A context can become less repetitive and less useful at the same time. That is the clearest finding from Carrier R03, “What survives the handoff?” The study follows evidence through selection, bounded memory, reference resolution, and reuse. Its results place limits on an appealing idea: that compact, diverse, verified context will automatically preserve what a later agent needs.
The fact that matters may be a small exception
Imagine a library collection log. The current pickup time is recorded, an earlier proposal is rejected, one item set is explicitly excluded, and the number of eventual borrowers is unknown. A general definition of a decision or constraint may help interpret the log. It cannot tell the next agent which item set was excluded. That answer depends on a particular sentence. Remove that sentence, and a responsible handoff must preserve the absence instead of filling it with a plausible guess.
This is the practical concern behind R03. A memory often preserves the broad topic and discards the exception, condition, or authority that changes the answer. Retrieval can make a similar mistake: five documents may all look relevant while leaving one requirement unsupported. We tested these stages separately because finding a document, preserving a fact, and resolving a reference establish different things.
Less redundancy did not deliver a better selection
The retrieval experiment reused all 48 public R02 tasks, their 192 requirements, and 666 AI-judged task–document pairs. It introduced one fixed method that balances similarity to the question and its requirements against similarity among selected documents. The method was frozen before opening this run’s grades. It was not tuned after seeing results. However, the benchmark was already public, its judged pools were incomplete, and no fresh holdout was added.
Across all tasks, coverage rose from 73.96% to 76.56%, while pooled ranking quality, measured by nDCG@5, fell from 0.692666 to 0.604827. On the original final36 partition, coverage fell from 73.61% to 72.92%, and nDCG fell from 0.674080 to 0.593081. The method gained 17 distinct task requirements and lost 12. The small all-task gain does not justify presenting it as a general improvement.
The documents were less similar to one another, as intended. But selections also contained more documents without agreed evidence for any requirement: an average of 2.54 out of five, compared with 1.92 for dense selection. A diversity score measures difference. It does not establish that the difference helps the task.
One statistics case shows both sides. The method found the missing idea of sequential statistical analysis, restoring support for how repeated looks affect a stopping rule. It also chose the binding of Isaac and Buddhist two-truths doctrine, both judged irrelevant to that task. In a lake-restoration case, environmental monitoring returned to the selection, alongside irrelevant material on memoization. A local repair is useful to inspect, but it cannot outweigh the aggregate loss by itself.
An exhaustive label-using oracle helps locate the remaining gap. Across 121,036 five-document combinations, every requirement with agreed support anywhere in its pool could fit into five slots. Yet 22 of 192 requirements lacked such support throughout their pools. Better selection could recover available evidence; it could not recover evidence missing from those candidates. The oracle sees the labels and is a diagnostic ceiling, never a deployable solution.
A handoff should preserve what is unknown
The conversation design has eight fictional logs, four formats, two evidence conditions, and two successive agent stages: 64 trajectories specifying 128 outputs. Each output answers six questions and writes memory under a 900-byte UTF-8 limit. One condition retains the log; the other replaces its sole scope fact with a relevant but non-answering definition. The second agent sees the first memory, resolved definitions, and an update changing two facts. It does not see the original log or first answers.
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.
The formats compare prose, structured inline material, generic references, and Carrier references under matched semantic content. The harness resolves references for the agents. These are not autonomous resolver trials. Generic references receive the same mechanics so that any ordinary benefit of identifiers is not mistakenly attributed to Carrier. Definitions are supplementary: the scenario logs determine answers. This experiment therefore cannot show that Carrier concepts are necessary.
Six scored fields let us distinguish accepted facts, correct updates, preserved facts, and justified unknowns. They do not grade every sentence in memory or establish a general hallucination rate. The agents use the same configured family, with no independent authentication of model builds. Cohorts process different cases sequentially, allowing carryover; access restrictions are instructions rather than a separate sandbox. There are no human receivers or private conversations in the study.
The reference is smaller; the evidence still has to arrive
Generic and Carrier references pass the same integrity fixtures when given identical validators and equal-length identifiers. Both reject stale or altered reference material. Both can also resolve an authentic definition beside an unsupported claim. A successful lookup shows which text was found. It does not prove what that text implies.
The cost result depends on reuse. For 32 records across 100 handoffs, inline evidence occupies 1,088,900 bytes. Repeating a cold lookup costs 1,531,900 bytes. A persistent cache starting empty lowers transfer to 383,500 bytes. Add full compact registry provisioning, and the total becomes 3,338,602 bytes. This charges a resolver’s full provision plus the receiving client’s first cache miss; preloading the complete registry directly at the receiver instead gives 3,327,002 bytes. These are canonical UTF-8 payload bytes, not model tokens. Every receiver still needs the resolved evidence; an isolated fresh receiver cannot simply inherit the accounting of a warm cache.
The scale test reaches one million keys by reusing 6,257 records with 6,256 distinct definition texts. It measures local dictionary behavior, excluding the network, full validation, semantic search, and model processing. More keys did not create more knowledge. The retained failures and earlier outputs make corrections inspectable, including normalization of historically absent relation lists before timing.
The next design must protect the marginal obligation
The next candidate is an evidence ledger: a compact record of what supports each required fact, which update supersedes it, and what remains unknown. Its value must be tested on fresh, harder tasks with a matched generic control and the same budget. The target is fewer lost obligations and unsupported answers while preserving relevance and correct updates. R03 supplies reasons to attempt that test, not evidence that it has already succeeded.
The R03 publication and reproducibility materials keep these distinctions visible. A handoff earns trust by retaining its qualifications, including the ones that limit the claim being made about the handoff itself.