| … | … | @@ -926,6 +926,39 @@ transfer to 1→2 on its own — only Law 2 does. |
| 926 | 926 | |
| 927 | 927 | --- |
| 928 | 928 | |
| 929 | ## The four-check tally — one mistake, made four times (2026-07-31) |
| 930 | |
| 931 | Every headline claim in Part C was eventually re-tested at **matched conditions**, because the first |
| 932 | version of each compared arms that differed in more than the variable of interest. The results: |
| 933 | |
| 934 | | claim | re-checked at | outcome | |
| 935 | |---|---|---| |
| 936 | | Test A — assortment variance beats compartment size | matched **realized size** | **survived** (+1.14, zero where the modes coincide, +2.98 where they diverge) | |
| 937 | | Test B T4 — variance destroys collective closure | matched **realized size** | **RETRACTED** — RAF tracks size in 8/8 cells regardless of mode | |
| 938 | | `chemoton.c` C3 — reproduces the decree's signature | matched **realized size** | **DOWNGRADED** — threshold 2.52× tighter (half-fall 4.6 vs 11.7) | |
| 939 | | `partition.c` — partitioning kills gating | matched **partitioning**, cost fixed | **survives, mechanism corrected** — it is edge loss, not topology | |
| 940 | |
| 941 | **All four failures had the identical shape: one knob moving two variables, with the effect credited |
| 942 | to the interesting one.** |
| 943 | |
| 944 | - maintenance cost moved fragmentation *and* everything else (`partition.c`) |
| 945 | - propagule mode changed assortment variance *and* realized compartment size (Tests A and B) |
| 946 | - division threshold changed the decree comparison *and* realized size (C3) |
| 947 | - severing changed edge count *and* component structure (the partition re-check) |
| 948 | |
| 949 | Each was caught the same way: hold the confounded variable fixed, or interpolate onto it, and ask |
| 950 | whether the effect survives. One survived untouched, one was retracted, one downgraded, one had its |
| 951 | mechanism corrected. |
| 952 | |
| 953 | This is more transferable than any single result in Part C. **The pre-registration discipline |
| 954 | ("never encode the answer in the update rule") catches encoded outcomes but does NOT catch confounded |
| 955 | comparisons.** A model can be perfectly honest about what it supplies and still attribute its effect |
| 956 | to the wrong variable. The matched-condition re-check is a separate discipline and has to be applied |
| 957 | separately — ideally at design time, by asking of every comparison: *what else changes when I move |
| 958 | this knob?* |
| 959 | |
| 960 | --- |
| 961 | |
| 929 | 962 | ## Standing methodological lesson (the operator enforced this all session) |
| 930 | 963 | |
| 931 | 964 | > **Never encode the answer in the update rule.** "Solve an equation you wrote" = engineering. |