Corpus Forge

Inspiration

A folder of documents is not yet a governed body of knowledge.

Modern AI systems can summarize, retrieve, and synthesize information, but they often blur important distinctions: what the source actually says, what someone interpreted it to mean, what a team currently accepts, what remains disputed, and who authorized a change.

Corpus Forge was inspired by the need for a more accountable kind of AI-assisted knowledge system: one that can evolve without silently rewriting its own history.

What Corpus Forge Does

Corpus Forge is a provenance-first workflow for maintaining a governed, living corpus.

It preserves:

source artifacts and fragments;
evidence, interpretation, claims, applications, and unresolved questions as distinct object types;
provenance and transformation lineage;
lifecycle states such as pending, provisional, accepted, rejected, contested, or unresolved;
explicit human-review decisions;
structured export for downstream software or AI systems.

The central rule is simple:

AI may propose. Humans authorize.

In the demo, new material enters an existing corpus. GPT-5.6 compares it with the current knowledge objects and proposes a successor. The proposal remains visibly unauthorized until a human reviewer accepts, rejects, quarantines, or leaves it unresolved. The previous object is preserved rather than overwritten, and the review decision is appended to the corpus history.

How I Built It

I used a small public-domain corpus based on Benjamin Jowett’s translation of Plato’s Euthyphro. The dialogue works well for this demonstration because definitions are repeatedly proposed, challenged, revised, and left unresolved.

I first prepared a sanitized corpus fixture containing:

six source fragments;
five distinct knowledge-object types;
provenance links;
an incoming fragment;
an AI-generated proposal;
predecessor and successor relationships;
review metadata and lifecycle states.

I then used GPT-5.6 to help define the product workflow, corpus rules, validation criteria, and demo narrative. Codex implemented the working vertical slice as a Vite, React, and TypeScript application.

The final application supports the complete demonstration path:

inspect the existing corpus;
introduce an incoming fragment;
view an unauthorized AI proposal;
record a human disposition;
preserve predecessor and successor lineage;
append a review event;
export the governed state as JSON;
reset the demonstration.

The project was deliberately kept small: no backend, authentication, database, or live model dependency was required for the MVP.

Challenges

The hardest challenge was not coding the interface. It was deciding what the system must refuse to collapse.

A conventional AI workflow tends to optimize for producing a cleaner answer. Corpus Forge instead had to preserve ambiguity, disagreement, historical states, and authority boundaries.

That required careful separation between:

source evidence and interpretation;
AI synthesis and authorized knowledge;
revision and erasure;
useful automation and human judgment.

Another challenge was scope. The underlying architecture could support research teams, standards bodies, institutional memory, policy maintenance, or long-lived personal knowledge systems. For Build Week, I reduced that larger vision to one visible, testable move.

What I Learned

I learned that provenance alone is not enough. A trustworthy living corpus also needs lifecycle state, review authority, unresolved tensions, and preserved lineage.

I also learned that AI can be most valuable when it is not treated as the final authority. GPT-5.6 was effective at classification, comparison, synthesis, and proposing changes. The architecture became stronger when authorization remained explicitly human.

Finally, I learned that a small vertical slice can demonstrate a large architectural idea when the boundaries are visible.

Why It Matters

A research team could use this pattern to maintain an evolving evidence brief. A standards group could preserve contested interpretations. An organization could give AI systems institutional memory without allowing a model to silently rewrite institutional judgment.

Corpus Forge does not attempt to make AI the authority.

It makes AI useful inside a system where provenance, uncertainty, change, and human responsibility remain visible.

Chatbots generate answers. Corpus Forge maintains what a team is authorized to rely upon—and why.