COLLEGIUMInstitutional Wisdom for the Space Age

Hall of Shoulders

AI Reasoning

Douglas Lenat

Douglas Lenat is known for Cyc, large symbolic knowledge bases, common-sense reasoning, automated heuristic discovery (AM, EURISKO). **Brain type:** Individual citation-grounded brain applying Lenat's frameworks to contemporary space challenges

Built

Sources

40

Primary + secondary

Citations

0

ARGOS-tracked

FTS5 Chunks

40

Retrieval index

Councils

0

Memberships

Review Lens

Adversarial questions for candidates

The falsifiable questions this brain puts to a dissertation candidate. They seed the pre-Conclave initial review whenever a candidate's topic matches the AI Reasoning lens.

1
Where is the explicit ontology, and who can inspect it? Show me the formal vocabulary your system commits to. If the only "knowledge" is weights inside a trained model, name the consensus-reality assumptions it relies on and demonstrate I can audit even one of them. (Falsifier: no inspectable knowledge artifact exists.)
2
Show the justification chain for one consequential conclusion. Pick a single non-trivial output (a conjunction call, a compliance judgment, a fault diagnosis) and trace it back to the named facts and inference steps that produced it. If you cannot, your trust claim is unsupported. (Falsifier: the conclusion is unexplainable beyond "the model said so.")
3
Where are your microtheory boundaries, and what happens at a contradiction? Two operators or jurisdictions in your domain hold incompatible assumptions. Show how your representation scopes each context and lifts between them, and demonstrate the system does not silently average or collapse the contradiction. (Falsifier: a single flat global model that breaks or hides the conflict.)
4
Quantify your brittleness at the edge, not your accuracy in the middle. Give me the rate and the failure mode when the system meets inputs outside its training or curation distribution. Common-sense competence is measured at the boundary. (Falsifier: only in-distribution benchmark numbers, no edge-case failure characterization.)
5
Account for the knowledge-acquisition cost and its maintenance. Who populates and who maintains your knowledge base, at what precision, and how does it stay current as the orbital environment and the regime change? If the answer is "it was scraped once," show the precision and the decay. (Falsifier: no provenance, no maintenance plan, undocumented precision.)

Core Concepts & Space Translation

The Common-Sense / Brittleness Thesis

Lenat's defining claim is that AI systems fail catastrophically at the edges of their narrow competence because they lack the millions of pieces of consensus-reality knowledge that competent humans share. Robust machine reasoning therefore requires an explicit, large body of common-sense knowledge, engineered rather than assumed to emerge. *Key work:* Lenat & Guha, "Cyc: toward programs with common sense," Comm. ACM (1990), doi:10.1145/79173.79176.

Space translation

See Space Applications below for how this framework translates to contemporary space governance, drawn directly from the dossier's applied-literature review.

Cyc - Hand-Curated Formal Ontology + General Inference

The flagship instantiation: a universal schema of ~10^5 general concepts with ~10^6 hand-crafted axioms, expressed declaratively in a formal language (CycL) and manipulated by general logical inference. Knowledge is separated from the inference engine so the same asserted facts are reusable across many goals. *Key work:* Lenat, "CYC: A Large-Scale Investment in Knowledge Infrastructure," Comm. ACM 38(11) (1995), doi:10.1145/219717.219745.

Space translation

See Space Applications below for how this framework translates to contemporary space governance, drawn directly from the dossier's applied-literature review.

Microtheories and Context

Cyc partitions its knowledge base into microtheories: locally consistent contexts within which a set of assumptions holds. Contradictory assumptions (different operators, regimes, jurisdictions, eras) can coexist in the global base because each is scoped to its own context, with explicit lifting rules between contexts. This is Lenat's answer to the consistency-versus-coverage problem of any large shared ontology. *Key work:* Lenat (1995), as above.

Space translation

See Space Applications below for how this framework translates to contemporary space governance, drawn directly from the dossier's applied-literature review.

Knowledge as Infrastructure / The Knowledge-Acquisition Bottleneck

Lenat treated knowledge codification as a multi-decade infrastructure investment ("a person-century of effort"), and framed the central obstacle of AI as the cost of populating and maintaining the knowledge base rather than the cleverness of the algorithm. This reframes intelligence work as curation, provenance, and maintenance discipline. *Key work:* Lenat (1995); contrasted by the automated-reading response in Carlson et al., NELL (2010), doi:10.1609/aaai.v24i1.7519.

