Most decisions don’t need a chain of thought. They need a fast, inexpensive call with an honest confidence. Tacet builds structured decision models: the tacit layer that agentic AI is missing.
My card was declined at checkout, but I can still log in and see my order history.
paymentsCard and billing problems0.00accountLogin and access0.00shippingDelivery issues0.00Based on the information provided, it appears likely that the order has been delivered. The tracking status indicates the package was left on the front porch this morning, which typically suggests a successful delivery. However, it’s worth noting that tracking information can occasionally be inaccurate, so you may want to confirm with the recipient. So, to answer your question: yes, it was most likely delivered.
“We can know more
Michael Polanyi, The Tacit Dimension, 1966
than we can tell.”
Experts can’t fully explain what they know: the nurse who senses a patient is slipping, the analyst who spots a fraud ring, the editor who knows a headline will land. Rules fail because that knowledge was never written down. Language models learned it from us, but they can only express it by talking. Tacet expresses it the way experts do: as a judgment, with a confidence.
| Property | Rules | Structured decision models | Language models |
|---|---|---|---|
| Learns judgment from examples, not written rules | — | ✓ | ✓ |
| Answer is always one of your options | ✓ | ✓ | Usually |
| Tells you how sure it is | — | ✓ | Sounds sure either way |
| Same input, same answer | ✓ | ✓ | Varies run to run |
| Fast and cheap enough for every event | ✓ | ✓ | — |
Tacet α is an alpha. We’re still measuring its latency and cost against matched workloads.
A support ticket, a payment, a listing, a pause on a phone call. Most real decisions are small, repeated millions of times, and due in milliseconds. They don’t need a plan or a paragraph. They need the right call, at the right cost, with an honest confidence.
Decided every couple of hundred milliseconds, on every live call.
State
Transcript so far: “I’d like to move my appointment to, um…” Silence: 640 ms. Last word was a filler.
no0.00yes0.00Why it needs a structured decision modelJump in too early and the caller feels cut off. Wait too long and the agent feels slow. The decision has to fit in the pause between words, which leaves no time to write and parse a sentence.
Illustrative outputs.
Today’s agents route every judgment through explicit reasoning, including the dozens of small calls inside a single task. It’s like asking an expert to write an essay before every glance. Tacet is the tacit layer: it makes the small calls in one pass each and leaves the hard, novel problems to your reasoning model.
Tap a short mark to see the check behind it.
Experts are trusted because they know how sure they are. A calibrated model’s 0.9 is right about nine times in ten, so you can set a threshold and know in advance how much runs on its own and how much reaches a person.
Simulated stream from a well-calibrated model. Drag the line to move the threshold. Measure on your own traffic before choosing one.
8,400 training questions and about 10 million trainable parameters, averaged over three seeds on 1,100 held-out questions.
Tacet runs on our servers. Send a state and your questions; get calibrated probabilities back as JSON.
from tacet import Tacet
client = Tacet()
result = client.decide(
model="tacet-alpha",
state="Tracking shows the parcel was left on the porch.",
questions={"delivered": {"type": "noul", "instructions": "Was it delivered?"}},
)
result.answers["delivered"].probs["yes"] # 0.97Most of what we know, we never say. Tacet puts that knowledge to work.
Read why we started Tacet