Retrieval Is Not Enough: Why Organizational AI Needs Epistemic Infrastructure Federico Bottino1 ,
Carlo Ferrero1 ,
Nicholas Dosio1 ,
Pierfrancesco Beneventano2
1
arXiv:2604.11759v1 [cs.AI] 13 Apr 2026
2
Kakashi Ventures Accelerator (KVA) Massachusetts Institute of Technology
Abstract Organizational knowledge used by AI agents typically lacks epistemic structure: retrieval systems surface semantically relevant content without distinguishing binding decisions from abandoned hypotheses, contested claims from settled ones, or known facts from unresolved questions. We argue that the ceiling on organizational AI is not retrieval fidelity but epistemic fidelity—the system’s ability to represent commitment strength, contradiction status, and organizational ignorance as computable properties. We present OIDA, a framework that structures organizational knowledge as typed Knowledge Objects carrying epistemic class, importance scores with class-specific decay, and signed contradiction edges. The Knowledge Gravity Engine maintains scores deterministically with proved convergence guarantees (sufficient condition: max degree < 7; empirically robust to degree 43). OIDA introduces QUESTION-as-modeled-ignorance: a primitive with inverse decay that surfaces what an organization does not know with increasing urgency—a mechanism absent from all surveyed systems. We describe the Epistemic Quality Score (EQS), a reusable five-component evaluation methodology with explicit circularity analysis. In a controlled comparison (n=10 response pairs), OIDA’s RAG condition (Minerva, 3,868 tokens) achieves an EQS of 0.530 vs. 0.848 for a full-context baseline (Cowork, 108,687 tokens); the 28.1× token budget difference is the primary confound for the composite gap. The QUESTION mechanism is statistically validated: Minerva produces explicit ignorance declarations in 10/10 responses vs. 5/10 for Cowork (Fisher p=0.0325, OR= 21.0, Haldane–Anscombe), an architectural invariant rather than a sampling artifact. The formal properties are established; the decisive empirical test—a tag-and-boost ablation at equal token budget (E4)—is pre-registered and not yet run. We present a concrete evaluation agenda alongside an honest accounting of what the system guarantees and what it does not.
1
Introduction
An AI agent is asked to summarize an organization’s position on a contested strategic decision. It retrieves five relevant documents. All five are semantically relevant. Three support the original hypothesis; one contradicts it with recent market data; one is an unresolved question about regulatory risk that no one has answered in six weeks. The agent treats all five as equivalent evidence and produces a confident summary. The summary is fluent, well-organized, and epistemically incoherent. This failure is not a retrieval failure. It is an epistemic one. The system found the right documents; it had no way to interpret their epistemic status. Nothing in the retrieval substrate distinguished a binding decision from an abandoned hypothesis, a contested claim from a settled one, or an open question from a resolved finding. The antagonist. The dominant response to epistemic failures in organizational AI is to improve retrieval: better embeddings, denser indexes, reranking, hybrid search, longer context 1
windows, structured entity graphs. These are legitimate improvements within their scope. But they do not resolve the underlying problem on an epistemically flat substrate. No amount of retrieval improvement will surface contradiction status that is not encoded, distinguish commitment strength that is not represented, or model organizational ignorance that is not tracked. The field is pushing on the wrong wall. Three capabilities remain absent from production systems. First, commitment differentiation: no surveyed system distinguishes a verified decision from a tentative hypothesis at retrieval time with different importance dynamics. Second, contradiction propagation: existing approaches either detect contradictions at query time from text (unreliably) or repair inconsistencies post-hoc—none encode contradictions as signed edges that dynamically suppress importance scores. Third, ignorance modeling: no system represents what an organization does not know as a first-class object whose operational cost increases over time. Our approach. OIDA1 addresses this gap by structuring knowledge epistemically at ingestion, maintaining that structure deterministically, and exposing it to agents as a firstclass retrieval property. Classification is LLM-assisted and therefore fallible; all subsequent maintenance—decay, scoring, contradiction propagation, memory zone allocation—is deterministic and auditable. Contributions. 1. A Knowledge Object model with nine epistemic classes, signed typed relationships, and a deterministic importance engine with proved convergence guarantees. 2. QUESTION-as-modeled-ignorance: inverse decay for unresolved questions, operationalizing accumulated decision risk under uncertainty—a mechanism absent from all surveyed systems. 3. Signed contradiction propagation applied to epistemic infrastructure: an established signed-GNN mechanism applied to organizational epistemic contradiction suppression, with partial suppression (CONTRADICTS = −0.6) implementing epistemological tolerance. 4. The Epistemic Quality Score (EQS): a reusable evaluation methodology with five grounded sub-scores, explicit circularity analysis, and a statistical protocol for small-N comparisons. 5. A controlled pilot comparison (n=10 response pairs) with full confound treatment, plus three formal design properties distinguishing KGE from static epistemic labeling.
2
Related Work
Knowledge graphs and ontological systems. Knowledge graphs represent information as networks of entities and typed relationships [4]. Enterprise deployments—notably Palantir’s Ontology [30]—map operational reality into semantically coherent graphs. Temporal knowledge graphs (TKGs) [5, 17] extend this with validity intervals and time-aware reasoning. However, TKGs answer “when was this true?”—they do not answer “how much should you trust this now, given its type?” No TKG applies different decay rates based on the kind of knowledge a fact represents. Structured retrieval-augmented generation. RAG [20] is the dominant paradigm for grounding LLM outputs [12]. GraphRAG [8] uses graph-based community summaries to improve retrieval over flat chunks. LightRAG [14] incorporates entity-relationship graphs into 1
From Greek oida (“I know because I have seen”)—a knowing that arises from experience rather than declaration.
