Experiments · E120
Yes. The file holds 671: the same 136 prototypes plus 535 database phases. But 18 data files have no recorded producer.
In the log: Reconciling the hull phase count, and what a fresh clone cannot do
book/assets/diagrams/exp/E120.svg).E120, second half: a fresh clone cannot rebuild this project, and the gap includes the default expansion.
data/ and runs/ are gitignored and zero files under either are tracked. Twenty-six
files sit in data/ on this machine. Checking each against every script that writes a file,
eighteen have no producer anywhere in the repository:
ce_12element.npz / .icet <- THE CURRENT DEFAULT EXPANSION
ce_12element_focused.npz / .icet
ce_8element_prior.npz / .icet
references_equilibrium.npz <- required by stage_b for 8 elements
references_equilibrium12.npz <- required by stage_b for 12 elements
offlattice_hull12_full.json (orphan, nothing reads it)
offlattice_hull.partial.jsonl
pair_interactions.npz
odt_sample.json
+ 7 connectome inputs (external downloads, legitimately not produced here)
Two of those are serious rather than untidy.
ce_12element.npz is the default expansion since E114 and cannot be rebuilt.
build_ce_multi.py writes data/ce_multi.{npz,icet} by default, not ce_12element; whatever
invocation produced the twelve-element fit is not recorded anywhere. Every result from E114
onward - the 1705-composition grid, the 55 feasible compositions, the E115 head-to-head -
rests on a file that exists only on this machine and whose provenance is not in the repo.
references_equilibrium*.npz have no writer at all. Nothing in scripts/ or forager/
calls savez on that stem. stage_b.build() refuses to start without one, so a clone cannot
even construct a verifier.
This is not a documentation gap. The 227 passing tests pass because untracked files happen
to be present, which is why CI, which installs neither icet nor data/, fails about 19 of
them. A green suite here is not evidence the project builds.
Not fixed this hour, and deliberately not papered over with a stub. Writing a plausible
producer that regenerates a different file would be worse than the current state: every
number in EXPERIMENTS.md would then be attributed to an expansion that did not produce it.
The honest repair is to (a) record the exact invocation that made ce_12element and
references_equilibrium12, from shell history or the file metadata, and (b) decide whether
data/ should be tracked, which for a 12 MB expansion is a real decision and the operator's
to make. Both are written into OVERNIGHT.md.
E120 result: prediction confirmed, and the bookkeeping fix uncovered a substantive one.
offlattice_hull12.json decomposes as 136 prototypes + 535 database phases - 193
alexandria, 167 mc3d, 132 mc3d-psol, 43 jarvis - which is 671. E86's 613 is stale, not
wrong: the hull was rebuilt with 58 further database phases after that entry was written
and the entry was never updated. E86 is superseded on this number, not withdrawn.
The substantive finding is what the breakdown shows about the other hull.
offlattice_hull.json, the eight-element hull that every pre-E114 result was differenced
against, holds 136 phases and all 136 are prototypes. It contains no database phase at
all. All 136 appear in the twelve-element hull as its prototype set, so the two are nested:
8-element hull 136 prototypes + 0 database
12-element hull 136 prototypes + 535 database
The twelve-element hull therefore adds 535 competitors the eight-element hull never had. A driving force is the distance below the hull, so a composition scored against the wider hull is scored against far more rivals, and the two numbers are not comparable. This is a difference in kind, not in element count.
This makes task 5 - "re-measure the eight-element record on twelve" - a larger claim than it appeared. It was written as though only the element set had changed. Both the expansion and the hull changed, and the hull change is the one that moves driving forces. Any comparison between a pre-E114 number and a current one is confounded by it unless the hull is held fixed.
Consequence for E117, which is load-bearing. Its 1705-composition grid, the 55 feasible compositions and the "twelve-fold enrichment" were all computed against the 671-phase hull. That is the right hull for a twelve-element search, so E117 stands. But its comparison to E114's eight-element collapse measurement crosses the hull boundary and should not be read as a like-for-like difference.
This entry is written in 3 separate places in the log, shown here in log order.
Two defects surfaced by the documentation pass, neither fixed when found.
The hull count disagrees with the record. E86 states the twelve-element hull stands at
613 phases - 136 prototypes plus 477 database. data/offlattice_hull12.json on disk holds
671. offlattice_hull12_full.json holds 1072 and is an orphan nothing reads.
