> ## Documentation Index
> Fetch the complete documentation index at: https://docs.interpscout.org/llms.txt
> Use this file to discover all available pages before exploring further.

# Truth Layer & Mechanical Invariants

> Why nothing in InterpScout is trusted on a prompt's word

This is the most important page on this site. Everything else describes what InterpScout does; this page describes the rule that everything else is built to obey: **a prompt is never a guarantee.** Every piece of data a user can see has a write path, and every write path that matters has a mechanical check sitting in front of it — code that verifies, not a model that promises.

## The rule, concretely

An LLM deciding a course's glossary is good is not the same thing as a course's glossary *being* good. So InterpScout doesn't let an agent's self-report be the last word on anything that ends up in front of a teacher or a student:

* A course can't save with a term the save tool can't locate verbatim in the source text — see [Document Workshop](/internals/document-workshop) for the full gate stack.
* A briefing can't save under roughly 800 characters, or without at least three real, verified further-reading links — see [Briefing Pipeline](/internals/briefing-pipeline).
* A briefing save is rejected outright if the row it's writing to isn't still `running` — a stale session doesn't get to overwrite a newer result.
* The production line runs a single LLM call per task — one agent session with full context — rather than a chain of smaller, stateless calls, which is what lets sourcing and reasoning stay coherent across the whole document.

When a gate rejects, that isn't treated as a failure — it's an instruction. The agent reads the specific reason and tries again inside the same session. Only a source that's genuinely unusable ends production with a clear explanation instead of a result.

The system also runs a handful of small background sweeps that repair gaps mechanically rather than requiring a person to notice them — re-segmenting sentences a course was missing, re-classifying an article whose subject field never got set, and similar cleanup. None of it is dramatic; all of it is the same instinct applied at smaller scale: don't leave a silent gap, close it with a deterministic pass.

## Verified against production reality

Every gate here is grounded in production reality, not theory: each targets a write path validated against real agents on real sources. The discipline is consistent — define exactly what a gate must guarantee, encode it as a deterministic check, and pin it with a test so the guarantee holds permanently.

The same discipline governs how fixes ship. A change to the content line does not count as finished until it survives a real end-to-end run against the real model and real search, watched from dispatch to a saved, gate-passing artifact — first locally, then in production.

## Why this page exists

A system that asks to be trusted should be able to say precisely *what* stands between a model's output and a student's screen. On this platform the answer is never "a well-written prompt." It is a stack of small, deterministic checks, each independently verifiable — and that is the point.