Knowledge Tomography: Prompting Human and Agentic Entities to Reproduce Their State of Knowledge

Educational technology and agent evaluation share a measurement problem. Final answers, multiple-choice scores, and polished deliverables sample thin output surfaces. They often fail to distinguish fluent generation from a recoverable, transformable state of knowledge—especially under AI-assisted work where the surface answer may not originate in the evaluated mind.

We introduce knowledge tomography as a methodological frame: prompt human and agentic entities to try to reproduce their state of knowledge, capture the process of that reproduction, and treat the resulting artifacts as projections from which a fuller picture of configuration can be reconstructed. Tomography is not the end goal. It is the measurement substrate required for knowledge induction technology—systems that actively transform configuration toward useful proximity with calibrated effort.

This working paper is a methods and design document. Section 2 frames tomography against induction. Section 3 defines the construct. Section 4 states methodological principles for humans and agents. Section 5 outlines a planned study validating the Think Aloud Protocol (TAP) as an initial tomography tool. We do not claim completed empirical results.

Knowledge induction tech names the long-horizon goal: methods and systems that induce desired knowledge configurations—moving an entity through configuration space toward states where target material can be retrieved, applied, and transformed with high reliability. Induction is transformation under guidance: shorter paths, less wasted effort, preserved depth.

Induction without tomography is blind steering. If we cannot externalize and reconstruct what an entity currently knows—and does not know—we cannot choose interventions, estimate proximity, or evaluate whether a transformation succeeded beyond a single test item. Quizzes and final products remain useful outcomes, but they are insufficient control signals for closed-loop induction.

Knowledge tomography therefore sits upstream of induction product roadmaps. Software attention loops, Socratic facilitation, proof-of-work verification, and later world models or biofeedback layers all depend on richer state estimates than pass/fail. This paper fixes the measurement vocabulary so induction research can cite a named family of methods rather than ad hoc “ask them to explain.”

• Ultimate aim: knowledge induction—guided transformation of knowledge configuration.
• Prerequisite: tomographic measurement of current state via reproduce-your-knowledge prompts.
• Scope of this paper: define tomography; outline TAP as first tool; leave full induction systems to future work.

We define knowledge tomography as methodologies that prompt human as well as agentic entities to try to reproduce their state of knowledge, capturing multi-angle process evidence from which that state can be reconstructed and compared to a target configuration.

Three commitments distinguish the construct. First, the primary act is reproduction: the entity is asked to re-externalize what it holds—definitions, causal structure, procedures, examples, boundary cases, and transfer—not merely to select among options. Second, subjects include both humans and agentic systems; the same family of prompts and scoring ontology should apply so co-evaluation is possible. Third, each elicitation is a projection (a tomographic slice), incomplete alone; designed batteries of prompts yield reconstruction of configuration, gaps, and proximity.

The medical metaphor is deliberate but bounded. We do not claim neural imaging. We claim a measurement program: controlled projections → reconstruction of an unobservable internal configuration (operationalized as knowledge state) → uncertainty estimates and gap maps. Proof of Work (PoW) dual-stream traces, embeddings, and Map of Knowledge regions are engineering substrates that can carry tomographic slices; they are not synonyms for the methodology itself.

• Core verb: reproduce (re-externalize) the entity’s state of knowledge under prompt.
• Subjects: human learners and agentic entities under matched targets.
• Output: inspectable process artifacts (slices) for multi-angle reconstruction, not finals alone.
• Relation to induction: tomography measures; induction transforms.

A tomography protocol specifies a knowledge target, a prompt battery designed to force reproduction across facets of that target, capture rules for process artifacts, and a reconstruction / scoring procedure. Facets typically include definition, mechanism or procedure, worked example, counterexample or edge case, transfer, and self-reported uncertainty. Incomplete or contradictory reproduction is signal, not noise.

For human subjects, concurrent verbalization and selective thought interfaces (for example TAP Stash/Submit) preserve spontaneous crystallization alongside deliberate dialogue participation. For agentic entities, matched prompts request the same facets with explicit instructions to externalize intermediate reasoning and to withhold or revise under a dual-stream ontology where available. In both cases, polished final answers alone are insufficient tomography.

Construct validity requires that prompts target reproduction of held knowledge rather than open-book retrieval or tool-mediated generation that bypasses the entity under test. Environments may include tools; scoring focuses on what is externalized as the entity’s attempted reconstruction of its own state. Protocol purity, consent for humans, and disclosure of agent scaffolding are mandatory reporting items.

• Prompt batteries force multi-facet reproduction of a labeled knowledge target.
• Capture process: verbal/text traces, timing, revisions, stash vs submit where instrumented.
• Humans and agents share facet ontology and scoring rubrics for comparable slices.
• Reconstruction metrics: completeness, accuracy, gap localization, transfer, rater agreement.

We plan a validation study that treats the Think Aloud Protocol (TAP)—in particular the selective Stash/Submit interface documented in our companion methods paper—as an initial knowledge tomography tool. The study asks whether TAP-elicited dual-stream Proof of Work reconstructs human knowledge state more faithfully than final-answer-only baselines, and whether the same elicitation frame extends to agentic entities under matched targets. Results are not claimed here; the outline below fixes aims, design, and success criteria.

Reproduction is incomplete and can be reactive; prompting changes what is said. Tomographic reconstruction is a proxy for knowledge state, not a full neural readout. Agentic “knowledge” is tool- and context-dependent; we report scaffolding and do not equate weights with human understanding.

TAP validation will not exhaust the space of tomography methods (written protocols, structured interviews, interactive probes, multi-session longitudinal slices). Claiming TAP as an initial tool is not a claim of uniqueness or optimality.

Human subjects research requires informed consent, clear retention of stashed content, and separation of research corpora from production telemetry. Comparative agent studies must avoid overclaiming human-likeness. Induction systems built on tomographic estimates inherit measurement error; uncertainty should remain first-class in any closed loop.

Knowledge tomography is the family of methodologies that prompt human as well as agentic entities to try to reproduce their state of knowledge, yielding multi-angle process projections for reconstruction. It is the measurement counterpart to knowledge induction technology, which aims to transform configuration toward useful proximity with less wasted effort.

We position TAP Stash/Submit as an initial tomography tool and outline a planned validation study with concrete aims, conditions, reconstruction metrics, and an agentic pilot arm. Completing that study is future work; this paper fixes the definitional frame and experimental outline so measurement research can proceed with transparent constructs.