kadubon/collective-intelligence-overlay: v0.4.3: bounded near-transfer sensitivity pilot

Abstract

v0.4.3 adds a bounded, preregistered near-transfer sensitivity pilot and offline resource endpoints. Its 86 package implementation files remain byte-for-byte identical to the published 0.4.2 runtime used in the study; distribution metadata selects 0.4.3. The real CPU Gemma4:e4b pilot completed 54 requests and 26,256 measured/charged tokens, with no missing usage. Locked SQL, numerical calibration and formation were 0/6, 1/6 and 0/6. The entrance gate is assay_not_ready: confirmation was not started. Full/Empty effects, CIO versus ordinary-memory benefits, formation benefits and break-even remain unestimated. Fixed-cost attribution is incomplete. The sensitivity result and software validation are separate outcomes. The corrected full native run 37144034341 passed 12 native profiles, mixed Python, four cross readers and ready. The initial run 37141746108 failed a new unit test's checkout-relative source path; its original failure is retained. The corrected test passed locally and outside checkout, and the corrected native gate verifies the changed scientific sources. Both candidate origins produced the exact same content pair. The annotated v0.4.3 tag on main used the standard hash-bound selector. OIDC publication run 37155346028 restored the original pair and gate through the existing pypi environment, without a rebuild or repeated native workload. Actual PyPI 0.4.3 files match the candidate hashes. A fresh no-cache/no-config actual-index installation outside the checkout matched all 86 package files and passed CLI/optional imports plus the real PostgreSQL/OPA/MCP/A2A/MAF three-process demo: ACCEPT, 117.00, REQUALIFY after environment change, REJECT after dependency withdrawal. The owned verification cluster was stopped. The single research archive contains the new protocol, frozen sources, all model requests/responses, signed records/CAS, corrected reader, full-denominator CSV, English/Japanese reports, failure records, local regressions, native evidence and actual publication checks. Signed/raw scientific bytes are unchanged; unsigned metadata path redactions have explicit mappings. Precompression layout/privacy/guard stops and repairs are retained. The archive contains no old 0.4.2 raw repack or model weights. Its retained upstream assembly license-notices ZIP has the original native-gate hash. With the recorded dependencies installed, python review-offline.py NEW_OUTPUT_DIRECTORY verifies the saved experiment using frozen sources and the corrected reader, with external connections blocked and no inference. The original failed reader remains in the frozen snapshot. Installation of dependencies is a separate prerequisite. Hashes establish content identity and contract consistency, not truth or protection against a malicious operator. The study does not establish distant-domain generalization, benefit or equivalence. Actual verification details are recorded in the release documentation.