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 five-component evaluation methodology with explicit circularity analysis. In a controlled comparison (n=10 response pairs), OIDA's RAG condition (3,868 tokens) achieves EQS 0.530 vs.\ 0.848 for a full-context baseline (108,687 tokens); the 28.1× token budget difference is the primary confound. The QUESTION mechanism is statistically validated (Fisher p=0.0325, OR=21.0). The formal properties are established; the decisive ablation at equal token budget (E4) is pre-registered and not yet run.
Figures & tables
Figure 1: OIDA as epistemic memory represented in graph form. Nodes carry epistemic type, provenance, and temporal metadata; edges carry relation type and weight; retrieval combines semantic similarity with graph state.
Type
Initial priority
Temporal behavior
Role
DECISION
1.00
Stable until superseded
Binding choice or commitment
CONSTRAINT
0.90
Stable until changed
Hard structural boundary
EVIDENCE
0.80
Slow decay
Verifiable supporting or refuting material
NARRATIVE
0.70
Stable contextual anchor
Persistent interpretive context
PLAN
0.65
Time-bounded decay
Structured intention with horizon
EVALUATION
0.55
Moderate decay
Informed qualitative assessment
Table 1: Epistemic node types. Initial priorities and temporal behaviors are design priors, not validated optima.
Corpus
Domain
Files / size
Intended stress tests
A / ClearPath
Consulting / operations
46 / 318 KB
Process redesign, bottleneck evidence, unresolved audit workflow, and conflicting organizational observations
Investment decisions, due-diligence gaps, prior-deal lessons, and market-size contradiction
Table 2: Released organizational epistemic-stress corpora. The public repository describes the corpora as synthetic but structurally realistic organizational knowledge bases.
Table 3: Evaluation scope in this draft. The release contains three corpora; the reported analysis uses Corpus A and Corpus C.
Sub-score
Weight
Structured
Full context
Δ
ECA
0.20
0.667
0.807
+0.140
CP
0.25
0.550
0.862
+0.312
CR
0.20
0.418
0.860
+0.442
EC
0.20
0.578
0.788
+0.210
DE
0.15
0.582
0.843
+0.262
EQS
1.00
0.5574
0.8329
+0.2755
Table 4: Corpus A / ClearPath aggregate EQS over three queries. The largest gap is contextual recall, so the result is dominated by breadth rather than a clean test of graph structure.
Metric
Structured
Full context
Ratio / gap
Mean input tokens
2,623
110,539
42.1 ×
Aggregate EQS
0.5574
0.8329
+0.2755
Contextual recall
0.418
0.860
+0.442
Table 5: Corpus A quality–cost summary. Structured retrieval is lower quality in this regime but uses far fewer input tokens.
Condition
Gap section
Interpretation
Structured graph retrieval
10 / 10
Open questions are foregrounded by the representation
Full-context baseline
5 / 10
Gaps appear when salient in the raw context
Table 6: ClearPath knowledge-gap surfacing. This is the pilot’s clearest positive signal, but it remains a small- n behavioral observation.
Sub-score
Weight
Structured
Full context
Δ
ECA
0.20
0.617
0.817
+0.200
CP
0.25
0.550
0.865
+0.315
CR
0.20
0.455
0.863
+0.408
EC
0.20
0.548
0.822
+0.273
DE
0.15
0.520
0.848
+0.328
EQS
1.00
0.5395
0.8438
+0.3043
Table 7: Corpus C / Vertex Minds aggregate EQS. The largest gap is contextual recall, consistent with the full-context baseline’s breadth advantage.
Corpus
Struct. tok.
Full tok.
Struct. EQS
Full EQS
A / ClearPath
2,623
110,539
0.5574
0.8329
C / Vertex Minds
1,944
109,303
0.5395
0.8438
Table 8: In-window quality–cost summary for the two corpus reports analyzed here. Full context is higher quality; structured retrieval uses much less input context.
Figure 2: In-window quality–cost comparison for Corpus A and Corpus C. Full-context access is higher quality in this regime, but uses much more input context. The figure diagnoses the tradeoff rather than claiming a win for structured retrieval.
Figure 3: Corpus C sub-score gap decomposition. The largest gap is contextual recall, so the aggregate result is dominated by context breadth.
Query
Structured retrieval behavior
Full-context behavior
Diagnostic lesson
NovaTech market size and reliability
May present one TAM figure and miss the pitch-deck vs. dossier contradiction
Recovers both figures and treats them as unverified hypotheses
Graph retrieval helps only if both sides of a tension are extracted and retrieved
Table 9: Corpus C contradiction-sensitive query. The example shows why extraction and retrieval completeness are load-bearing for typed graph memory.
Quantity
How to estimate it
Why it matters
Window pressure
Tokens in the task-relevant corpus divided by usable context budget
If pressure is low, full context may be best
Epistemic heterogeneity
Fraction of units that are decisions, plans, hypotheses, evidence, or open questions
Flat chunks are less harmful when all records play the same role
Conflict / supersession rate
Fraction of claim-like units that contradict or replace earlier units
Signed edges matter only when conflicts or changes are common
Open-question rate
Number of unresolved information needs per document or per query
Open-question nodes matter only if uncertainty must be surfaced
Multi-document dependency
Number of source documents needed for a correct answer
Graph state helps most when answers require chains or tensions
Extraction quality
Precision and recall of node types and edge types on a small audit set
Bad graph structure can be worse than no graph structure
Table 10: Pre-deployment diagnostics. These quantities estimate whether epistemic graph memory is likely to help on a deployment dataset.
Condition
What it tests
Decision rule
Full context
Best in-window answer quality when the corpus fits
Use if cost and latency are acceptable
Dense / hybrid RAG
Standard low-cost retrieval without typed graph state
Use if queries are mostly factual and single-hop
Static typed graph
Value of typed nodes and signed edges without time dynamics
Use if state-sensitive queries improve at equal budget
Dynamic typed graph
Added value of time-dependent node weights
Use only if it beats the static typed graph
Table 11: Minimal adoption experiment. Each condition isolates a different source of value.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Type
Coeff.
Semantics ( A→B : A acts on B)
SUPPORTS
+1.0
A provides evidence strengthening B
BASED_ON
+0.8
A is the logical grounding of B
IMPLEMENTS
+0.7
A operationally realizes B
SUPERSEDES
+0.6
A replaces B; B is demoted, not deleted
REFINES
+0.5
A narrows B without contradiction
DERIVES_FROM
+0.5
A follows logically from B
Appendix
Table 12: Edge vocabulary with signed semantic coefficients. Coefficients are implementation priors.
Department of Computer Science, Aalborg University, Aalborg, Denmark · Department of Informatics, University of Oslo, Oslo, Norway · School of Computing, Macquarie University, Sydney, Australia