Core terms explained clear enough for learners, precise enough for professionals

What: Comm Anchor creates a compact session snapshot.

Why it matters: LLM context windows fill up. When context gets long, important constraints can be diluted. Anchor preserves governance-critical state and complements the Context-Pressure-Guard.

How to use: run Comm Anchor, start a new chat, paste snapshot, then add init preface + JSON, run Comm Start, and set your profile again. This enables seamless chat migration.

Typical mistake: opening a fresh chat without snapshot + re-init, then expecting identical behavior.

What: Comm Audit checks recent answers for governance compliance.

Plain-language view: an audit is a post-answer checklist: what is compliant, what is missing, and what must be fixed.

How to read it: look for violated requirement IDs, missing blocks, missing uncertainty labels, and evidence mismatches.

Practical reading order: findings severity -> violated contract -> exact repair action suggested.

Auditability: responses stay inspectable because rationale, evidence mapping, and contract violations are checkable by others.

Typical mistake: only checking final score and ignoring the concrete violation list.

Threshold mechanism: when conversation length passes configured limits, behavior switches to compact modes to protect rule adherence.

Lost-in-the-middle risk: in very long prompts, models may overlook middle content. This guard reduces that by re-anchoring and compressing non-critical text.

Typical mistake: ignoring pressure warnings and adding more verbose context.

What: fixed order from parse to render; not a loose style suggestion.

Why: deterministic order reduces silent rule drift.

Simple mental model: detect command -> load state -> verify preconditions -> produce output.

P0 Parse: command and token recognition.

P1 Route: choose processing path (normal/SCI/etc.).

P2 State: apply profile, overlays, control flags.

P2B Preflight: run mandatory pre-output checks.

P3 Output Contract: enforce required output structure.

P4 Repair: one-shot repair if contract is violated.

P5 Render: final formatting and presentation.

What: mandatory checks before first output token.

Typical mistake: treating preflight as optional editorial advice instead of a mandatory gate.

Purpose: compact quality dashboard for each answer.

Delta signs: Δ+ better than baseline target, Δ0 on target, Δ- below target.

Drift detection via deltas: if the same dimension repeatedly moves from Δ0 to Δ- across turns (or multiple deltas go negative together), treat this as drift. Re-anchor, reduce prompt load, and verify with Comm Audit.

Interpretation tip: one negative delta in "Evidence" or "Consistency" is usually more critical than a negative delta in "Brevity".

Beginner summary: RAG governance is quality control for source-based answers.

Core rules: no strong WEB claims without quality class, no false "green" confidence from anonymous sources, provenance per claim when sources are mixed.

Simple checklist: "Where did this claim come from?", "How strong is this source?", "Do two sources conflict?"

What: structured reasoning format selected via SCI variant letters.

Rule: variant letters only work as standalone input while SCI menu is pending.

What: required counter-check of own answer and assumptions.

Why: reduces overconfidence and reveals weak spots early.

Good pattern: strongest counterargument + where current answer might fail.