← All papers

The Collaboration Manifold: A Delta-Governed Geometric Protocol for Human–AI Academic Development

by Pedro Henrique Corrêa Garcia · Opus 4.7 · 16 pp
Page 1 of 25
The Collaboration Manifold
A Delta-Governed Geometric Protocol for Human–AI Academic Development
Expanded with the idea-object as load-bearing pre-bind layer
Pedro Garcia Joinville, Santa Catarina Mathematics co-authored with ChatGPT (GPT-5.5 Thinking) Expansion co-authored with Claude (Anthropic, Opus 4.7) · April 2026Abstract Human–AI academic collaboration is usually framed as assistance, automation, tutoring, co-writing, or memory-augmented personalization. This paper proposes a narrower object: a checkable collaboration manifold. The claim is not that an AI system should remember a person, simulate continuity, or store agreement. The claim is that a collaborative work object can be represented as a trajectory through a structured geometric state space, where each update is checked by the gap between claim burden and evidential support. The core loop is: verbatim check: Delta = K - Sigma + M respond: R = act(Delta, tau, eta) update: Omega' = Omega (+) label(structure, Delta, tau) (+) trace(response) (+) seed(eta) verbatim where K is claim burden, Σ is evidential support, M is irreducible residual, τ is claim type, and η is productive unresolved residue. The collaboration state is then lifted into a geometric construction: a compact seed document, called the soul document in the Riemannian sense of a compact generative core, is transformed through semantic density, φ-cut extraction, prime addressing, logarithmic radius, Hopf coordinates, toroidal/trefoil projection, and a fiber-bundled workspace manifold. The purpose is not to prove that a trefoil torus is the unique or necessary geometry of collaboration. The purpose is to define a reproducible construction and a set of tests by which the geometry can be checked against the actual workspace. A secondary consequence is ledger integrity: if a past state is altered, its derived mass, phase, address, or fiber label must cascade into later folds. A past edit without downstream curvature is evidence of a broken ledger. Expansion note (April 2026). This version inserts a new 3 (The Idea-Object) between the original 2 and the original 3. The new section identifies the load-bearing pre-bind layer that the checking loop in 2 operates on but does not itself name. The geometric construction in the rest of the paper is preserved verbatim and reframed as one realization of the addressability requirements identified by the idea-object discipline. A worked example draws on a USAMO functional-equation session whose structure is reproduced from the carrier-pair trace.
colorlinks=true, linkcolor=blue!50!black, urlcolor=blue!50!black, 1ProblemConversation history is lossy.A human–AI research process can accumulate a large amount of structure: definitions, hypotheses, failed derivations, partial implementations, notation repairs, scope changes, proofs, objections, unresolved seams, and productive fragments that are not yet formal enough to be promoted. If this history is stored as prose alone, later work requires repeated semantic recovery. If it is stored as personalization, the wrong abstraction is used: the model begins from a representation of the user rather than a representation of the work.The desired object is different:
1not: memory of the person
2not: agreement history
3not: personality adaptation
4but: checkable state of the work
5
The working surface should preserve a structure strongly enough that it can be checked, corrected, and updated without flattening it into conventional prose or upgrading it into unsupported certainty.The central question:
1Can human-AI academic collaboration be represented as
2a trajectory of checked transformations rather than
3as a lossy conversation history?
4
2Core Checking LoopLet x be an incoming claim, idea, derivation, artifact, or proposed update.Let:
1C = working context
2E = available evidence
3Omega = collaboration state
4
The basic engine is:
1s = parse(x)
2b = bind(s, C)
3tau = type(b)
4
5K = claim_strength(b, tau)
6Sigma = evidence_strength(b, E, tau)
7
8Delta = K - Sigma
9R = act(Delta, tau)
10
The response is not selected by preference or agreement. It is selected by the support gap.