the same line read "H_mix(T) inflection",
because I derived the annotation from the method field — and method names a technique
(CE+MC), not a measurement. It was one of three rows with a value passing the filter, so a
third of the surviving scorecard was the original error wearing the fix's clothes.
Relabelled hmix_inflection. Two consequences: the module docstring's claim that "every
entry is … the same quantity the ladder estimates, by the same route" is deleted as false,
and test_the_two_independent_determinations_of_MoNbTaVW_agree — which called FC17's 750 K
and WAK23's 742 K "the tightest external calibration point the ladder has" — is rewritten:
they are an enthalpy inflection and a Landau instability temperature agreeing to 8 K by
coincidence, and treating that as corroboration was the same error again. New guard:
test_no_entry_contradicts_its_own_protocol fails if a row's protocol names a measurement
other than the one it is labelled with (procedure words are excluded — SOB20's protocol
names only its sweep, which is a gap in the source, not a contradiction).
E209 — local competition. FORAGER_LOCAL_COMPETITION=1, elitist + fly. Bars:
fly depth ≤ −72 (from −66.8, i.e. reaching elitist's level) and finds ≥ 260
(within one seed-sd of 284). Either alone is a partial pass; failing both falsifies the
mechanism as tuned here, not the idea — the niche grid (12 × 10 over VEC 4–11 and δ 0–13)
is MapElites' and has never been calibrated for this reward, and a partial pass would make
the grid the next thing to vary. A rise in finds with no depth gain would mean the
mechanism is acting as a novelty bonus rather than a quality-diversity one.
Queue: both wait on E207 (the hysteresis test), which holds the CPU. E207 first because it decides whether rung 1's protocol or the model is at fault, and that decision outranks either of these.
Scorecard defect, found by review and fixed (2026-09-20 21:34). Yesterday's like-for-like fix
carried a bug of exactly the kind it existed to prevent: FC17's entry was annotated
observable="cv_peak" while its own protocol on the same line read "H_mix(T) inflection",
because I derived the annotation from the method field — and method names a technique
(CE+MC), not a measurement. It was one of three rows with a value passing the filter, so a
third of the surviving scorecard was the original error wearing the fix's clothes.
Relabelled hmix_inflection. Two consequences: the module docstring's claim that "every
entry is … the same quantity the ladder estimates, by the same route" is deleted as false,
and test_the_two_independent_determinations_of_MoNbTaVW_agree — which called FC17's 750 K
and WAK23's 742 K "the tightest external calibration point the ladder has" — is rewritten:
they are an enthalpy inflection and a Landau instability temperature agreeing to 8 K by
coincidence, and treating that as corroboration was the same error again. New guard:
test_no_entry_contradicts_its_own_protocol fails if a row's protocol names a measurement
other than the one it is labelled with (procedure words are excluded — SOB20's protocol
names only its sweep, which is a gap in the source, not a contradiction).
Built with PRISMWebsite and visualizations made using Claude