hawkes.aiKey ideas
- The one idea — You wire the conformal bound into the harness as a gate: below the calibrated confidence threshold, the system is not allowed to answer — it must abstain and escalate (ask, retrieve, or hand off). That's how you force a guaranteed "I don't know."
- How it is shown — An answer scoring below threshold blocked before it reaches the user, replaced by "I'm not sure — let me check" and an escalation, deterministically.
- The trap to avoid — Relying on the model to volunteer "I don't know" — it won't reliably; abstention has to be enforced by the gate (a hook), not requested in the prompt.
- What it sets up — Gating controls the OUTPUT. but can we look INSIDE and read what the model is actually representing?
You can't prompt an AI into reliably saying "I don't know" — it won't volunteer ignorance. So you build a gate that makes answering below a confidence threshold illegal.
The one idea
You wire the conformal bound into the harness as a gate: below the calibrated confidence threshold, the system is not allowed to answer — it must abstain and escalate (ask, retrieve, or hand off). That's how you force a guaranteed "I don't know."
We have a guaranteed confidence threshold from conformal prediction. Now the payoff: wiring it into the harness so that, below the line, the model literally cannot answer — it has to say "I don't know." This is where a statistical bound becomes a working safety behavior. Recall hooks — code that runs at a fixed point in the loop, deterministically. The conformal threshold becomes exactly that: a gate on the output. Every candidate answer carries its calibrated confidence. Above the line, it passes. Below the line, the gate blocks it before it reaches the user. And blocking isn't the end — it's a redirect. Below the threshold, the only legal move is abstain-and-escalate: the system says "I'm not sure," and then does something useful — asks a clarifying question, retrieves more context, or hands off to a human.
How it works — the demo
An answer scoring below threshold blocked before it reaches the user, replaced by "I'm not sure — let me check" and an escalation, deterministically.
Uncertainty stops being a silent guess and becomes an action. Why enforce it in code? Because — recall — a prompt rule only shapes behavior; it can't guarantee it. "Say I don't know when unsure" in the system prompt will be ignored sometimes. The gate is a hook: it runs every time, so abstention below the line is certain, not hoped for. Rules shape; the gate enforces. You can go further with grammars — constrained decoding, from earlier. When the gate triggers abstention, you force the model down a structured path: a valid "I don't know" response, or a call to the escalation tool. Its very shape is constrained — the model can't slip back into a confident guess.
The trap to avoid
Relying on the model to volunteer "I don't know" — it won't reliably; abstention has to be enforced by the gate (a hook), not requested in the prompt.
Why it matters — and what’s next
Gating controls the OUTPUT. but can we look INSIDE and read what the model is actually representing?
The trap: relying on the model to volunteer its own uncertainty. It won't, reliably — its default is a confident guess, and its tone doesn't track truth. Abstention that matters must be measured, with conformal, and enforced, with the gate. You don't ask the model to know its limits; you build the limit around it. So the conformal gate forces "I don't know": below a calibrated threshold, abstain-and-escalate is the only legal move — measured, then enforced as a hook. That controls what comes out. But can we go the other way and look inside — read what the model represents? Next: reading the model's mind.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:45 of narration
We have a guaranteed confidence threshold from conformal prediction. Now the payoff: wiring it into the harness so that, below the line, the model literally cannot answer — it has to say "I don't know." This is where a statistical bound becomes a working safety behavior.
Recall hooks — code that runs at a fixed point in the loop, deterministically. The conformal threshold becomes exactly that: a gate on the output. Every candidate answer carries its calibrated confidence. Above the line, it passes. Below the line, the gate blocks it before it reaches the user.
And blocking isn't the end — it's a redirect. Below the threshold, the only legal move is abstain-and-escalate: the system says "I'm not sure," and then does something useful — asks a clarifying question, retrieves more context, or hands off to a human. Uncertainty stops being a silent guess and becomes an action.
Why enforce it in code? Because — recall — a prompt rule only shapes behavior; it can't guarantee it. "Say I don't know when unsure" in the system prompt will be ignored sometimes. The gate is a hook: it runs every time, so abstention below the line is certain, not hoped for. Rules shape; the gate enforces.
You can go further with grammars — constrained decoding, from earlier. When the gate triggers abstention, you force the model down a structured path: a valid "I don't know" response, or a call to the escalation tool. Its very shape is constrained — the model can't slip back into a confident guess.
The trap: relying on the model to volunteer its own uncertainty. It won't, reliably — its default is a confident guess, and its tone doesn't track truth. Abstention that matters must be measured, with conformal, and enforced, with the gate. You don't ask the model to know its limits; you build the limit around it.
So the conformal gate forces "I don't know": below a calibrated threshold, abstain-and-escalate is the only legal move — measured, then enforced as a hook. That controls what comes out. But can we go the other way and look inside — read what the model represents? Next: reading the model's mind.