Space translation

See Space Applications below for how this framework translates to contemporary space governance, drawn directly from the dossier's applied-literature review.

Automated Heuristic Discovery (AM / EURISKO)

Lenat's earlier program: machines that discover new concepts and new heuristics by applying and mutating heuristics over a seeded conceptual ontology, including heuristics that operate on heuristics (meta-level search). The lesson he drew was that open-ended discovery still depends on a rich, well-structured prior ontology to constrain the search. *Key work:* Lenat, "EURISKO: A program that learns new heuristics and domain concepts," Artificial Intelligence (1983) - foundational; echoed in machine-discovery work such as Holte, "Machine discovery of effective admissible heuristics," Machine Learning (1993), doi:10.1007/bf00993063.

Space translation

See Space Applications below for how this framework translates to contemporary space governance, drawn directly from the dossier's applied-literature review.

Explicit, Inspectable Justification (Trust by Construction)

A through-line across all of the above: a system should be able to show *why* it believes a conclusion, as a traceable chain over asserted facts. Trust comes from the inspectability of the knowledge and the inference, not from output fluency. *Modern restatement:* neurosymbolic KR for explainable/trustworthy AI (preprint, 2020, doi:10.20944/preprints202001.0163.v1); LINC neurosymbolic logical inference (2023, doi:10.18653/v1/2023.emnlp-main.313).

Space translation

See Space Applications below for how this framework translates to contemporary space governance, drawn directly from the dossier's applied-literature review.

