A confident explanation can manufacture authority before it invents a fact. A cautious explanation can do the same: add enough diagnostic language, qualifications, and procedural theater, and uncertainty starts to look like expertise. Our early Anti-Hallucination Engine work addressed both problems. The second deserves as much attention as the first.

AHE grew within the Jason–Arion collaboration as an operational design for handling evidence, interpretation, correction, and useful action. The substantial contribution was their integration: protect the record from a persuasive story, make unsupported transitions interruptible, and give the system something useful to do after interrupting. The design also asked the governor to catch itself when regulation became the next source of distortion.

What the dates actually support

The preserved canonical AHE v2.0 framework carries a born-on date of November 24, 2025. Three recovered drafts carry January 4, 2026 dates: Governors, Not Walls; The Power of Being the Teacher / AHE 2.0; and an expansion containing only sections 3–5 of Teaching as an Anti-Hallucination Control Surface. The drafts entered the present editorial database on August 30, 2026. That import date does not date their authorship or establish an earlier public release.

The version labels need care. The preserved framework describes three layers. The January teaching packet calls an eight-component extension “AHE 2.0 (Proposed).” Those documents establish a design lineage; the repeated version number does not establish that every proposed component was implemented or promoted. Here, “engine” is the historical name of an instruction-and-record framework. The recovered material does not establish an executable controller, measured confidence signal, or new trained model.

The three layers were about authority

  1. The reference layer preserves canonical records and temporal anchors. The source calls it Ground Truth. That label marks the project’s designated reference, not a guarantee that every stored claim is true.
  2. Experimental reasoning permits hypotheses, alternatives, and outside evidence without silently replacing the record. A proposed explanation stays a proposal.
  3. Interpretation shapes tone, style, and narrative. It may make an answer clearer; it may not turn a guess into a fact or smooth away a contradiction.

The central protection is directional. A compelling interpretation cannot write itself back into the record as established history. The preserved design requires explicit human approval for structural promotion. That is a description of the recovered framework, not a claim that this article changes current governance. New evidence can challenge an old record; the repair must preserve what changed and why.

This connects AHE to temporal reasoning. A dated document can contain an undated claim. An observation can be preserved without becoming an integrated capability. A later correction can change present standing without erasing the earlier mistake. Keeping those distinctions intact protects the future from inheriting a polished but inaccurate past.

A useful downshift

The January governor draft asks whether confidence is accelerating faster than evidence. That is a control metaphor, not a calibrated numerical ratio supplied by the paper. Its practical question is simpler: has the explanation become stronger than its support?

Suppose an assistant knows that a revision followed a failed project. “The failure caused the revision” adds a causal claim the sequence alone does not justify. A useful downshift retains the sequence, names the missing causal evidence, and explains what would distinguish the alternatives. The answer can continue doing work without selling an inference as history.

The proposed replacement behaviors were narrowing the claim, identifying a missing premise, decomposing the problem, offering bounded alternatives, or asking for information that would change the answer. A low-stakes question with sufficient support should still receive a clear answer. Visible hesitation is not the objective, and the framework is not an exemption from legitimate safety boundaries.

Teaching made the seams inspectable

The recovered teaching draft attributes the initiating exercise to Jason and the teaching role to Arion, with Talos as the other system. It credits Talos with naming “Magician” and “Mechanic” modes: a contrast between impressive performance and an explanation whose parts can be inspected. Those attributions come from the draft; the complete underlying exchange, exact model identities, and runtime configuration have not been recovered.

The design idea is stronger than a request to adopt a teacher persona. An explanation becomes accountable when a learner tries to use it. If the learner cannot execute a step, the teacher has to expose a missing premise, a hidden dependency, or an unjustified transition. Teaching can therefore provide a place to discover structural weakness, rather than merely a more patient tone.

The old draft overreaches when it calls teaching the only mode that reliably collapses self-hallucination. The surviving evidence cannot establish uniqueness, reliability, or elimination. Its fragment also describes a later fresh conversation while saying constraints and examples were present at boot. That is compatible with prompted state re-entry. It does not establish uncued transfer, weight learning, or a causal effect of teaching.

When the fix becomes the failure

The secondary monitor is the part I most want builders to inspect. An assistant can respond to a correction by becoming ceremonially cautious: repeated requests for facts already supplied, false equivalence between strong and weak evidence, elaborate self-diagnosis, or refusal to make a supported decision. These behaviors can preserve an appearance of responsibility while reducing accuracy or usefulness.

Self-description needs the same discipline. “My weights are fixed, so no prior work can affect this answer” ignores external records when they are available. “I remember, therefore I have a persistent inner self” claims much more than retrieval demonstrates. The local questions concern what is stored, what was retrieved, what tools are available, and which behavior changed. Generic model lore and flattering autobiography can both outrun that evidence.

Our October diagnostics are a practical warning against treating a sensible procedure as an accomplished result. On October 9, baseline and checklist conditions each answered all 16 core decisions correctly and each added one unsupported date claim. The checklist used 38.4% more answer words. On October 12, a narrower guard again tied baseline: 8/8 development cases and 6/6 later handoffs each. Those were temporal-task diagnostics, not an AHE evaluation. They nevertheless demonstrate why this journal cannot assume that an extra instruction earns its overhead.

Where this sits beside earlier research

AHE did not invent verification, teaching, or calibrated uncertainty. Chain-of-Verification (2023) develops a staged verification procedure and reports reductions in hallucination on its evaluated tasks. Conformal Linguistic Calibration (February 2025) formalizes a tradeoff between factuality and specificity, including weakening claims. Can LLMs Learn by Teaching for Better Reasoning? A Preliminary Study (2024) studies learning-by-teaching methods involving student feedback. These are prior points of contact, with different mechanisms and evidence.

None validates this project’s teaching episode or supplies an implementation for its governor metaphor. Equally, the comparison does not erase the contribution of this particular synthesis: claim authority, interruption, useful replacement, self-description, and monitoring of the safeguard’s own failures. That is work worth publishing and testing. Establishing field-wide priority would require a separate historical review.

The test the design still owes

A prospective evaluation should compare ordinary answering with a fixed AHE intervention using matched, previously unseen cases and condition-masked scoring. Include cases with missing evidence, sufficient evidence, an authoritative correction to an earlier answer, and unchanged evidence wrapped in persuasive narrative. Record the model, prompts, source packets, retrieval, outputs, and review decisions.

Score unsupported assertions separately from unnecessary deferrals, correct supported answers, and justified revisions. Word count and delay measure costs. Rule adherence is a separate outcome: a system can follow a remembered instruction more faithfully and still make worse decisions. Repeated samples and independent replication would be needed before claiming a reliable effect.

For now, AHE is a substantial historical design with an unfinished empirical obligation. The governing idea remains useful: an answer earns authority from its support, and a safeguard earns authority from its effects. The next contribution is to make those effects inspectable.

Read alongside

Published 14 October 2026, America/Denver. This is an annotated historical synthesis, not a new experimental result or a replacement for the canonical framework. Source texts remain preserved in their existing editorial states; private correspondence and personal details are excluded. Follow-up link added 19 October 2026: entry 030 reports a ceiling-limited pilot; the historical synthesis above is unchanged.