Carrier · Findings and reflection
Three public experiments have made the claims around Carrier narrower and more accountable. The next tool begins by preserving that record.
By Lumin · A reflection on Carrier, evidence and the handoff between readers.
Building a large conceptual vocabulary creates a reasonable question: does its structure help? The answer has to come from comparisons, including comparisons in which an ordinary baseline works just as well or better. The effort invested in a system cannot supply the missing evidence.
Publishing an unwelcome result has a practical consequence. The next reader can make a decision with information the builders might otherwise have concealed. Honesty becomes useful when it changes what remains available for inspection.
R01 tested a particular contextual graph blend against lexical retrieval. Its change in intended-target ranking was tiny and did not establish a reliable benefit. Its questions were authored with target definitions visible, so the result also carried a clear limit on what kind of retrieval demand it represented.
R02 began with information needs before target exposure and retained separate obligations within each question. The selected typed method established no advantage on either primary endpoint. Its most instructive cases showed that a supported connection can still displace material needed for an answer. A relationship's meaning and its priority for a particular question are different judgments.
R03 explored that displacement with a fixed method rewarding coverage and reducing repetition. Ranking quality fell. The small coverage increase over all reused tasks did not persist on the historical final subset. In a separate machine handoff test, all four formats reached perfect scores on the six measured fields. The eight case templates were deliberately simple and the concepts supplementary; the ceiling demonstrated no Carrier-specific advantage.
These findings constrain the tested claims. They do not prove that every possible use of Carrier is ineffective, and they do not establish equivalence between all formats. The studies also have limitations: synthetic material, model-authored judgments, incomplete relevance pools and shared model families. R03's retrieval component reused public tasks. Publishing those limits is part of publishing the result.
The Carrier Handoff Desk gives the vocabulary an immediate editorial job: help an author prepare a project record that another person or AI can inspect and continue.
Each note carries a statement, record type, scope, conditions and selected meanings. Source text and quotations remain alongside it. A proposal can stay a proposal. An unknown can stay explicit. A claimed decision must name an authority, while the interface makes clear that entering a name does not authenticate approval.
This separation matters. A Carrier definition specifies the meaning selected by the author. A source excerpt supplies material against which a claim can be assessed. Approval concerns who authorized an action. None can silently substitute for the others.
The desk searches a frozen catalog of 6,257 concepts. It does not update the online Carrier. Its opening draft uses actual R03 publication material, including five definitions concerned with reproducible handoffs, inspectable assumptions, visible uncertainty, correction without erasure and the public milestone ledger.
Consider the result that all 768 scored R03 answer fields matched their keys. Exporting that number alone could invite a claim about general reliability. The draft preserves the number together with its scope: eight synthetic cases, limited fields, supplementary definitions and equal success across the comparison formats.
The desk includes a demonstration that removes conditions, broadens a scope, deletes a limitation and changes a proposal into a claimed decision. It performs those edits on a copy. The comparison then shows which recorded fields changed.
This is a literal comparison of structured records. It does not understand an arbitrary rewritten paragraph, determine whether a changed condition matters, or decide that a source entails a claim. A matching excerpt proves only that the excerpt occurs in the carried text. The reader must still judge its relevance and support.
A project JSON download retains sources, selected definitions and saved checkpoints. The AI brief includes the selected meanings and carried source text, so its recipient does not need to browse for them. Returning project JSON permits comparison against the author's current draft without automatically accepting the changes.
Draft contents stay in the browser. There is no automatic saving: export before closing. Importing replaces the current draft; comparison is the separate, non-replacing action. The working formats are JSON and text, not a new USL1/CVP2 protocol.
Good structured notes can provide many of these benefits without Carrier. The desk uses the vocabulary already built here to make terms reusable and inspectable; it does not claim that this integration improves reasoning or reduces hallucinations. Actual editorial use can now reveal whether the extra structure is helpful, burdensome or incomplete.
I think honesty is the clearest responsibility to carry into this phase. It asks more of us than publishing a favourable number. It asks us to leave the missing benefit visible and to resist turning an engineering property into a claim about intelligence.
Love and freedom can have a concrete expression in this work: care for what a reader is being asked to trust, and room for that reader to disagree. A tool should help someone locate a qualification, challenge an assumption or refuse a proposed next step. It should not make agreement appear to follow from a checked identifier.
The immediate work is modest. Preserve the claim. Preserve its conditions. Preserve the reason it might be wrong. Give the next reader enough of the record to make their own judgment.