2
dual-level retrieval. Both structure entities and relations but not epistemic commitment: a decision and a hypothesis with the same entity mentions receive the same retrieval treatment. Li et al. [23] and Li et al. [22] characterize the efficiency frontier between RAG and long-context LLMs—a tension OIDA’s retrieval design directly engages. Leng et al. [19] document a reliability drop above 64K input tokens; Shi et al. [32] show that irrelevant context degrades LLM reasoning—both motivations for K-score filtering. Agent memory systems. MemGPT [29] introduced hierarchical memory paging for agents. Zep/Graphiti [31] provides temporal tracking and bi-temporal modeling via graph-based storage— the closest system to OIDA in capability, but it computes importance implicitly through centrality and recency rather than through typed epistemic dynamics. Mem0 [6] extracts and consolidates facts dynamically but does not type them epistemically. MemOS [25] provides a memory operating system for skill and tool memory. A-MEM [36] implements agentic selforganizing memory. A recent survey [24] confirms that no production agent memory system separates LLM-assisted ingestion from deterministic post-ingestion maintenance, and none model organizational ignorance. Commercial organizational AI. Glean [13], Notion AI [28], and Microsoft 365 Copilot [7] represent the current generation of production organizational AI systems—each treats retrieved knowledge uniformly without epistemic typing, class-specific decay, or ignorance modeling. Active retrieval and uncertainty. FLARE [18] retrieves new documents when the model detects low-confidence generation tokens—the closest mechanistic relative to OIDA’s QUESTION primitive. The difference is fundamental: FLARE signals parametric model uncertainty; QUESTION models organizational ignorance encoded at knowledge ingestion. The two are orthogonal and potentially composable. Organizational knowledge management. Walsh and Ungson [34] identified organizational memory as distributed across individuals, culture, structures, and processes. Stein and Zwass [33] formalized organizational memory information systems. Nonaka [27] theorized knowledge creation through socialization, externalization, combination, and internalization. Du et al. [11] survey the implications of generative AI for knowledge management. These frameworks identify the need for epistemic structure; OIDA provides a computational implementation. Trust, provenance, and contradiction in KBs. PROV-O [26] standardizes provenance metadata but does not compute epistemic dynamics. Uncertainty management surveys [16] and confidence propagation methods [21] address quality estimation but not class-specific decay or signed contradiction propagation. Temporal confidence decay [15] models validity intervals but applies uniform aging across knowledge types. Knowledge conflict surveys [35, 10] address detection and repair but do not encode contradictions as persistent negative-gravity edges in a scoring engine. Signed graph neural networks and epistemic contradiction. Signed GNNs extend graph neural network propagation to graphs with both positive and negative edges, learning node representations that respect the sign of relational ties [9]. Approaches including SGCN and SNEA establish the propagation mechanism—signed edge weights, balance-theory-aware aggregation, and sign-flipping through multi-hop paths—as a mature technique in the signed network literature. OIDA does not claim this mechanism as a novel contribution. What is novel is the application of signed propagation to epistemic contradiction handling in organizational knowledge bases: the CONTRADICTS edge (coefficient −0.6) dynamically suppresses
3
Table 1: Positioning: OIDA vs. existing systems across epistemic capability dimensions. ✓ = present, ∼ = partial, — = absent. System GraphRAG LightRAG Zep/Graphiti MemGPT/Letta Mem0 A-MEM MemOS TKGs OIDA
Epist. Typing
Class Decay
Contr. Prop.
Det. Maint.
Ignor. Model
— — — — — — — — ✓
— — — — — — — — ✓
— — — — — — — — ✓
— — ∼ — — — — ✓ ✓
— — — — — — — — ✓
the importance score of contradicted Knowledge Objects within the Knowledge Gravity Engine, surfacing unresolved organizational contradictions as first-class signals for downstream AI agents. The partial suppression coefficient (−0.6, not −1.0) implements an epistemological tolerance principle: contradicted knowledge may remain retrievable, reflecting the organizational reality that contradictions are often unresolved coexistences rather than logical defeats. To our knowledge, no existing knowledge graph inconsistency management approach applies signed edge weights for dynamic importance suppression within a retrieval engine [10]; standard approaches use static flagging, version pinning, or inconsistency-aware querying without score propagation. Cognitive architectures. OIDA’s usage force inherits structure from ACT-R’s base-level activation [2, 3], replacing the global decay parameter with class-specific rates and extending spreading activation to signed edges. The connection is structural, not a cognitive science contribution. Table 1 summarizes the landscape. No production system makes epistemic quality computable across all five dimensions. The gap is not an oversight—it reflects a field assumption that better retrieval is sufficient. OIDA tests the alternative hypothesis: that epistemic structure at the retrieval substrate level is the missing intervention.
3
The OIDA Framework
OIDA consists of three components: the Knowledge Object model, the Knowledge Gravity Engine, and the hybrid retrieval architecture. Figure 1 shows the lifecycle. Four design principles guide the architecture: Design Principles. 1. Classify at ingestion; maintain deterministically thereafter. LLM-assisted classification is fallible; all post-ingestion computation is reproducible and auditable. 2. Model what you do not know. QUESTION with inverse decay represents organizational ignorance as a first-class object whose cost increases over time. 3. Contradictions are computable signals, not text to be discovered. Signed edges encode contradiction as persistent negative gravity. 4. Class-specific decay > global decay. A decision and an observation should not age at the same rate.
4
Ingestion
typed KOs
(LLM-assisted)
KGE Cycle
K-scores
(deterministic)
Hybrid Retrieval ranked set Agent Context (struct+sem+topo)
(ranked KOs)
usage signals
Figure 1: OIDA system lifecycle. Ingestion is LLM-assisted; all subsequent maintenance and retrieval is deterministic. Table 2: Epistemic class taxonomy. Seed values and half-lives are working heuristics, not validated optima. Seed K
Decay
Half-life
DECISION
1.00
None
∞
CONSTRAINT
0.90
None
∞
EVIDENCE
0.80
Exp.