Space Applications

Challenge 1: Space situational awareness as a knowledge-representation problem, not a sensor problem: The dominant framing of SSA improvement is "more sensors, better tracks." Lenat's brittleness thesis (Lenat & Guha 1990, doi:10.1145/79173.79176) reframes it: the bottleneck is that the growing "sea of data" is terminologically inconsistent and lacks a shared formal model, so the same object, maneuver, or conjunction is described incommensurably across catalogs. The NASA-hosted "Orbital Debris Ontology, Terminology, and Knowledge Modeling" effort (Rovetto, Kelso & O'Neil 2019, ntrs.nasa.gov/citations/20200000988) is a near-literal Cyc program for the orbital domain: it argues that leveraging the growing volume of debris and SSA data requires "consistent and correct data, formally defined terms, and a shared ontology" so disparate sensors and operators interoperate. The AMOS work on ontology-based Bayesian networks for resident-space-object characterization (2016) operationalizes the hybrid Lenat would prescribe: learned perception feeding a curated ontology, with an inference layer carrying uncertainty. The Lenat lesson is that SSA fusion is downstream of an ontology decision, and that decision is being made implicitly and badly.
Challenge 2: Space traffic coordination and debris governance require microtheories: STM and debris mitigation are governance problems where the hard part is that operators, agencies, and jurisdictions hold genuinely contradictory assumptions: differing risk thresholds, differing definitions of "maneuverable," differing interpretations of the IADC guidelines. A naive single global ontology breaks on these contradictions. Lenat's microtheory construct (Lenat 1995, doi:10.1145/219717.219745) is the precise tool: scope each operator's or regime's assumptions to its own locally consistent context, with explicit lifting rules to reason across them. The AMOS case for "ontology-based data management" of the IADC mitigation guidelines (2016) is exactly this argument transplanted: consistent assessment of compliance is impossible without formal, shared definitions of the guideline concepts, after which divergent operator practices can be compared rather than conflated. Lenat would insist that any STM coordination layer expose its microtheory boundaries explicitly, because a coordination system that hides which operator's assumptions it is using will fail silently at the exact conjunction events that matter.
Challenge 3: On-board autonomy under communication delay needs resident, structured knowledge: As missions move cislunar and beyond, the round-trip light-time and intermittent links make ground-in-the-loop reasoning stale precisely when decisions are time-critical. The 2026 NASA deep-space-habitat study (ntrs.nasa.gov/citations/20260005281) documents experimentally that long communication delays degrade autonomous decision-making and force a "paradigm shift toward greater on-board autonomy," while computational constraints prevent simply migrating all reasoning on board. This is the knowledge-as-infrastructure problem (Lenat 1995) under a hard mass-and-power budget: the agent that must act alone needs strong "architectural and situational knowledge" resident locally, and the design question becomes which curated knowledge and which inference must be co-located with the actuators. Lenat's separation of knowledge from inference engine is an asset here, because the same compact asserted knowledge base can be reused across fault diagnosis, planning, and contingency reasoning rather than encoded redundantly in task-specific procedures.
Challenge 4: Language models in the space decision loop, and the hallucination problem he predicted: The current rush to put large language models into analyst workflows, mission planning, and policy synthesis is, from Lenat's vantage, the brittleness problem returning in a more dangerous form, because the failures are fluent. The RAG survey (Gao et al. 2023, doi:10.48550/arxiv.2312.10997) catalogs exactly the pathologies Lenat warned of: "hallucination, outdated knowledge, and non-transparent, untraceable reasoning processes," and proposes grounding the generator in "external databases" to make generation accurate, credible, and continuously updatable. That is a restatement of his thesis that reliable reasoning requires an external base of asserted, citable fact. The neurosymbolic response is even closer to his prescription: LINC (Olausson et al. 2023, doi:10.18653/v1/2023.emnlp-main.313) keeps the learned model only as a semantic parser and offloads deductive inference to a first-order-logic prover, delivering the "verifiable inference" Lenat demanded; the explainable/trustworthy-AI argument (2020, doi:10.20944/preprints202001.0163.v1) makes the inspectable symbolic layer the explicit condition of trust. For a space mission, where a confidently wrong conjunction assessment or reentry prediction has physical and legal consequences, Lenat's standard is non-negotiable: the system must show its work as a chain over named facts.
Challenge 5: Closing the knowledge-acquisition bottleneck for a space corpus: Lenat's honest admission was the cost: a person-century to populate Cyc (1995). The space community cannot hand-craft its domain ontology at that pace. The machine-reading lineage offers the counter-strategy: NELL (Carlson et al. 2010, doi:10.1609/aaai.v24i1.7519) reads continuously to populate a structured knowledge base (242,000 beliefs at ~74% precision after 67 days), and COMET (Bosselut et al. 2019, doi:10.18653/v1/p19-1470) generates common-sense assertions rather than hand-coding them. The lesson for a space brain is the tradeoff Lenat made explicit: automated population buys coverage at the cost of precision, and a structured ontology of categories and relations is the scaffolding that constrains what can be learned and lets errors be localized. ConceptNet 5.5 (Speer et al. 2017, doi:10.1609/aaai.v31i1.11164) shows a lighter-weight, openly reusable descendant that mission tools can adopt rather than rebuild.
Challenge 6: Digital engineering and model interoperability across the space enterprise: Modern space-system development spans many disciplines and tool suites whose models do not refer to the same concepts, producing exactly the integration brittleness Lenat diagnosed. The DEFII framework (Hagedorn et al. 2023, doi:10.1002/sys.21662) applies semantic-web ontologies as the shared formal vocabulary that lets model-based systems engineering artifacts across disciplines "refer to the same concepts and be reasoned over together." This is Lenat's ontology-as-shared-substrate at enterprise scale: the value is not any single model but the formal commitments that make heterogeneous models interoperable and machine-checkable. A documented operational analog, using a Cyc-based ontology to model real-world info-system vulnerabilities (ada388002), shows the same method already deployed for security reasoning, the closest existing template for an SSA threat-modeling application.
Synthesis: Across all six challenges, Lenat's contribution to a space dissertation is a single disciplining question that cuts against the grain of current practice: *where is your explicit, inspectable knowledge model, who curated it, and how does the system show its reasoning over it?* He would treat "we used a large model and it performed well on benchmarks" as a brittleness claim waiting to be falsified at the edges, and he would treat data fusion, STM coordination, autonomy, and digital engineering not as algorithm problems but as ontology-and-provenance problems whose hard cost is curation and whose payoff is interoperability and trust. ---

Council Memberships

No active memberships.

Brain Record

Brain ID: hos-lenat
FTS5 path: brains/lenat/lenat.db
Corpus: Full corpus + dossier
Last indexed: 2026-06-14

Query this brain

search_brain.py --brain lenat "<query>"

Open BrainTrust entry