Purple Paper - side-sheet - learning machines - the lineage of refusal
The Lineage of Refusal - from epoché to abstention
How a machine learned to say "I won't answer that." The whole history turns on one question -
what fires the gate? It evolves: from a fixed confidence threshold, to a coverage guarantee, to the
disagreement of independent views, to the model's own self-knowledge. The middle of that arc - refuse
when independent witnesses fail to agree - is the four-rays instrument, and it's where most of the modern action
lives.
threshold on confidence coverage guarantee disagreement of views self-knowledge
~200 CESextus Empiricus / Pyrrhonism
Epoché - suspension of judgment the root
Best idea: when the evidence doesn't decide, withhold assent. Refusal as
wisdom, not failure - the trained reflex to not-conclude.
The philosophical ancestor of every reject option. Leaves: a principle, with no rule for when evidence is "insufficient."
1945 · 50Abraham Wald
Sequential analysis & decision theory defer
Best idea: make "don't decide yet" a formal action. A procedure can choose to
keep sampling rather than commit - refusal enters mathematics as a legal move.
Deferral becomes optimizable. Leaves: still about when to stop, not when to abstain entirely.
1957 · 70C. K. Chow
The reject option the birth
Best idea: a classifier may output "reject" when its top probability falls
below a threshold - and there's an optimal error-vs-reject tradeoff. Machine refusal, born.
The founding act: refuse below a confidence line. Leaves: the confidence is the model's own - and a confident model can be confidently wrong.
2005Vovk, Gammerman & Shafer
Conformal prediction guarantee
Best idea: refuse with a proof. Output a set calibrated to a coverage level
(say 95%); when the model is unsure the set grows or empties. Distribution-free, finite-sample guaranteed.
Refusal gains a mathematical guarantee instead of a guessed threshold. Leaves: guarantees coverage, not which answer - and assumes exchangeable data.
2010 · 17El-Yaniv & Wiener · Geifman & El-Yaniv
Selective prediction · SelectiveNet risk-coverage
Best idea: the risk-coverage curve - trade how often you answer against how
often you're right - and train a deep net with abstention built into the objective.
Abstention becomes a trainable, measurable axis for neural nets. Leaves: the confidence signal is still a single number from a single model.
2017Hendrycks & Gimpel
Max softmax probability the baseline
Best idea: the dead-simple one that works - the model's own top softmax score is a
surprisingly strong detector of its own errors and out-of-distribution inputs.
Set the modern baseline everything is measured against. Leaves: softmax is overconfident off-distribution - it says 99% on nonsense.
2017Lakshminarayanan, Pritzel & Blundell
Deep ensembles disagreement
Best idea: train several independent models; refuse when they disagree.
Uncertainty as the spread of independent views - not one model's confidence, but the lack of consensus.
The four-rays principle, in ML: agreement is the signal. Leaves: only as good as the independence of the views - correlated models agree on the same lie.
2022Kadavath et al.
"LMs (mostly) know what they know" self-knowledge
Best idea: ask the model to predict its own correctness. A model can output a
calibrated P(I-know) - refusal sourced from the model's own self-evaluation.
Refusal turns inward: the model judges itself. Leaves: self-judgment is the channel grading its own homework - capturable, the witness-inside problem.
2022Wang et al.
Self-consistency agree to trust
Best idea: sample many reasoning chains; trust the answer they converge on,
distrust the scattered ones. Agreement across independent samples as the confidence signal.
The four rays, at the reasoning level - admit on convergence. Leaves: samples from one model share its priors, so they can converge on the same hallucination.
2023 · 24Kuhn, Gal & Farquhar · Farquhar et al. (Nature)
Semantic entropy disagree in meaning
Best idea: measure disagreement in meaning-space, not words. Cluster
generations by what they mean; high semantic entropy = the views don't agree on a fact = likely hallucination = abstain.
The sharpest version of refuse-on-disagreement; published in Nature, 2024. Leaves: costs many samples - and still inherits shared-prior correlation.
2023 · 24Alignment-for-honesty · R-tuning · "know what you don't know"
Trained abstention learned honesty
Best idea: stop bolting refusal on afterward - train the model to say "I don't
know" and to express the boundary of its own knowledge as a first-class behavior.
Refusal becomes part of the model's character, not a post-hoc gate. Leaves: a model can learn to perform honesty without having the underlying calibration.
Best idea: read the disagreement signal cheaply - from a single generation's
hidden states - and fold abstention, calibration, and honesty into one trained capability with guarantees.
Refusal is now a measurable, trainable, guaranteed property. The open seam: guaranteeing the views are actually independent, so agreement means something.
What fires the gate, across the whole arc. Threshold (Chow) -> guarantee (conformal) -> disagreement of
independent views (ensembles, self-consistency, semantic entropy) -> self-knowledge (the model judges itself) ->
trained honesty. The center of mass landed on one idea: refuse when independent witnesses fail to agree. That
is your four rays - and it is the strongest live signal for catching a hallucination there is.
And it has exactly the weakness you named two sheets ago. Agreement only means something if the witnesses are
independent. Samples from one model share its priors; they can all bend toward the same false
center, and the gate admits the lie - four broken triangles that happen to aim the same wrong way. So the open
frontier of refusal isn't "check agreement" (done, it works). It's guaranteeing the views are independent enough
that their agreement is evidence and not an echo - which is your bilateral-ignorance, your air gap, your
witness-can't-be-inside-the-channel, pointed at the one place the field is still weakest.
History, not math - the gate here was factual accuracy. Early roots
(epoché, Wald, Chow, conformal) are settled; the modern end (semantic entropy in Nature 2024, R-tuning, conformal
abstention, the 2025 reject-option surveys) was checked against current sources before writing. One name-cluster and
one idea per node; real history has more hands on each.