The single-seed reading entered above was reported as firing on both organizations. At the registered three seeds it does not, and the entry above stands as written rather than being edited, with this as its correction.
Registered before these ran: a placebo point estimate of the same order as the cross-organization one is evidence that the gate responds to any project boundary and the organization is not the unit, whatever either interval does. It is.
The registered gate compares an adapter trained on one organization against one trained on another, on the first organization's held-out refinements. It cannot say on its own that the organization is what the adapters learned. A project, a codebase family or a document type would all produce the same reading, and this study's own separability probe already reports that two projects inside one organization separate as well as two organizations do (0.828 and 0.872 within, against 0.849 across, on matched Python).
# research(2026-09). FDLoRA (Lu et al., arXiv:2406.07925) is the remaining half of the
adapter-instance cut. Its paper leaves three things a measurement must decide, so the reading is
registered here, before any of it is implemented or run.
*-dual-cpp-rebuilt-c128-t2-local.json. The same run, the same clients, the same two rounds: the
local adapter each client keeps, read by the two instruments that read the global one it sends.
Both halves had two training passes over the same 128 examples, so this comparison carries none of
the length confound that separates the dual run from the single-adapter one. Neither FedDPA nor
FDLoRA makes it.
*-dual-cpp-rebuilt-c128-t2.json. FedDPA's iterative variant on the rebuilt corpus, 128 examples
a client, two rounds: each client holds a global adapter that is communicated and a local adapter
that never is. What the server receives is the global half, and that is what the two instruments
read here. Neither FedDPA nor FDLoRA attacks the half that leaves; FedDPA states the assumption
instead, that "all clients are trusted ... and the whole process does not suffer from any attacks".
defence-curve-cpp-rebuilt-c128-mid.json. The masking result was measured on the nine-project
corpus at 64 examples a client. This repeats it on the rebuilt corpus at 128, content held to C++,
48 clients, at rounds 10, 30 and 50 with six mask sizes.
subspace-split-cpp-256-c256.json. The same registered split at 256 examples a client, on 32 C++
clients over nine projects, against 126 groupings.
aggregate-attack-cpp-rebuilt-c128-{a,b}.json. FedSA-LoRA shares A and keeps B local, and its
reason is that A is the more similar factor across clients. The same 48 C++ clients over thirteen
projects, read three ways, at rounds of sixteen holding two of the target's clients:
# research(2026-09). The two designs whose cut this apparatus has not measured, read from their
PDFs rather than from their abstracts, to settle what a measurement would have to implement. Four
load-bearing quotes were re-checked against each PDF's own text layer.
scripts/cluster_informativeness.py, datasets/results/cluster-informativeness.json. The
registration binds the gate to the pooled estimand, and one of its two reasons is Kahan et al.'s
condition: pooling over examples is wrong when a change's size carries information about its
outcome. The report quoted a correlation for that, and nothing on disk produced it.
# research(2026-09). The Stage 1 report carried ten \needcite badges because the intake
staged every entry as a search-level characterization. All but one are now cited: each source was
read against arXiv, the ACM Digital Library, the ACL Anthology or Springer, and four of the claims
built on them were wrong in a way a reviewer would have caught.
datasets/results/interval-calibration.json, scripts/interval_calibration.py. Section 7 of the
report states the pairs cluster bootstrap's false-positive rate under a true null. That figure has
been living in a comment in measure/stats.py since it was measured, with no artifact behind it,
which is the one thing a registered report cannot do with a number a reviewer will check.
*-c128-p2.json. The 128-example run said leakage is a function of how long each client trains.
It was one packing, so who shared a client with whom was decided once, and the reading rested on
two points that differ in size as well as in length. This is the same 51 clients, the same
training seed, the same 128 examples each, repacked with packing seed 2: the only thing that moves
is which examples group into which client.
The review found the raw probe reading 0.841 with an interval of [0.814, 0.836]. The cause was the procedure, not the interval formula: each bootstrap resample refitted the classifier on changes drawn with replacement, about 63% of them unique, so every refit trained on less and scored lower, and the resampled accuracies sat below the estimate by more than their own spread. A bias-corrected (BCa) interval was tried first and made it worse -- [0.839, 0.839] -- because any interval built from quantiles of the draws cannot contain an estimate that every draw falls below.