# Source → concept → construction → limit

Foray 238, first essai, 18 October 2026. Primary methods were read before construction; this is not a replication of the four papers. No external dataset or private corpus was used.

## 1. J. R. Quinlan (1986), Induction of Decision Trees

[Publisher PDF](https://link.springer.com/content/pdf/10.1007/BF00116251.pdf), *Machine Learning* 1, 81–106. Read §§3–4, especially the induction task, recursive partitioning, original windowing and information-gain criterion (printed pp. 85–91).

- **Concept:** classification rules can be induced from attribute-labelled examples, with greedy information gain selecting successive tests; training fit and unseen-case induction are distinct.
- **Construction:** `fit_tree` is a full-data categorical variant whose actual fitted branches and training support are displayed. It never receives the authored adequacy function.
- **Limit:** original ID3's windowing is omitted; no pruning or treatment of uncertain measurements is attempted. The attribute vocabulary already contains the information needed to express the constructed criterion. Nothing here tests concept discovery from professional documents.

## 2. T. M. Cover and P. E. Hart (1967), Nearest Neighbor Pattern Classification

[Author's institutional PDF](https://isl.stanford.edu/~cover/papers/transIT/0021cove.pdf), *IEEE Transactions on Information Theory* 13(1), 21–27. Read introduction and §§II–III: observed feature vectors, labelled reference samples, distance and nearest-neighbour decision.

- **Concept:** nearby stored examples can support classification without fitting an explicit tree.
- **Construction:** nearest-history comparison uses the identical feedback corpus, unweighted Hamming distance and an openly declared vote over all nearest ties. The page exposes actual neighbours for each case.
- **Limit:** the metric and tie handling are this essai's choices. No asymptotic bound from the paper is transferred to a tiny, categorical, deterministic world. This is not an embedding, semantic search or LLM-history experiment.

## 3. N. Shinn et al. (2023), Reflexion: Language Agents with Verbal Reinforcement Learning

[Primary full text, v4](https://arxiv.org/html/2303.11366v4). Read §3 and Algorithm 1, the actor/evaluator/self-reflection roles, and Appendix B.1's WebShop limitation.

- **Concept:** feedback, a retained lesson and a subsequent attempt can be separate parts of a loop; selecting an evaluator remains consequential.
- **Construction:** feedback labels, the adapting classifier and the scoring criterion are explicitly separated. Biased reviewer labels make that separation visible.
- **Limit:** there is no verbal reflection, LLM actor or Reflexion replication here. The paper also reports an unsuccessful WebShop setting; a feedback loop alone is not a sufficient explanation of improvement.

## 4. A. Madaan et al. (2023), Self-Refine: Iterative Refinement with Self-Feedback

[Primary full text, v2](https://arxiv.org/html/2303.17651v2). Read §2, Algorithm 1 and equations 1–4; §3's evaluation procedure and task-dependent limitations.

- **Concept:** generation, feedback and revision can improve a current output without model training.
- **Construction:** this distinction sets the experimental question: can a feedback-derived lesson help a separate held-out case, and remain appropriate after the intended use changes?
- **Limit:** the desk does not perform self-feedback or answer refinement. Its across-case rule induction provides no evidence for or against Self-Refine's reported language-model results.

## What is original to this essai

The review-desk scene, feature vocabulary, two use-specific adequacy rubrics, simulated presentation bias, grouped split and correction schedule are authored constructions. They are neither independently validated professional standards nor observed reviewer behaviour. The implementation and local freeze receipt make this bounded test repeatable; they do not establish universal research novelty.