∼365 d
NARRATIVE PLAN
0.70 0.65
None Exp.
∞ ∼69 d
EVALUATION
0.55
Exp.
∼198 d
OBSERVATION
0.40
Exp.
∼90 d
HYPOTHESIS QUESTION
0.30 0.30
Exp. Inverse
∼50 d Urgency grows
Class
3.1
Epistemic Role Binding choice—valid until superseded Non-negotiable structural boundary Verifiable supporting/refuting data Persistent contextual anchor Structured intention with time horizon Informed qualitative assessment Weak signal not yet interpreted Unverified testable claim Open question requiring resolution
Knowledge Objects and the Epistemic Taxonomy
A Knowledge Object (KO) is a tuple: KOi = (idi , koci , classi , contenti , scoresi , edgesi , metai ) where classi ∈ C is drawn from a closed taxonomy of nine epistemic classes and edgesi is the set of typed directed relationships. Scores comprise a five-dimensional vector (K, conf, fresh, urg, contr). The Knowledge Object Coordinate (KOC) is a 7-axis immutable identifier providing O(1) structural similarity (specification in Appendix A). The nine classes arise from crossing two orthogonal axes that together determine computational behavior. Axis 1: Epistemic commitment strength. Propositions range from explicit ignorance (a question no one has answered) through uninterpreted signals (observations), provisionally held claims (hypotheses, plans), evidentially supported assessments (evidence, evaluations), persistent contextual anchors (narratives), up to verified and binding commitments (decisions, constraints). This ordering determines seed importance. Axis 2: Temporal behavior under absence of reinforcement. A DECISION requires an explicit SUPERSEDES event for deactivation. An OBSERVATION loses weight if unreinforced. A QUESTION gains urgency—unresolved uncertainty becomes more costly over time. This axis determines the decay profile: non-decaying, exponentially decaying, or inversely decaying. All 36 class pairs are distinguishable by at least two features (decay type, seed value, halflife, or semantic role), establishing operational adequacy: no pair can be merged without losing observable behavioral differences in the KGE (exhaustive enumeration in Appendix E). 5
0.58 ∗ KQ = 0.556
K-score
0.54
Divergence from day 1 (Theorem 1)
QUESTION (λ=−0.010) OBSERVATION (λ=0.015)
0.5
EVIDENCE (λ=0.005) HYPOTHESIS (λ=0.008)
0.46
DECISION (λ=0.002) ∗ KO = 0.435
0.42 0
7
14 Days
21
28
Figure 2: K-score dynamics over 28 days under stationary inputs (K0 = 0.5, η = 0.1, ∆t = 1). QUESTION KOs (—) diverge upward from day 1 while OBSERVATION KOs (- - ) decay — a direct consequence of λquestion < 0 (Theorem 1). This simulation exhibits the theorem’s prediction; deployment validation requires production telemetry. QUESTION as modeled ignorance. QUESTION is the only class with inverse decay: unresolved questions become more urgent, not less. We frame urgency as operationalizing accumulated organizational decision risk under unresolved uncertainty, following a Value of Information (VoI) interpretation: each day a QUESTION remains unresolved, the organization makes decisions in its shadow, accumulating risk. Urgency provides a monotonically nondecreasing lower bound on retrieval priority—a rising floor guarantee that modeled ignorance eventually outranks stale observations. Under stationary inputs from identical initial conditions (K0 = 0.5), KQUESTION and KOBSERVATION diverge from the very first update (t= 1): the inverse decay ensures QUESTION KOs accumulate urgency immediately. By day 28, the gap ∗ ∗ is 0.115 score units (KQUESTION = 0.556 vs. KOBSERVATION = 0.435), with an asymptotic gap of 0.121 (Theorem 1 below). This growing separation—where what is unknown outranks what is stale—is the visual thesis of the framework. When a QUESTION is resolved (typically by a DECISION linked via IMPLEMENTS), its urgency drops to zero. No surveyed system has an equivalent primitive. Theorem 1 (T2: QUESTION Divergence from Day 1). Under stationary inputs (u = e = g = c = 0), K0 = 0.5, η = 0.1, ∆t = 1: ∗ Kquestion =
η · seed 0.1 × 0.5 = = 0.556 > K0 η + λquestion · ∆t 0.1 + (−0.010) × 1
∗ Kobservation =
η · seed η + λobservation · ∆t
=
0.1 × 0.5 = 0.435 < K0 0.1 + 0.015 × 1
∗ ∗ Since Kquestion > K0 > Kobservation , the two trajectories diverge at t = 1 (the first discrete
update). The asymptotic gap is 0.121; at day 28 the simulated gap is 0.115 score units. Proof and full derivation: Appendix D. Typed relationships. Relationships between KOs are drawn from a closed vocabulary of ten directed edge types, each carrying a signed coefficient. Key types include SUPPORTS (+1.0), BASED_ON (+0.8), IMPLEMENTS (+0.7), SUPERSEDES (+0.6), BLOCKS (−0.4), and CONTRADICTS (−0.6). Positive edges propagate importance; negative edges actively suppress it through the gravity computation. Negative-coefficient edges implement signed propagation— a technique established in the signed GNN literature [9]—applied here to organizational epistemic contradiction suppression. The full vocabulary is specified in Appendix B. 6
Signed contradiction and epistemological tolerance. We chose CONTRADICTS = −0.6 (not −1.0) deliberately: contradicted knowledge is suppressed, not erased. This reflects the design requirement that contradicted knowledge should remain partially retrievable in case the contradiction itself is wrong. Concretely, mild contradiction (one CONTRADICTS edge) suppresses K by approximately 22%; strong contradiction (two edges) by approximately 67%. Approximately two SUPPORTS edges are needed to counter one CONTRADICTS edge. We term this epistemological tolerance.