Predicted: the 671 decomposes as 136 prototypes plus 535 database entries - the same prototype set, more database phases - because the hull was rebuilt after E86 was written and the entry was never updated. On that reading E86 is stale, not wrong, and the fix is to record the current number rather than to rebuild anything.
Falsified if the prototype count is not 136, which would mean two different hull-building paths exist and produce different prototype sets, and then the question is which of them produced the file every driving force in this project has been differenced against.
E120 result: prediction confirmed, and the bookkeeping fix uncovered a substantive one.
offlattice_hull12.json decomposes as 136 prototypes + 535 database phases - 193
alexandria, 167 mc3d, 132 mc3d-psol, 43 jarvis - which is 671. E86's 613 is stale, not
wrong: the hull was rebuilt with 58 further database phases after that entry was written
and the entry was never updated. E86 is superseded on this number, not withdrawn.
The substantive finding is what the breakdown shows about the other hull.
offlattice_hull.json, the eight-element hull that every pre-E114 result was differenced
against, holds 136 phases and all 136 are prototypes. It contains no database phase at
all. All 136 appear in the twelve-element hull as its prototype set, so the two are nested:
8-element hull 136 prototypes + 0 database
12-element hull 136 prototypes + 535 database
The twelve-element hull therefore adds 535 competitors the eight-element hull never had. A driving force is the distance below the hull, so a composition scored against the wider hull is scored against far more rivals, and the two numbers are not comparable. This is a difference in kind, not in element count.
This makes task 5 - "re-measure the eight-element record on twelve" - a larger claim than it appeared. It was written as though only the element set had changed. Both the expansion and the hull changed, and the hull change is the one that moves driving forces. Any comparison between a pre-E114 number and a current one is confounded by it unless the hull is held fixed.
Consequence for E117, which is load-bearing. Its 1705-composition grid, the 55 feasible compositions and the "twelve-fold enrichment" were all computed against the 671-phase hull. That is the right hull for a twelve-element search, so E117 stands. But its comparison to E114's eight-element collapse measurement crosses the hull boundary and should not be read as a like-for-like difference.
E120, second half: a fresh clone cannot rebuild this project, and the gap includes the default expansion.
data/ and runs/ are gitignored and zero files under either are tracked. Twenty-six
files sit in data/ on this machine. Checking each against every script that writes a file,
eighteen have no producer anywhere in the repository:
ce_12element.npz / .icet <- THE CURRENT DEFAULT EXPANSION
ce_12element_focused.npz / .icet
ce_8element_prior.npz / .icet
references_equilibrium.npz <- required by stage_b for 8 elements
references_equilibrium12.npz <- required by stage_b for 12 elements
offlattice_hull12_full.json (orphan, nothing reads it)
offlattice_hull.partial.jsonl
pair_interactions.npz
odt_sample.json
+ 7 connectome inputs (external downloads, legitimately not produced here)
Two of those are serious rather than untidy.
ce_12element.npz is the default expansion since E114 and cannot be rebuilt.
build_ce_multi.py writes data/ce_multi.{npz,icet} by default, not ce_12element; whatever
invocation produced the twelve-element fit is not recorded anywhere. Every result from E114
onward - the 1705-composition grid, the 55 feasible compositions, the E115 head-to-head -
rests on a file that exists only on this machine and whose provenance is not in the repo.
references_equilibrium*.npz have no writer at all. Nothing in scripts/ or forager/
calls savez on that stem. stage_b.build() refuses to start without one, so a clone cannot
even construct a verifier.
This is not a documentation gap. The 227 passing tests pass because untracked files happen
to be present, which is why CI, which installs neither icet nor data/, fails about 19 of
them. A green suite here is not evidence the project builds.
Not fixed this hour, and deliberately not papered over with a stub. Writing a plausible
producer that regenerates a different file would be worse than the current state: every
number in EXPERIMENTS.md would then be attributed to an expansion that did not produce it.
The honest repair is to (a) record the exact invocation that made ce_12element and
references_equilibrium12, from shell history or the file metadata, and (b) decide whether
data/ should be tracked, which for a 12 MB expansion is a real decision and the operator's
to make. Both are written into OVERNIGHT.md.
Built with PRISMWebsite and visualizations made using Claude