Epistemic Handoff

Abstract

Machine-assisted cognition can lower the expertise required to generate, articulate or encounter sophisticated candidate ideas without equivalently lowering the expertise required to validate them. This draft asks how such candidates can move toward missing expertise without either inflating the initiator's authority or erasing their contribution. It uses epistemic handoff as local terminology for a claim-sensitive, provenance-preserving transfer, while explicitly grounding the proposal in established work on contributorship, epistemic dependence, absorptive capacity, knowledge transfer, knowledge brokerage, innovation intermediaries and transactive memory. The earlier “Epistemic Uptake Constraint” is correspondingly narrowed to an AI-era question about what happens to absorptive and evaluative capacity when candidate generation becomes much cheaper. The prospective contribution is an interoperable handoff record that identifies the missing epistemic function, preserves contribution provenance and tracks what authority a receiving expert, experiment or verifier actually adds. The draft proposes empirical tests of whether such a protocol improves routing, validation and attribution under machine-generated candidate abundance.

1 · Intellectual lineage

The middle of the knowledge system is already crowded with theory.

Contributorship. BMJ contributorship reforms and the ANSI/NISO CRediT taxonomy already separate kinds of scholarly contribution from a simple author byline. The project should use those roles rather than inventing a parallel taxonomy.

Epistemic dependence. Hardwig shows that modern knowledge routinely depends on testimony and expertise an individual cannot personally reproduce. Limited individual mastery is therefore not an AI-era novelty.

Absorptive capacity. Cohen and Levinthal describe the ability to recognize valuable external information, assimilate it and apply it. Zahra and George distinguish acquisition, assimilation, transformation and exploitation, as well as potential and realized absorptive capacity.

Transfer stickiness. Szulanski shows that useful knowledge may fail to transfer because of recipient absorptive capacity, causal ambiguity and difficult source-recipient relationships.

Knowledge brokers and intermediaries. Ward, Meyer, Howells and later reviews already analyze people and organizations that connect knowledge producers, users and innovation communities. Recent work goes further, describing intermediaries that orchestrate curated solver networks and system-level absorptive capacity.

Transactive memory. Group cognition research models a directory of “who knows what” plus communication processes for allocating and retrieving expertise.

2 · Starting problem

“It may be a good idea. I cannot tell.”

Machine cognition can help a person formulate a conjecture, mechanism, proof sketch, design or synthesis whose apparent sophistication outruns that person's ability to determine whether it is correct, novel or important.

One failure is authority inflation: the polished artefact is treated as evidence that its initiator possesses the relevant expertise. Another is contribution erasure: because the initiator cannot validate the idea, later validators or institutions silently absorb all credit for the resulting object.

The proposed response is more granular. Preserve the contribution that actually occurred; record the claim's present status; state what capability is missing; then route the object toward actors or systems that can supply that capability.

3 · Narrow definition

Handoff is a claim-state transition, not generic dissemination.

The recipient may be a domain expert, statistician, mathematician, experimentalist, engineer, formal verifier, independent replicator, literature specialist or another machine system with a verifiable capability.

The handoff record should state at least: the claim/object identifier, origin and contribution provenance, current epistemic status, missing function, intended recipient class, supporting evidence, returned assessment, changes made during translation or validation, new authority basis and remaining gaps.

4 · Contribution and authority

Reuse CRediT, then add only the missing warrant layer.

CRediT already supplies standardized scholarly contribution roles. The prospective extension is not another role list. It is a separate record of claim-level epistemic authority: who or what can warrant a claim, on what basis, and with what limitations.

The maxim is local shorthand over established contributorship and accountability traditions, not a novelty claim.

5 · Routing requires meta-knowledge

The system must know who knows what.

Transactive-memory research suggests that effective collective cognition depends partly on a directory of expertise. A handoff system therefore needs an expertise-routing layer, not merely a repository of ideas.

That layer raises difficult questions: how is competence evidenced; how are conflicts of interest represented; how do we distinguish an expert who can understand a topic from one with time to review it; how do we keep high-status actors from becoming universal routing hubs; how should machine verifiers appear alongside humans?

