Skip to content

Planned before any adapter d_org on s is read: H2 as a Qt cell against a C++-matched foreign adapter

# research(2026-10). Decided in the research-question review of 2026-10-14 (AJ's full-auto grant; the host questions at the end stay his), and revised the same day after an independent review. The mock review panel ranked H2 resting on two pending permissions as the likeliest reason for a continuity acceptance (papers PR #85, org-house-style/REVIEW-PANEL.md, item 1), and the registered rule makes H2 exploratory with near certainty: Chromium's infra-dev list has sent only an automatic reply to the request of 2026-09-23, and its freeze by 2026-10-23 is also blocked on cutting chromium/src below the project and on sub-month fetching past the host's 10,000-result cap (Plan A).

What was already known. No adapter d_org on s exists on any corpus: the decomposition pilots ran no H2 cell. Two related readings do, and both were in view when this was decided: the retrieval comparator's sibling-minus-foreign readings on OpenStack and Wikimedia (2026-10-06), and RQ2's source-attribution pilots on held Qt and AOSP client updates, which found no organization signal within C++ beyond content and project at 64 examples a client and one for AOSP, not Qt, at 128 and 256 (2026-09-18 and Plan F). The second is why AOSP comes last in item 1.

Checked and rejected.

  • A language-matched pair inside the admitted hosts. Wikimedia's registered halves hold 153 and 60 .py training examples of 2,094 and 2,095 (split-criteria-wikimedia.json), and OpenStack's hold no PHP, so neither can train a foreign adapter at N = 1,850 in the other's language. Wikimedia's sibling-minus-foreign retrieval reading against OpenStack, +0.0640 [+0.0434, +0.0853] at k = 1 (2026-10-06 entry), bounds the language confound from above, since it also holds any organization effect.
  • OpenDev's namespaces as organizations. It reopens H1's registered OpenStack cell (splits, K, simulations), and the namespaces share a host, CI and foundation.
  • Full Chromium by 2026-10-23. Three blockers in nine days, at the scale the study chose to wait on (2026-10-08 entry).

Fixed now, before any adapter d_org on s is read:

  1. Cells and source. If Qt and Chromium are both admitted as evaluated organizations (rule unchanged), H2 is that pair, each cell matched as in 3. Otherwise, if Qt is admitted, H2 is one confirmatory cell, Qt, against the first of these sources whose host is admitted for this use when the first H2 development contrast is read, and whose pool covers every matched draw in 3: (a) Chromium, with its maintainers' permission; (b) Fuchsia, from Qt's own training-window months, live and never read by this study; © AOSP, changes created 2024-10-01 to 2025-03-26, before aosp-main went read-only. Fuchsia's and Chromium's git hosts serve AOSP's robots.txt line for line (read 2026-10-14). Otherwise H2 has no confirmatory cell, as before.
  2. Estimand. "C++" means CPP_SUFFIXES throughout (.h included, .c not). The confirmatory H2 cell reads Qt's held-out C++ examples, the estimand cpp_supplement computes today; the whole-hunk d_org is reported beside it and decides nothing. H1's Qt verdict stays on all of Qt's examples. The joint shares, the total s(O_a) - s(P_x) and a restricted d_proj are read for Qt on its C++ examples, reported only, since the joint bootstrap needs one example set.
  3. Matched foreign arm. For each run and each evaluated half, the foreign adapter trains on N examples from the source: as many C++ examples as that run's sibling draw holds, the rest from the source's other files, drawn with the run's training seed. Sibling and foreign then match in total and in C++ examples. The pool has to cover the draws of all 40 admissible splits. It is deduplicated against Qt like any organization, and the foreign pool against Qt's evaluated test window joins check 4 at the sibling's threshold (2% at Jaccard 0.7); a pool over it leaves H2 exploratory and does not halt H1. Filtering foreign training data toward the target has precedent in Turhan et al.'s nearest-neighbour filter of cross-company data (EMSE 14(5), 2009, doi:10.1007/s10664-008-9103-7).
  4. AOSP's projects, if AOSP is the source: the projects in its manifest at android-15.0.0_r1 whose tree at that tag is more than half C++ by code-file count (read from git, no review data), taken in descending order of their NoteDb change count until the pool covers every draw. The ten projects fetched over REST on 2026-09-18 are too few for a draw in this span (counted on held data, which enters no figure).
  5. Qt's admission needs its terms in writing and its cells' bounds simulated and committed before submission; the gate already refuses a cell without bounds. H1 then has three cells, so every cell's simulation is rerun at three.
  6. K and bounds for the H2 cell come from its own development pilot and simulation. A cell whose detectable effect is δ or more at K = 40 is read supported or inconclusive, never bounded. Qt's joint read on its C++ examples takes the larger of its H1 and H2 K, since every run trains own and sibling adapters. The H2 test-window read over partitions goes through require_test_read.

Superseded later on 2026-10-14. The second sentence of item 6 ("never bounded") gave way to the bounded reading at the SESOI (entry below); the rest of item 6 stands. 7. Vocabulary. Both hypotheses are reread on identifier-renamed references (panel item 5), which matters most for H2: Qt's sibling knows Qt's API and no foreign adapter does. The foreign source's reviewer overlap with Qt is reported beside the exposure analysis.

Superseded on 2026-10-14. No googlesource git host is in GIT_PERMITTED until its maintainers grant permission in writing (entry of that date).

Open, AJ's: which hosts are admitted for the foreign draw (Fuchsia's and Chromium's git hosts before infra-dev answers; AOSP's is in GIT_PERMITTED), and whether the pending HSRO protocol (IRB0000393) needs an amendment naming the source.