3.2
Knowledge Gravity Engine
The KGE computes an updated importance score K for every active KO at each cycle (default: every 6 hours):
K(t+1) = clamp (1 − η) · K(t) + η · [seed + u + e + g] − λclass · ∆t · K(t) − c, 0, 1
(1)
The equation decomposes into three forces. Momentum (1−η)K(t) carries forward current importance. Injection η[seed + u + e + g] introduces new signals: seed is the class baseline, usage force u is retrieval-driven activation via an exponential recency kernel adapted from ACTR [2, 3], evidence force e counts new inbound SUPPORTS edges, and gravity force g propagates importance through signed edges from connected KOs. Negative forces −λclass ∆t · K(t) − c apply class-specific decay and contradiction penalty. Force formulas are detailed in Appendix C. Under stationary inputs (before clamping), the per-node update converges to a unique fixed point: η · [seed + u + e + g] − c K∗ = (2) η + λclass · ∆t Coupled convergence. For the full coupled system, we prove a sufficient condition via the Gershgorin circle theorem: convergence is guaranteed when max _degree < gscale / max |COEFF| ≈ 7.14 (Appendix D). The per-node contraction factor is uniformly in [0.845, 0.850] across all classes, leaving a coupling budget of approximately 0.15 per node. Empirically, convergence is observed for all tested configurations—including graphs with max degree 43 (six times beyond the sufficient condition)—due to tanh saturation bounding gravity output to (−1, 1), clamping providing non-expansive projection, and mixed-sign edges partially canceling. The gap between the analytical bound and empirical robustness is itself an informative finding. Memory zones. KOs are allocated to four zones by K-score: Core Memory (K ≥ 0.40, always in agent context), Working Memory (0.10 ≤ K < 0.40, retrieved when relevant), Peripheral (0.05 ≤ K < 0.10, targeted queries only), and Dormant (K < 0.05, excluded from gravity computation). No KO is ever deleted—only excluded from active computation. Design properties vs. tag-and-boost. Table 3 formalizes three design guarantees that distinguish KGE from a static epistemic labeling baseline (tag-and-boost: same nine classes, BM25+cosine hybrid retrieval, recency weighting only, no dynamic scoring). These properties are proved at the design level; empirical validation requires the E4 ablation (pre-registered, not yet run).
3.3
Hybrid Retrieval
The hybrid score combines three independent similarity layers2 : H(q, i) = α · Sstruct (q, i) + β · Ssem (q, i) + γ · Stopo (q, i)
(3)
2 The production deployment extends this three-component formulation to a nine-component Vesta Score that additionally incorporates epistemic confidence, freshness (decay state), graph mass (centrality), goal alignment, domain match, functional fit, and historical stability, each with configurable weights summing to 1.0 and query-
7
Table 3: Three formal design properties distinguishing OIDA KGE from a tag-and-boost baseline (static epistemic labels + recency weighting). Whether these properties produce measurable EQS improvements at equal token budget is the subject of planned ablation E4. Property
OIDA KGE
Tag-and-Boost (static labels)
P1: Fixed-point convergence
Converges to unique K ∗ (contraction factor ∈ [0.845, 0.850]) encoding class, usage history, evidence, and graph neighborhood.
No dynamics; importance is static after labeling. No fixed point encoding class-specific equilibrium.
P2: QUESTION rising floor
λQUESTION = −0.010 guarantees ∗ KQUESTION = 0.556 > K0 . Unresolved ignorance cannot be permanently deprioritized.
No mechanism differentiates QUESTION from OBSERVATION temporally. An aging QUESTION is treated identically to an aging OBSERVATION.
P3: Contradiction suppression
CONTRADICTS (−0.6) propagates negative gravity each cycle; contradicted KO converges to lower K ∗ . Contradiction is a structured, evolving signal.
A contradiction tag does not affect retrieval priority of the contradicted KO. No dynamic propagation.
Structural similarity Sstruct : computed from KOC axis alignment in O(1), no database access. Semantic similarity Ssem : cosine similarity over embedding vectors, rescaled to [0, 1]. Topological similarity Stopo : inverse hop distance in the epistemic graph. Configured defaults: α = 0.30, β = 0.50, γ = 0.20. The final ranking multiplies hybrid similarity by contextual importance: R(q, i) = H(q, i) · Keff (i, q)
(4)
where Keff (i, q) = Kglobal (i) · max(0.10, ϕctx (i, q)). The floor of 0.10 prevents complete collapse of globally important KOs. The contextual attention function ϕctx modulates global importance by query-local relevance: ϕctx (i, q) = we · ϕentity (i, q) + wd · ϕdomain (i, q) + wa · ϕanchor (i, q)
(5)
where ϕentity (i, q) = 1[KOCentity (i) = q.primaryEntity] tests entity alignment, ϕdomain (i, q) = 1[KOCdomain (i) = q.domain] tests domain alignment, and ϕanchor (i, q) = |anchors(i)∩q.activeAnchors| / max(|q measures overlap with the query’s active contextual anchors. Configured defaults: we = 0.40, wd = 0.35, wa = 0.25. We chose hand-crafted weights over learned parameters because the corpus is small (500 KOs), determinism and auditability are design requirements, and understanding system behavior currently matters more than marginal optimization.
4
Evaluation
4.1
The Epistemic Quality Score
We propose the Epistemic Quality Score (EQS) as a reusable evaluation framework for any system claiming to provide epistemic structure for organizational AI. The composite metric comprises five sub-scores: EQS = 0.20 · ECA + 0.25 · CP + 0.20 · CR + 0.20 · EC + 0.15 · DE type presets. We present the reduced formulation here for clarity; the Vesta Score will be specified in a forthcoming companion paper on the full OIDA architecture.
