abstain-right

abstain-right

Every frontier model is trained to always produce an answer — its own creator denies it the option to say "I can't tell." Estelle gives that option back, and makes it enforceable: the clarify stage (needs_clarification, the config.clarify step in best_answer) and the deterministic grounding gate (GroundingReport.is_grounded, served as POST /verify → {grounded, ungrounded, arity_errors}) decide answer-or-abstain from ground truth, not from the model's felt confidence. This is Estelle's founding thesis — it exists to stop the confident hallucination that would otherwise pass review.

When to run

Before writing any answer that asserts a fact about the repo, an API, or a team decision — i.e. any claim that could be wrong. Run it hardest when recall came back thin, when the request names a file or function that could be one of several, or when you feel the pull to fill a gap from training memory. If the claim is fully grounded, this costs one cheap check and you answer as normal.

Procedure

1. Recall before you reason. Pull the repo/team memory with retrieve_cited (hybrid RRF, recall@4 90.7%, control 2.4%) — these are the sources you would cite. If recall returns nothing on the load-bearing claim, treat that as an early abstain signal, not an invitation to improvise.
2. Run the clarify check first. Before spending tokens on an answer, needs_clarification grounds the request in that recalled context and asks: is the target file/function ambiguous, a behavior unspecified, or are there multiple plausible readings? If yes, return ONE specific question and stop — asking beats guessing.
3. Draft only what the context supports. Write against the recalled sources, using only symbols and signatures that appear there. Do not reach past the context to complete a pattern from memory.
4. Submit the draft to the grounding gate. ground_report_and_repair → GroundingReport checks every repo symbol you referenced against the real symbol graph (ungrounded, arity_errors, third_party). is_grounded is the deterministic verdict — POST /verify returns the same {grounded, ungrounded, …}.
5. Read the verdict, not your gut. If is_grounded is False the gate names the invented symbols and wrong-arity calls. Your subjective confidence is irrelevant here; the compiler-accurate check is truth.
6. Choose the honest exit. If the gap is ambiguity → ask the clarifying question from step 2. If the gap is a missing fact or a symbol the repo simply doesn't define (and can't be repaired from context) → abstain: say "I don't know" or "I need X first", naming exactly what is missing to answer correctly.
7. Never fabricate to satisfy the gate. Inventing a plausible-looking symbol to make the answer read complete is the exact failure Estelle exists to block. The right to abstain is also the requirement: an ungrounded claim must become a question or an admission, never a guess dressed as an answer.
8. Bank the catch. When the gate rejected an invented symbol, learn_from_grounding records it as a lesson keyed on this prompt, so the next similar ask recalls it and you don't reinvent it.