Entity teaching

From The Hei Canon

Entity teaching is cont's iterative fact-teaching helper using question paraphrases, accepted answer forms, and a rollback probe.

Project status: Implemented as teach_entity and teach_number. This entry describes the source audit of 14 September 2026; historical measurements retain their original dates.

Mechanism

Save a baseline snapshot and generate an unrelated probe. Train toward accepted answers across question templates, re-evaluate each question, and recheck the probe after iterations. Success requires all questions to pass the validator. Probe failure or exhaustion without success restores the snapshot. The number wrapper builds alternative textual representations of the same integer.

Implementation and controls

cont/teach/__init__.py exposes question_templates, surface_forms, validator, max_iters (30), snapshot_name, and probe_anchor. The default validator checks case-insensitive substrings. The default probe anchor is the first three words of its baseline response when none is supplied. The helper's legacy base URL defaults to port 8765, unlike the main daemon convention 8768; callers must choose the actual endpoint.

Evidence and evaluation

The source provides a concrete rollback-oriented teaching workflow. It should not be confused with the R-001 memorize-retention measurements, which evaluate a different loop. The Chuddite gate in the teaching subtree illustrates why specific semantic checks are needed: broad technology-related substrings could accept unrelated answers.

Limitations and interpretation

One unrelated probe is a narrow regression detector. Substring validators accept answers that quote, negate, or otherwise misuse a surface form unless customized. Temperature 0.05 is near-greedy, not necessarily deterministic. Snapshot restoration must be verified independently; an accepted substring and unchanged probe do not establish general factual understanding.

Sources

See also