corpora-builder Is Not augment-it Tooling

Short, sharp orientation for any agent that touches corpora-builder, augment-it, dididecks-ai, or memopop-ai. The layout of this tree invites a wrong inference, and the wrong inference leads to bad architectural suggestions.

Origin — inheritance, not overlap

Rule. corpora-builder was built inside augment-it and later split out into its own repo. Treat it as a product that was born there, not as tooling augment-it owns and not as a peer that happens to overlap.

Why it was there: both centered on the same use case — augmenting record sets for a client — and the first corpus corpora-builder needed was one augment-it already had. As it built out, the owner recognized it is useful as a standalone app and as one that integrates into the didi.sh venture-capital suite alongside dididecks-ai and memopop-ai. That is when it earned its own repo.

What follows, and this is the part agents get wrong:

• The corpus at augment-it/clients/<slug>/corpus/ is a legacy of origin, not a design decision. Nobody chose to put a corpus inside a record-enrichment app; it was already there when corpora-builder split out. Do not reason backwards from its location to a principle.
• They are currently disconnected in fact. As of 2026-08-22 the reach-edu corpus (892 files, git submodule lossless-group/augment-reach-edu) is edited through augment-it’s tooling, while corpora-builder has never written a byte to it or to its R2 bucket. If a session finds this confusing, that is the correct reaction — it is unfinished, not intentional.
• Moving a corpus toward corpora-builder is a handover, not a copy. Frame it that way in any plan.
• Never propose a shared package. Three consumers — augment-it (record sets), dididecks-ai (decks), memopop-ai (memos) — makes shared code tempting and it remains prohibited. Integration is a service with a verb contract and file formats (the source.* vocabulary), never an import across the three apps. See [[Preferred-Stack]] and the standing no-shared-dependency rule.

Purpose — no guesswork on factual claims

Rule. The corpus is a hard boundary, not a retrieval convenience. When generating anything factual for a client-facing artifact, claims come from the corpus or they do not get made.

Why, in the owner’s framing (2026-08-22):

Because of the rise of agents, ever more capable models at lower and lower token cost, we can assume that millions of people will bump into the fact that they are drafting documents of consequence and can’t risk hallucination and the inaccuracies that come from it. The only real solution for now is to tightly restrict drafting documents and decks to a “corpus” or “corpora” — no guesswork allowed for factual-level claims.

This is the thesis the whole product rests on, and it is why corpora-builder is foundational to the suite rather than a peer within it. dididecks-ai and memopop-ai both generate documents of consequence; under this rule the corpus is not a feature they integrate, it is the substrate they are only trustworthy on top of.

How to apply:

• Do not fill gaps from training data when authoring memo or deck content for a client. A missing fact is a capture task, not a paraphrase opportunity.
• Provenance is per claim, not per document. A corpus boundary constrains retrieval; it does not stop a model interpolating between retrieved facts, and a sentence routinely mixes a checkable claim with a judgment. Hex-code citations and the LFM citation conventions are the enforcement mechanism, not formatting.
• A weak corpus turns hallucination into confident citation of bad sources, which is harder to catch because it looks rigorous. That makes the quality scan load-bearing rather than hygiene — treat corpus quality as the product’s trustworthiness, and surface it rather than burying it.
• Say what is missing. “The corpus does not support this claim” is a correct and useful output. Filling the hole silently is the failure this whole system exists to prevent.

Related

• ../plans/Sync-Corpora-to-R2-and-Show-Clients-What-Changed.md — where the origin was recorded and what it changes
• ../../corpora-builder/context-v/explorations/Corpora-Builder-System-Design.md — the operator wishlist and the domain model
• ../explorations/A-Syncbox-For-Client-Document-Folders.md — the substrate and history reasoning around it
• ../../../context-v/reminders/Check-The-Substrate-Before-Reasoning-On-Top-Of-It.md — the tree-wide guardrail from the same session