HOLDLINE calls a list of places, asks them all the same thing, and gives you back only the answers the call actually established. The rest come back empty, with the reason.
Nobody has confirmed a bed. So a discharge coordinator rings a care home and meets this, before a person says a word — and then does it again, seven more times.
| One placement, by hand | On the phone | What you end up with |
|---|---|---|
| Oakfield Care Home | 6:10 | bed, nursing level |
| Riverside Nursing | 4:48 | no answer |
| Belmont House | 7:22 | bed only |
| Five more on the list | 29:40 | three reached |
| Eight homes, one placement | 48:00 | five answers you can act on |
Five placements a week puts that near four hours on the telephone. Those are assumptions, written down so you can change them. The hours are not the expensive part — a “yes, we have a bed” the call never established is, because the patient stays another night and the coordinator starts again tomorrow.
A call result tells you what the model concluded. It does not tell you what was spoken, and nothing in it separates asked and not understood from never asked at all. Watch one call land.
verifiedThe agent raised it and a usable answer came back — quoted, so you can read it yourself.
asked_but_unclearAsked. No usable answer.
unattributedSomething was asked and answered, but no probe for this field matches it. Withheld, and not accused of being invented.
never_askedNothing was asked that this value could answer. A value present here is flagged and withheld.
no_transcriptNothing to check against, so nothing is credited.
400 labelled cases, offline, no calls placed. Every figure below comes back from one
command: npm run eval.
| What a caller ends up believing | Answers returned | Never established | |
|---|---|---|---|
Trusting structured_result |
355 | 178 | 50.1% wrong |
| Through the gate | 133 | 0 | 0.0% wrong |
A benchmark containing only the cases a system handles is marketing. That last figure is the cost of the zero above it, and it prints on every run.
Each was fixed in the same order: add the case to the corpus, measure the damage, then change the check. No transcript, recording or call id from any of them is kept in the repository.
A call returned a whole sentence saying nothing had been established. The gate called it verified — its usable-value check knew only short tokens like unknown. Every value in 400 cases was one word long, so the corpus could not see its own worst case.
A phone tree read our outbound caller id aloud. The transcript kept it in national notation and our redaction matched E.164 only, so a full number reached a file the tool calls shareable. One notation is not every notation.
The agent opened by naming the very thing it had been sent to find out, asked nothing, and those words happened to sit inside the probe. The gate credited a field on less evidence than it claims to require — the exact failure it exists to prevent, from the other direction.
A multi-recipient call returns no transcript turns; the same numbers dialled one at a time do. The batch could verify nothing, so a batch is now one call per place. It reproduces in one command: npm run probe:fanout.
The gate credits a field only when a turn actually asked for it. The compiled task carried a goal, a routing hint and a disclosure — and nothing telling the agent to ask. Two calls came back with the agent having put no question of its own. The tests could not see it: the fake returns transcripts already full of questions.
The agent asked, in plain interrogative order, and the turn was cut off before the question mark. The gate called it never asked and threw the field away — while crediting three other turns from the same call, each about the call rather than the subject, because each carried a mark. Measured: 44 of 44 genuine questions eaten before the fix, 0 of 44 after, with statements of purpose still caught 44/44.
Our fake transport was built from the documentation, so it returned transcripts the platform does not. A fake built from documentation models the documentation. Only a telephone models the telephone.
The console runs the same engine against a local fake. A console reachable off its own machine is locked into simulation, whatever else is set — so a visitor cannot spend credits or ring anybody.
No artifact of ours can prove this dials, so place the call yourself — your CALL-E account, your number. The plan runs with no key and reaches no network.
npm run call -- \
--ask "Are you accepting new patients?" \
--to '+1...' \
--field accepting="accepting new patients" \
--live
--live prints the masked numbers and waits for you to type
LIVE. Drop it and nothing dials. An established field comes back with
the turn that established it, quoted; one the call never covered comes back empty
with the reason.