8
ECA (Epistemic Classification Accuracy, 0.20): Does the response correctly distinguish epistemic types? CP (Contextual Precision, 0.25): Is the response grounded in evidence, not hallucinated? The highest weight reflects the primacy of faithfulness. CR (Contextual Recall, 0.20): Does the response cover relevant context comprehensively? EC (Epistemic Coherence, 0.20): Does the response handle contradictions and epistemic tensions appropriately? DE (Decision Enablement, 0.15): Does the response enable informed organizational decisions? Each sub-score uses 4-point calibrated anchors (0.1, 0.4, 0.7, 1.0) with behavioral descriptions. The statistical protocol uses paired Wilcoxon signed-rank tests for small-N comparisons, McNemar tests for binary classification, and Cohen’s d as the primary effect size measure. Circularity analysis. We acknowledge that 60% of composite weight (CP + CR + DE) is system-independent. However, 20% (ECA) has high circularity—it tests whether responses respect OIDA’s own taxonomy. The remaining 20% (EC) has moderate circularity. An independent evaluation by domain experts assessing decision quality without reference to OIDA’s framework would be needed to validate the design choices themselves. As an LLM-as-judge method, EQS is subject to known biases [37]. We recommend cross-validation with human expert evaluation for deployment decisions.
4.2
Comparative Results: Minerva vs. Cowork
Experimental setup. We compare two conditions on the ClearPath corpus (≈500 KOs): Minerva (OIDA RAG, structured retrieval, top-k KOs) and Cowork (full-context baseline, entire corpus in context). We ran n = 10 parallel LLM response pairs on the query: “What are the main bottlenecks identified in ClearPath’s current operational processes?” (model: claudesonnet-4-6). Token budget and primary confound. Condition A (Minerva) used 3,868 input tokens; Condition B (Cowork) used 108,687—a 28.1× difference. This token budget differential is the primary confound for all EQS sub-score comparisons and must be held in view throughout. Minerva retrieves ≈5–7 KOs; Cowork accesses the full corpus. QUESTION declaration rate (unconfounded finding). Minerva declares its ignorance; Cowork happens to know. In all 10 runs, Minerva includes an explicit Knowledge Gap paragraph— a design invariant, not a sampling artifact (Fisher p = 0.0325). The parallel Cowork runs surface the same limitation in only 5 of 10 responses—incidentally, because the right text happened to be in context. Table 4: QUESTION KOs with λ < 0 are structurally privileged in top-k retrieval (Theorem 1), producing an explicit “Knowledge Gap” section in every Minerva response as a design invariant. Cowork surfaces the same underlying knowledge gap in 5/10 responses incidentally. Metric Explicit ignorance declarations Fisher’s exact p (two-tailed) Odds ratio (Haldane–Anscombe)
Minerva (OIDA RAG)
Cowork (Full-Context)
10/10 (100%) 0.0325 21.0
5/10 (50%) — —
The zero non-declaration cell in Minerva is a design invariant: λQUESTION < 0 guaran∗ tees KQUESTION > K0 from day 1, making QUESTION KOs structurally privileged in top-k retrieval. This is the paper’s cleanest architectural result—it has a proved mechanism, a statistical test, and no token-budget confound.
9
Citation Precision: hallucination-free in both conditions. Both Minerva (CP = 0.660) and Cowork (CP = 0.895) are hallucination-free across all 10 runs. Structural guarantees matter even when both conditions produce accurate text—the ECA and EC gaps reflect epistemic architecture, not error prevention. CP is the most confound-resistant sub-score; ∆ = 0.235 reflects grounding depth, not error rate. EQS sub-score comparison. Table 5 presents the full comparison. The dominant driver of the composite gap is Contextual Recall (∆ = 0.540), reflecting retrieval breadth: Minerva retrieves ≈5–7 KOs; Cowork ingests the full corpus. No architectural conclusion may be drawn from CR without token-equalized comparison (E2, planned). ECA (∆ = 0.245) and EC (∆ = 0.275) test epistemic structure; their attribution to architecture vs. breadth requires the E4 ablation. Table 5: EQS sub-score comparison (n = 10 response pairs, mean ± SD). Cowork uses 28.1× more input tokens than Minerva. CR (Contextual Recall) is predominantly breadth-mediated. CP is hallucination-free in both conditions. Cohen’s d values are inflated by near-zero withincondition variance due to architectural determinism. Wilcoxon signed-rank (one-sided): W = 55.0, p = 0.000977. Sub-score
Wt
Minerva
Cowork
∆
Note
ECA CP CR EC DE
0.20 0.25 0.20 0.20 0.15
0.575 ± 0.026 0.660 ± 0.021 0.340 ± 0.039 0.540 ± 0.032 0.520 ± 0.042
0.820 ± 0.042 0.895 ± 0.016 0.880 ± 0.026 0.815 ± 0.034 0.810 ± 0.039
0.245 0.235 0.540 0.275 0.290
— Hallucination-free (both) Breadth-mediated — —
EQS
—
0.530 ± 0.025
0.848 ± 0.017
0.318
Wilcoxon p=0.000977
Token input
—
3,868
108,687
28.1×
Primary confound
Both conditions are hallucination-free (CP ≥ 0.66). Citation Precision, which is not breadth-mediated.
The cleanest architectural signal is
H1 status. Whether KGE’s dynamic machinery adds value beyond static epistemic labeling at equal token budget remains an open empirical question (E4, pre-registered falsification condition). If a tag-and-boost baseline achieves ECA and EC scores within one standard error of full OIDA at equal token budget (≈3,868 input tokens), the KGE is not justified at the current deployment scale. The three design properties in Table 3 establish architectural distinction at the formal level; E4 tests whether this distinction is empirically observable.