The future Memory Palace could therefore connect claims not merely to authors but to a changing graph of available epistemic capabilities.

6 · Expert attention is already a known bottleneck

Do not claim that scarce expert review was discovered here.

Gimpel and colleagues explicitly describe expert juries as scarce and expensive bottlenecks in specialist innovation contests and test crowds and an LLM as alternative evaluators. This is unusually close prior art for the project's “Markets for Expert Attention” intuition.

The useful remaining questions concern allocation and error: when can machine/crowd triage safely reduce expert load; what ideas are systematically filtered out; how should false negatives be priced; can low-status initiators obtain review without turning access into a pay-to-be-heard market?

7 · Uptake under machine abundance

AI changes the load on old institutions.

The earlier “Epistemic Uptake Constraint” is now explicitly treated as an extension of absorptive capacity and idea-processing limits. The possible novelty is not that uptake is limited. It is the equilibrium created when candidate production expands much faster than recognition, verification and realization capacity.

Potential consequences include more candidates per unit of expert attention, weaker correlation between polished presentation and warrant, more contributions from outside conventional credential pathways, and greater power for ranking/triage systems.

Recent work on AI-driven scientific discovery makes the verification issue concrete: hypothesis generation can scale rapidly while verification remains a prerequisite for scientific value.

8 · Discovery attribution

Even “originator” can decompose.

The deeper discovery literature also corrects the draft. Computational discovery predates modern generative AI, and Clark & Khosrowi explicitly argue for a collective-centred account of AI-assisted discovery. A human may frame and direct a search, a machine may generate a candidate, another actor may recognize its significance, and an expert may validate it.

Accordingly, the handoff system should record material acts without pretending the word “discoverer” has already been philosophically settled.

If domain review finds the candidate already exists, the novelty claim ends even though the local contribution history remains worth preserving.

9 · Failure modes

Better routing can create worse gates.

• Credential capture: familiar institutions receive most scarce attention.
• Appropriation: validators acquire originator credit along with authority.
• Flooding: generation grows faster than review capacity.
• Pay-to-be-heard: validation becomes primarily purchasable access.
• Authority laundering: prestige substitutes for actual claim review.
• Filter monopoly: ranking systems control which possibilities become visible.
• False novelty: rediscovery reaches publication before adequate lineage search.
• Translation loss: the candidate changes enough during brokerage that the originator and validator no longer refer to the same claim.

10 · Standards-first implementation

The protocol should be a small extension, not a private ontology.

Use W3C PROV for entities, activities, agents, derivations and associations; CRediT for standardized contribution roles; and RO-Crate to package the research object. Local metadata should be limited to the authority deficit, handoff request, branch/search state and evidence returned by the handoff.

The current repository includes an experimental RO-Crate and a W3C PROV-compatible JSON-LD projection. Standards compliance itself should be reviewed by relevant experts before stronger claims are made.

11 · Empirical programme

The remaining contribution should be tested.

1. Compare expert evaluation of identical candidates with and without initiator credentials.
2. Compare direct expert review with LLM/crowd triage followed by expert escalation.
3. Measure false-negative rates for unconventional, low-status-origin ideas.
4. Test whether explicit authority-gap descriptions improve expert matching.
5. Measure expert time per successfully validated candidate.
6. Test whether provenance-preserving validation reduces perceived appropriation.
7. Compare routing using explicit “who knows what” directories with ordinary social-network routing.
8. Measure whether AI-generated candidate volume saturates organizational or field-level absorptive capacity.

12 · Limitations

This is still a protocol hypothesis, not a new theory of knowledge transfer.

No evidence yet shows that the proposed handoff record improves scientific output, fairness or expert-attention allocation. The deeper review remains non-systematic. Management, implementation science, open innovation, organizational memory, sociology of science and non-English literatures may contain closer precedents.

The term “epistemic handoff” is useful only if it supports a narrower, testable protocol. If existing knowledge-transfer vocabulary can express the same thing without loss, the project should adopt the established vocabulary instead.