Research update · 5 October 2026
Carrier R03 examines what survives a machine handoff
The new Carrier R03 publication, “What survives the handoff?”, examines evidence selection, bounded machine memory, reference integrity, and transfer costs. Its retrieval finding is negative: a fixed requirement-aware method reduced redundancy but lowered pooled ranking quality, while its small coverage gain across all tasks did not persist in the original final subset.
The experiment reuses 48 public R02 tasks with 192 requirements and 666 AI-judged task–document pairs. Across all tasks, agreed coverage increases from 0.739583 to 0.765625, while nDCG@5 declines from 0.692666 to 0.604827. On the original final36, both measures decline. The method recovers 17 requirements and loses 12. Because the candidate pools are incomplete and no fresh holdout was added, these are exploratory findings.
A complementary handoff design crosses eight fictional cases, four formats, and two evidence conditions over two successive agent stages. It specifies 64 trajectories and 128 outputs, each with six exact questions and a 900-byte memory limit. The displaced condition removes a sole supporting fact; the appropriate answer then becomes unknown.
768 scored fields across 128 actual agent outputs: 100.0% matched the frozen keys. There were 0 unsupported non-unknown field answers and 0 incorrect abstentions. Updated-field accuracy was 100.0%; unchanged-field accuracy after the update was 100.0%. Requested memory-limit violations: 0; mean memory size 708.4 bytes, maximum 859 bytes.
The concepts were supplementary and the eight case templates were deliberately simple. Equal success across formats is a ceiling on this test, not proof of equivalence, Carrier superiority or general hallucination reduction. Only six closed-world answer fields per output are scored; every free-prose memory claim is not independently graded.
Generic and Carrier references receive matched resolution mechanics. A separate engineering test finds identical integrity and byte costs under equal-length identifiers. Caching can reduce repeated evidence transfer, but cold lookups add overhead, and full provisioning changes the total. The accounting uses canonical UTF-8 bytes, not model tokens.
The scale experiment reaches one million keys by reusing 6,257 records containing 6,256 distinct definition texts. Its local dictionary timings do not measure full resolver latency, semantic retrieval, or knowledge growth. The release preserves an initial schema-related failure, its normalization amendment, earlier runs, and corrected labels.
The study includes no human receivers or private conversations and does not establish a general hallucination rate or Carrier necessity. The proposed next step is an evidence ledger that protects marginal requirements and explicit unknowns, evaluated on a fresh, harder holdout with matched controls. Protocols, first outputs, scoring rules, costs, and reproduction materials accompany the report.