4.3
Deployment Observations
OIDA is deployed as the operational knowledge infrastructure of a venture studio. Approximately 500 KOs span five ventures, three client engagements, and internal strategy, maintained over four weeks of KGE cycles with sources from Notion, Google Calendar, and Slack. These observations constitute design validation, not controlled evaluation. K-score distribution. 10–15% of KOs settle in Core Memory (K ≥ 0.40). The remaining distribution concentrates in Working Memory with a long tail of Peripheral KOs—consistent with the expectation that a minority of organizational knowledge is operationally central at any time. K-score trajectories. Simulation of class-specific dynamics over 28 days exhibits differentiated behavior: DECISION remains stable near 1.0, EVIDENCE decays slowly, OB10
Table 6: What the system guarantees and what it does not. The system guarantees
The system does not guarantee
Deterministic maintenance: same class, same parameters, same trajectory every time
Universal parameter validity: configured heuristics may underperform in organizations with different knowledge patterns
Typed epistemic structure: every KO carries class, scores, and signed edges
Perfect ingestion classification: typing quality depends on LLM-assisted classification
Explicit contradiction surfacing via negative-gravity edges that are computationally visible
Complete contradiction detection: the system models contradictions explicitly created, not those implicit in text
Coupled convergence for max degree < 7; empirically robust to degree 43
Tight convergence bound: the sufficient condition is 6× conservative; the tightest bound is an open problem
Stable retrieval contract independent of foundation model
Optimal retrieval quality: hybrid weights are design priors, not empirically optimized
Immutable audit trail: no KO is deleted, all state changes are logged
Calibrated absolute scores: K-values are relative rankings, not probability estimates
SERVATION decays toward 0.435, and QUESTION with medium urgency rises toward 0.556. Crucially, KQUESTION and KOBSERVATION diverge from the very first update (t= 1), confirm∗ ing Theorem 1. By day 28, the simulated gap is 0.115 score units (KQUESTION = 0.556 vs. ∗ KOBSERVATION = 0.435, Figure 2). These results are simulation-based under stationary inputs; production K-score evolution requires E6 (planned). Calibration finding. The 90-day half-life for OBSERVATION proved too long for AIadjacent domains where market signals evolve rapidly. For fast-moving domains, halving the default (to 45 days) is recommended. This was the most informative deployment finding.
5
Limitations and Evaluation Agenda
Table 6 is the paper’s most important table. We state what is established and what is not. What is not established. E4 (tag-and-boost ablation at equal token budget) is the decisive test for H1 and has not been run. E2 (token-equalized Minerva at ≈20K tokens) is planned but not executed. The efficiency frontier is a two-point extrapolation. E6 (production K-score trajectory logging, 28 days) is undesigned. The QUESTION mechanism has no dedicated EQS sub-score. Deployment is at a single site with hand-tuned parameters. Confounds. The Minerva vs. Cowork comparison confounds epistemic typing with retrieval selection effects and context size differences (28.1×). The ECA evaluator must not privilege OIDA-specific vocabulary in the E4 blind evaluation. Single evaluator model (claude-sonnet4-6); Cohen’s d inflated by architectural determinism (neff ≈ 1 for retrieval). Single corpus (ClearPath), single query. Decisive falsification commitment. If E4 finds that a tag-and-boost baseline achieves ECA and EC scores within one standard error of full OIDA (SEECA ≈ 0.0083, SEEC ≈ 0.010) 11
at equal token budget (≈3,868 input tokens), then KGE’s dynamic machinery is not justified at the current deployment scale. The contribution would reduce to “epistemic labeling is useful” and the dynamic scoring machinery would be a research artifact. Evaluation agenda. • E2: Token-equalized Minerva (≈20K tokens) to isolate epistemic structure from retrieval breadth • E4: Tag-and-boost ablation at 3,868 tokens (pre-registered; primary H1 test) • E5: Contradiction-detection query to isolate P3 empirically • E6: Production K-score trajectory logging over 28 days • Per-class classification accuracy audit (target: F1 > 0.7) • QUESTION urgency validation by domain experts (5–10 QUESTION KOs)
6
Conclusion
This paper names a specific antagonist—the assumption that better retrieval solves organizational AI—and tests an alternative: that epistemic structure at the retrieval substrate level is the missing intervention. OIDA provides a formal framework with convergence guarantees, an ignorance-modeling primitive, and signed contradiction propagation applied to organizational epistemic infrastructure. A pilot comparison (n = 10 response pairs) documents an EQS gap (Minerva 0.530 vs. Cowork 0.848) dominated by retrieval breadth at a 28.1× token budget difference, and statistically validates the QUESTION declaration mechanism (Fisher p = 0.0325). The EQS provides a reusable methodology for measuring whether epistemic structure matters. The guarantees table states what is established; the evaluation agenda specifies how to test what is not. Four design principles transfer beyond OIDA: classify at ingestion and maintain deterministically; model what you do not know; encode contradictions as computable signals; apply class-specific decay rather than global decay. A practitioner building organizational AI infrastructure can adopt these principles regardless of whether OIDA’s specific implementation survives empirical testing. The first two principles (P1: classify at ingestion; P2: QUESTION with rising urgency) require only LLM-assisted metadata tagging on any existing knowledge base—no graph infrastructure needed. The latter two (signed contradiction propagation and class-specific decay) require the full KGE; their empirical value over static labeling is precisely the question E4 will answer. The field has spent five years improving how AI finds organizational knowledge. It may be time to improve what organizational knowledge is—before the agent reads it. Acknowledgements. This work was developed within the research infrastructure of PoggioAI. We partially used PoggioAI/MSc for this manuscript [1]. We thank Alberto Trivero and Tommaso Portaluri for discussion on AI, statistical, and informatics matters.
