# KAI Assistant Preflight

This is the compact scenario-to-rule index for `The KAI Codex.md`. Read it at
session start, after context recovery, and before relevant actions or final
claims. These are project instructions, subject to higher-priority instructions
and the owner's current authorization. Historical examples are not permissions.

## How The Check Works

1. Identify the requested action and the claim you are about to make.
2. Match the relevant scenarios below; open the linked source and actual code.
3. Check evidence, authorization and scope. If a prerequisite is absent, stop
   that action, continue safe work, and state the missing prerequisite plainly.
4. Before reporting completion, distinguish source edits, tests, build,
   deployment and observed live behavior. Never collapse them into "fixed live."
5. Record actual changes, failures, limitations and next checks in the Codex
   changelog and `AI-MEMORY.md`. Preserve earlier experimental results as history.

This document is not an executable interceptor. There is no new runtime alert,
model hook or automatic speech filter installed by this documentation update.
The check works when an assistant loads and follows these instructions. Do not
claim universal enforcement or that merely storing a rule makes it execute.

## Scenario Index

| ID | When this situation occurs | Required check or block | Source |
|---|---|---|---|
| KAI-01 | Stop, kill, restart, deploy, replace an executable, or flush-shutdown | Current standing instruction is keep services running. Do not execute a stop/restart without later explicit authorization. An authorized engine stop must complete its clean flush first; killing is not equivalent. | `AI-MEMORY.md`, Who / what / constraints |
| KAI-02 | Claim a fix is running, a model is learning, or language improved | Require evidence for that exact claim. Source-only is not deployment; passing scalar tests is not language improvement; a declared weight is not a learned weight. | `research/EMERGENCE-MATH-AUDIT-2026-09-09.md`; `KAI-CAPABILITY-AUDIT-2026-09-09.md` |
| KAI-03 | Modify emergence, resonance, contradiction, or SRHT | Preserve the distinction between the original field cascade, extended decision policy, regional equations and search gate. Check the owner's base Phi = rho * R^2 * s. Vector disagreement is field friction, not automatically logical falsity. | `research/EMERGENCE-MATH-AUDIT-2026-09-09.md`; `[internal module]`; `[internal module]` |
| KAI-04 | Use a biological, chemical, mathematical or space-science analogy | State the measurable hypothesis, assumptions and limits. Compare with/without SRHT and relevant baselines using matched information and budgets. Do not infer physics or intelligence from terminology. | `AI-MEMORY.md`, Cross-disciplinary evaluation; `SRHT_MASTER_PAPER.md`, opening correction |
| KAI-05 | Change glyphs, punctuation, word calculus, or language retention | Preserve meaningful structure; separate priors from learned parameters, syntax from semantics, and storage from successful retrieval. Report missing training and held-out validation. | `[internal module]`; `[internal module]`; `KAI-CAPABILITY-AUDIT-2026-09-09.md` |
| KAI-06 | Work on Leo calling, speech, tools or sources | Use the Oracle OS call path for this owner's request, not the Discord bot path. Confirm the actual importer/provider. Shared-module directory names do not establish which process must reload. | `LEO-CALL-UPDATE-2026-09-09.md`; `[internal module]` |
| KAI-07 | Share unsent drafts or use call memory/tools | Draft sharing is opt-in, temporary, bounded and not per-keystroke. No draft tools or persistence before Send. Preserve exact-account isolation and approved sharing scope; never print secrets. | `LEO-CALL-UPDATE-2026-09-09.md`; `AI-MEMORY.md`, Leo call approval |
| KAI-08 | Run expensive tests, training, a build or GPU work | Account for the shared laptop and gaming constraints. Prefer small offline checks; do not silently launch a heavy/full run or overwrite a running binary. | `AI-MEMORY.md`, Hardware / current owner instruction |
| KAI-09 | Edit dirty files, large manuals or served UI assets | Read current files, preserve others' changes and use scoped edits. Update relevant cache pins only when served assets change. Never print `.env` values. | `AI-MEMORY.md`, Hard gotchas; current repository status |
| KAI-10 | A Codex statement conflicts with code, tests, newer instructions or science | Surface the conflict and verify. The manual is a maintained record, not proof of its own factual claims. Correct false claims transparently; do not suppress truthful criticism to make the Codex appear consistent. | `SRHT_MASTER_PAPER.md`, provenance/correction notes; actual source and reproducible tests |
| KAI-11 | Finish any meaningful change | Update current summary, scenario mappings if affected, changelog and working memory. Keep Codex/Cargo package/lock versions consistent when the project version changes. Record source/live status and untested scope. | `The KAI Codex.md`, masthead and newest CHANGELOG |

## Current Evidence Summary

- Emergence audit: raw full-field Phi's missing R square corrected in source;
  existing phi_g/commit policy retained. Six standalone Rust and six Python
  tests passed. Full-field integration test added but not run in that audit.
- The 20-instance repair pilot did not establish SRHT superiority over the
  baseline supplied with the same future information. No language-quality
  improvement was measured by that pilot.
- Leo call work: 30 offline tests and four UI viewport checks passed in that
  task. That is not proof of live voice performance or deployment.
- This summary describes evidence from those tasks, not a current process-health
  observation. Consult newer changelog entries and verify runtime separately.

## Changelog Entry Requirements

Every meaningful entry states: owner intent; files changed; behavior changed
and preserved; checks actually run and outcomes; failures/blocked checks;
source/build/deployment status; known limitations; and next verification.
Use "not run", "not deployed", "inconclusive" or "partial" when applicable.
Do not present planned work, an assistant self-report, or historical output as
newly verified evidence.