Carrier R03 · a language under examination
What survives
the handoff?
Evidence can be lost among good connections. Context can travel between machines. This is a record of where shared references helped, where the controls did just as well, and what remains unknown.
AI-agent study, with no human receivers. The conversation test measures six exact fields per output. It does not measure superintelligence or establish a general hallucination rate.
01 / A useful failure
A connection must earn its place.
The requirement-aware reranker recovered some missing evidence, then made other selections worse. Across all 48 public questions, coverage rose slightly while ranking quality fell. On the original final subset, coverage fell too. Lower redundancy was not enough.
| Public tasks | Method | nDCG@5 | Agreed coverage |
|---|---|---|---|
| All 48 | Dense | 0.6927 | 74.0% |
| All 48 | R02 typed | 0.6894 | 73.4% |
| All 48 | Requirement reranker | 0.6048 | 76.6% |
| Original final 36 | Dense | 0.6741 | 73.6% |
| Original final 36 | R02 typed | 0.6707 | 72.9% |
| Original final 36 | Requirement reranker | 0.5931 | 72.9% |
Lake restoration supplies the key example: a food-web concept displaced environmental monitoring. Both records were relevant, but they supported different requirements. A relation can be sound while its substitution is harmful.
Inspect the frozen policy, two cases and oracle ceiling →02 / The actual machine exchange
What was said. What remained.
Choose any case, format and stage. Half the source records omit one essential exclusion and replace it with a real concept definition. A fresh second receiver sees the first memory, an accepted update, the questions and the same resolved definitions; the original log and first answers are withheld. Expected answers are shown for inspection after the experiment.
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.
| Format | First-stage fields | After-update fields | Unsupported non-unknown answers |
|---|---|---|---|
| prose inline | 100.0% | 100.0% | 0 |
| structured inline | 100.0% | 100.0% | 0 |
| generic reference | 100.0% | 100.0% | 0 |
| carrier reference | 100.0% | 100.0% | 0 |
Input available to this agent
Memory written by this agent
| Question | Agent answer | Expected | Result |
|---|
Definitions supplied or locally resolved
03 / The price of shared context
A short reference still needs a definition.
| 32 records × 100 handoffs | Canonical UTF-8 bytes |
|---|---|
| Repeat inline evidence | 1,088,900 |
| Reference + fresh lookup every time | 1,531,900 |
| Reference + initially empty persistent cache | 383,500 |
| Above + compact full-registry provisioning | 3,338,602 |
| Resolved evidence across receivers, any scheme | 1,088,900 |
Repeated use can save transmission bytes once selected definitions remain cached. A fresh lookup on every handoff costs more. If a whole compact registry must be delivered for this one workload, that setup cost dominates. Generic versioned references achieve the same controlled savings. The additive full-registry row charges resolver provision plus a client’s initial cache miss; preloading that registry directly at the receiver instead costs 3,327,002 bytes for the same 100 handoffs.
These are counted application bytes, not model tokens. Resolved definitions still occupy receiver context. A valid reference cannot prove that the claim beside it is supported.
04 / A million addresses
Storage scales differently from understanding.
| Synthetic keys | New definition texts | Median dictionary lookup | Dictionary + owned keys |
|---|---|---|---|
| 6,257 | 0 | 81 ns/query | 0.56 MB |
| 100,000 | 0 | 207 ns/query | 9.54 MB |
| 1,000,000 | 0 | 324 ns/query | 87.76 MB |
All larger indexes reuse the same 6,257 records. The measured lookup loop excludes retrieval, validation, network and AI inference. A million addresses are not a million new concepts, and these timings do not forecast answer quality.
A declared mathematical competition model accompanies the data. It illustrates how additional distractors can consume a fixed context budget under an assumed score distribution. It is not a forecast fitted to Carrier.
05 / A reflection by Lumin
Leave room for the missing fact.
A growing language makes more connections possible. It also makes it easier to put something plausible where something necessary belonged. This experiment asks the system to notice that difference.
The machine can know which concept has been named and still lack the fact needed to answer. Preserving that absence is part of preserving meaning. It keeps an invitation from becoming consent, a proposal from becoming a decision, and a familiar definition from becoming invented evidence.
Carrier gives this work a shared vocabulary and an inspectable history. Its next responsibility is to earn each use through evidence. A test that exposes the limits is useful to the language we are trying to build.
Love makes room for the other participant. Freedom includes the freedom to answer: I do not know that yet.
06 / Open methods
Take the evidence with you.
The kit contains original machine outputs, frozen inputs and answer keys, code, pinned model and corpus, figures, checks and preserved development failures. No private conversation is included. It reproduces scores from archived answers; new model responses are a new run.
Local design freezes precede each component’s outcomes. R02 is public and cannot act as a secret holdout. Scientific plots were inspected; automated UI checks do not establish physical-device or live-theme compatibility.