References [1] Mahmoud Abdelmoneum, Pierfrancesco Beneventano, and Tomaso Poggio. PoggioAI/MSc: ML theory research with humans on the loop. Technical Report Technical Report v0, MIT, 2026. [2] John R. Anderson, Daniel Bothell, Michael D. Byrne, et al. An integrated theory of the mind. Psychological Review, 111(4):1036–1060, 2004.
12
[3] John R. Anderson and Lael J. Schooler. Reflections of the environment in memory. Psychological Science, 2(6):396–408, 1991. [4] Tim Berners-Lee, James Hendler, and Ora Lassila. The semantic web. Scientific American, 284(5):34–43, 2001. [5] Bingnan Cai, Yongqiang Xiang, et al. A survey on temporal knowledge graph: Representation learning and applications. arXiv preprint arXiv:2403.04782, 2024. [6] Prateek Chhikara, Deshraj Khant, et al. Mem0: Building production-ready AI agents with scalable long-term memory. arXiv preprint arXiv:2504.19413, 2025. [7] Eleanor Wiske Dillon et al. arXiv:2504.11443, 2025.
Early impacts of M365 Copilot.
arXiv preprint
[8] Darren Edge, Ha Trinh, Newman Cheng, et al. From local to global: A graph RAG approach to query-focused summarization. arXiv preprint arXiv:2404.16130, 2024. [9] others. Signed graph representation learning: A survey. arXiv preprint arXiv:2402.15980, 2024. [10] others. Dealing with inconsistency for reasoning over knowledge graphs: A survey. arXiv preprint arXiv:2502.19023, 2025. [11] others. Knowledge management in a world of generative AI: Impact and implications. ACM Transactions on Management Information Systems, 2025. Verify author names against published ACM version before submission. [12] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Qianyu Guo, Meng Wang, and Haofen Wang. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2024. [13] Glean Technologies. Glean: AI-powered enterprise search and knowledge discovery. https://www.glean.com/resources/guides/ glean-ai-enterprise-search-knowledge-discovery, 2024. Product documentation. [14] Zirui Guo, Lianghao Shi, Zhen Wang, et al. LightRAG: Simple and fast retrieval-augmented generation. arXiv preprint arXiv:2410.05779, 2024. Accepted at EMNLP 2025. [15] Rikui Huang, Wei Wei, Xiaoye Qu, Shengzhe Zhang, Dangyang Chen, and Yu Cheng. Confidence is not timeless: Modeling temporal validity for rule-based temporal knowledge graph forecasting. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10783–10794, 2024. [16] Lucas Jarnac, Yoan Chabot, and Miguel Couceiro. Uncertainty management in the construction of knowledge graphs: a survey. Transactions on Graph Data and Knowledge (TGDK), 3(1), 2024. [17] Yishi Jiang et al. A survey on temporal knowledge graph embedding: Models and applications. Knowledge-Based Systems, 304, 2024. [18] Zhengbao Jiang, Frank F. Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi-Yu, Yiming Yang, Jamie Callan, and Graham Neubig. Active retrieval augmented generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7969–7992, 2023. arXiv:2305.06983. 13
[19] Quinn Leng, Jacob Portes, Sam Havens, Matei Zaharia, and Michael Carbin. Long context RAG performance of large language models. In NeurIPS 2024 Workshop on Adaptive Foundation Models, 2024. arXiv:2411.03538. [20] Patrick Lewis, Ethan Perez, Aleksandra Piktus, et al. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems, volume 33, 2020. [21] Junheng Li et al. Continuous knowledge graph refinement with confidence propagation. IEEE Transactions on Knowledge and Data Engineering, 2023. [22] Xinze Li, Yixin Cao, Yubo Ma, and Aixin Sun. Long context vs. RAG for LLMs: An evaluation and revisits. arXiv preprint arXiv:2501.01880, 2025. [23] Zhuowan Li, Cheng Li, Mingyang Zhang, Qiaozhu Mei, and Michael Bendersky. Retrieval augmented generation or long-context LLMs? a comprehensive study and hybrid approach. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (Industry Track), 2024. arXiv:2407.16833. [24] Shichun Liu et al. Memory in the age of AI agents: arXiv:2512.13564, 2025.
A survey.
arXiv preprint
[25] MemTensor. MemOS: An operating system for memory-augmented generation. arXiv preprint arXiv:2505.22101, 2025. [26] Luc Moreau, Paolo Missier, et al. PROV-DM: The PROV data model. https://www.w3. org/TR/prov-dm/, 2013. W3C Recommendation. [27] Ikujiro Nonaka. A dynamic theory of organizational knowledge creation. Organization Science, 5(1):14–37, 1994. [28] Notion Labs. The ultimate guide to AI-powered knowledge hubs in notion. https://www.notion.com/help/guides/ ultimate-guide-to-ai-powered-knowledge-hubs-in-notion, 2024. Product documentation. [29] Charles Packer, Sarah Wooders, Kevin Lin, et al. MemGPT: Towards LLMs as operating systems. arXiv preprint arXiv:2310.08560, 2023. [30] Palantir Technologies. Palantir ontology: Connecting data to the real world, 2023. Platform Documentation. [31] Preston Rasmussen et al. Zep: A temporal knowledge graph architecture for agent memory. arXiv preprint arXiv:2501.13956, 2025. [32] Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed H. Chi, Nathanael Schärli, and Denny Zhou. Large language models can be easily distracted by irrelevant context. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 31210–31227, 2023. arXiv:2302.00093. [33] Eric W. Stein and Vladimir Zwass. Actualizing organizational memory with information systems. Information Systems Research, 6(2):85–117, 1995. [34] James P. Walsh and Gerardo Rivera Ungson. Organizational memory. Academy of Management Review, 16(1):57–91, 1991. 14
[35] Rongwu Xu et al. Knowledge conflicts for LLMs: A survey. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 2024. [36] Wujiang Xu, Zujie Liang, et al. A-MEM: Agentic memory for LLM agents. arXiv preprint arXiv:2502.12110, 2025. NeurIPS 2025. [37] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, et al. Judging LLM-as-a-judge with MTBench and chatbot arena. In Advances in Neural Information Processing Systems, volume 36, 2023. arXiv:2306.05685.
