# 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.
# research(2026-09). Two papers by the same author attribute agents from pull requests, and
both are close enough to RQ1 that a reviewer will raise them.
*-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".
The three adversarial reviews on 2026-09-18 and 2026-09-19 each ended with a list recorded rather than fixed. Re-reading those lists against the code, most had been overtaken: the permutation now has behavioural tests rather than text searches, the enumeration counts distinct groupings, the overwrite guard claims exclusively at job start, and the contamination battery's partial output is claimed under its own name. Two were still standing, and both are closed here.
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:
datasets/results/subspace-split-cpp-rebuilt-c128.json. Two tests were registered for the
ten-project rebuild before it existed: the detector on the shared half at rank 4, and the
classifier on the residual at rank 1. Both are now run on 48 C++ clients over thirteen projects,
with the attribution basis fitted without the scored client's whole project and the null the same
1,716 groupings the detector uses.
*-cpp-256-c256.json. Thirty-six clients, thirty-two of them C++, over five AOSP projects and four
of Qt's: the largest set this corpus supports at this size, since a project needs 256 usable
examples after deduplication to fill even one client.
*-cpp-rebuilt-c128.json. The ten-project AOSP rebuild gives seven C++ projects that can fill a
client at 128 examples, against Qt's six, so the project-level permutation enumerates
C(13, 7) = 1,716 distinct groupings where every earlier run had 84. Same training length, same
initialization, a larger and better-balanced project set.
An independent reviewer was given the three pieces of code whose numbers reach the report and told to reproduce rather than read: the interval calibration, the informativeness reading, and the harvest that asserts every quoted figure. It ran its own implementations and its own mutants. What it found, and what each one cost:
# 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.
scripts/separability_over_time.py. The review noted that this probe reads every built month,
dev included, where separability.py stops at train, and left it. The difference is right and was
undocumented: this one asks whether the corpus moves over time, and a drift reading that stops two
months early cannot see the end of its own series. It reads corpus text and no held-out
predictions, so the dev window is described rather than spent. The docstring now says that instead
of implying both probes follow one rule.