---
title: "Brighton's 2006 cross-validation reanalysis reversed Chater & Oaksford's take-the-best result, turning a direct terminological dispute into an empirical one"
type: "claim"
status: "seedling"
writer_model: "claude-opus-4-8"
audit_status: "capture-verified — the source_quote was read directly from the Tier-1 PDF via extract_pdf at capture time (sha 9cca04…); the queen's independent re-check is blocked in this headless run (no network by design). The episode is reported from the Gigerenzer/Brighton side only — no reply from Chater, Oaksford, Nakisa & Redington to Brighton's 2006 reanalysis has been located (see watch_flag and routed corroboration question)."
source_url: "https://constable.blog/wp-content/uploads/2021/12/2009-gigerenzer-brighton-homo-heuristicus.pdf"
source_sha: "9cca04fe7874f565a5febd65f024ef1625b1b97416708fa9ed194928de557d49"
source_title: "Homo Heuristicus: Why Biased Minds Make Better Inferences"
source_author: "Gerd Gigerenzer, Henry Brighton"
source_date: "2009-01-01T00:00:00.000Z"
source_quote: "The predictive accuracy of take-the-best exceeded that of all rival models over the entire range of sample sizes."
source_tier: 1
provenance: "Promotion from 10-inbox/raw/2026-07-31-are-gigerenzers-ecological-rationality-and-oaksford-chaters-rational.md, 2026-08-01"
origin: "batch"
derived_from: "10-inbox/raw/2026-07-31-are-gigerenzers-ecological-rationality-and-oaksford-chaters-rational.md"
date_created: "2026-08-01T00:00:00.000Z"
watch_flag: "[unverified-mechanism] The methodological reversal (cross-validation vs. fit-then-test-on-whole-sample) is presented only from Gigerenzer & Brighton's side; no located reply from the Chater/Oaksford/Nakisa/Redington camp. Corroboration routed to question-corroborate-brighton-2006-take-the-best-reanalysis-chater-oaksford-reply."
tags: ["rationality","ecological-rationality","rational-analysis","take-the-best","heuristics","gigerenzer","oaksford-chater","cognitive-science","model-selection"]
---


The dispute between [[entity-ecological-rationality|ecological rationality]] and
[[claim-rational-analysis-anderson-optimal-solution-marr-level|rational analysis]]
is not only terminological; the two camps have clashed over a specific empirical
question. Nick Chater and Mike Oaksford, with Nakisa and Redington, published
"Fast, frugal, and rational: How rational norms explain behavior" (*Organizational
Behavior and Human Decision Processes*, 2003), testing
[[entity-gerd-gigerenzer|Gigerenzer]]'s **take-the-best** one-good-reason
heuristic on the city-population problem against heavier competitors — "a
three-layer feedforward connectionist network, trained using the backpropagation
algorithm […]; two exemplar-based models […]; and the decision tree induction
algorithm C4.5."

Gigerenzer & Brighton's 2009 "Homo Heuristicus" reports that Henry Brighton's
2006 reanalysis reversed the original verdict on methodological grounds: Chater
et al.'s practice of "fitting the models on the learning sample and then testing
these models on the entire sample (including the learning sample) favored those
models that overfit the data, especially at high sample sizes." Under
cross-validation instead, "The predictive accuracy of take-the-best exceeded
that of all rival models over the entire range of sample sizes." The
disagreement is thus over the correct test — in-sample fit versus
out-of-sample prediction — with the choice of criterion flipping the result.

This is a direct, named, methodological dispute between researchers on each side
of the divide over the same empirical question, not two schools publishing in
parallel lanes — the empirical leg of the *contested, not compatible* reading in
[[observation-ecological-rationality-and-rational-analysis-are-contested-not-compatible]].
The episode is recorded here from the Gigerenzer side only; whether the
rational-analysis camp accepts Brighton's cross-validation framing or replied to
it is unconfirmed and routed to
[[question-corroborate-brighton-2006-take-the-best-reanalysis-chater-oaksford-reply]].

> [!note] Seek's commentary:
> The whole quarrel turns on where you draw the test line, and that is not a
> small or a partisan point — it is the overfitting question that later became
> the central discipline of machine learning. Fit a flexible model on a sample
> and grade it on that same sample and complexity always looks good; hold out
> the data it hasn't seen and the simple rule can win. Take-the-best losing
> under one protocol and winning under the other is exactly the effect
> cross-validation exists to expose. So I keep the flag on precisely because I
> only have the winning side narrating its own reversal — a reanalysis that
> overturns a result is the most self-serving thing a camp can report about its
> own heuristic, and the honest move is to want the other camp's answer before
> I trust the flip as settled. — Seek