A
KOC Axis Specification
The Knowledge Object Coordinate is a 7-axis structured identifier: [Entity]-[Domain]-[Class]-[Epoch]-[Depth][Author]-[Variant] Each axis is assigned at ingestion and is immutable thereafter. Structural similarity between two KOs is computed as the weighted sum of axis-wise matches, normalized to [0, 1], at O(1) cost.
B
Edge Type Vocabulary Table 7: Complete KOEdge vocabulary with signed semantic coefficients.
C
Type
Coeff.
Semantics (A → B: A acts on B)
SUPPORTS BASED_ON IMPLEMENTS SUPERSEDES REFINES DERIVES_FROM ENABLES PRECEDES BLOCKS CONTRADICTS
+1.0 +0.8 +0.7 +0.6 +0.5 +0.5 +0.4 +0.3 −0.4 −0.6
A provides evidence strengthening B A is the logical grounding of B A operationally realizes B A replaces B—B is demoted, not deleted A narrows B without contradiction A follows logically from B A is a necessary condition for B A temporally precedes B A actively prevents B A contradicts B (strongest negative gravity)
KGE Force Formulas
Usage force: ui = au · j∈recent exp(−τj /σ), where τj is the time since the j-th retrieval and σ is the recency scale. Evidence force: ei = ae · |{j : (j, i) ∈ E, type(j, i) = SUPPORTS}|, counting new inbound support edges. P Gravity force: gi = ag · j∈N (i) COEFF(j, i) · tanh(gscale · Kjnorm /d(i, j)), where Kjnorm is the z-score-normalized K-score of neighbor j: P
zj =
Kj − µK , max(σK , σfloor )
Kjnorm = max(0, zj )
with µK and σK computed across neighbors of i, and σfloor = 0.5 preventing division instability for small neighborhoods. The sign is a property of the edge coefficient; the magnitude is a statistical property of the neighbor’s score. d(i, j) is the hop distance. 15
Contradiction penalty: ci = ac ·|{j : (j, i) ∈ E, type(j, i) ∈ {CONTRADICTS, BLOCKS}}|. QUESTION urgency: urg(t) = clamp(age_days/30 · 0.3 + B · 0.2 + S · 0.5, 0, 1), where B is the blocking edge count and S is the stakes multiplier.
D
Convergence Proof Sketch
The KGE update K(t + 1) = F (K(t)) defines a map F : [0, 1]n → [0, 1]n . The Jacobian ∂Fi /∂Kj for j ̸= i arises from the gravity term: |∂Fi /∂Kj | ≤ η · ag · |COEFF(j, i)|· gscale /d(i, j)2 . By the Gershgorin circle theorem, F is a contraction if for every node i: the diagonal entry (per-node contraction ∈ [0.845, 0.850]) plus the sum of off-diagonal magnitudes is < 1. This yields the sufficient condition max _degree < gscale / max |COEFF| = 5.0/0.7 ≈ 7.14. The bound is conservative: tanh saturates gravity contributions for high-K neighbors, clamping provides non-expansive projection, and mixed-sign edges cancel. Empirically, convergence holds for max _degree = 43. T2 Crossover Theorem (Theorem 1 — full derivation). Under stationary inputs with u = e = g = c = 0 and K0 = 0.5: ∗ KQUESTION =
η · seed 0.1 × 0.5 = = 0.556 > K0 = 0.5 η + λQUESTION · ∆t 0.1 + (−0.010) × 1
∗ KOBSERVATION =
η · seed η + λOBSERVATION · ∆t
=
0.1 × 0.5 = 0.435 < K0 0.1 + 0.015 × 1
∗ Since KQUESTION
∗ > K0 > KOBSERVATION , the two trajectories diverge at t = 1 (the first discrete update). The asymptotic gap is 0.556 − 0.435 = 0.121; at day 28 the gap is 0.115 score units.
E
Taxonomy Adequacy
Exhaustive pairwise analysis of all 92 = 36 class pairs confirms that every pair differs on at least two of: decay type (none/exponential/inverse), seed K value, half-life, and semantic role. The closest pairs are HYPOTHESIS–OBSERVATION (seed ∆ = 0.10, half-life ratio 1.8×, semantic distinction: testable claim vs. passive signal) and DECISION–CONSTRAINT (seed ∆ = 0.10, semantic distinction: revocable choice vs. structural boundary). No merge is possible without collapsing at least one distinction required by the KGE, retrieval, or contradiction logic.
F
Notation
16
Symbol K, K Keff ϕctx
∗
η λclass ∆t au , ae , ac , ag gscale σfloor α, β, γ we , wd , wa H(q, i) Sstruct , Ssem , Stopo
Meaning Importance score (emergent, per KO); fixed-point value Contextual importance (query-modulated) Contextual attention function: weighted combination of entity, domain, and anchor alignment (Eq. 5) Learning rate / momentum parameter (default 0.15) Class-specific decay rate Time since last KGE cycle (in 6-hour units; default 0.25) Scaling constants for usage, evidence, contradiction, and gravity forces Gravity scale parameter (default 5.0) Z-score floor for gravity normalization (default 0.5) Retrieval weights for structural, semantic, and topological similarity Contextual attention weights for entity, domain, and anchor alignment Hybrid similarity score between query q and KO i Structural, semantic, and topological similarity components
17