39. Challenges and Limitations

Caution

The most important chapter in this document. A proposal that cannot articulate its own weaknesses more precisely than its critics can has not been thought through. This chapter states the framework’s limitations without mitigation-first framing — mitigations are noted where they exist, but the limitation is stated first and plainly.

39.1 Limitation Severity Map #

quadrantChart
    title Limitations by Severity and Tractability
    x-axis "Tractable with engineering" --> "Fundamental"
    y-axis "Manageable impact" --> "Programme-threatening"
    quadrant-1 "Accept or do not adopt"
    quadrant-2 "Invest heavily"
    quadrant-3 "Standard engineering"
    quadrant-4 "Design around"
    "Combinatorial explosion": [0.35, 0.85]
    "Spurious correlation": [0.85, 0.80]
    "No causal guarantee": [0.95, 0.55]
    "Entity resolution ceiling": [0.72, 0.82]
    "Cold start / no ground truth": [0.55, 0.70]
    "Privacy amplification": [0.80, 0.88]
    "Skills scarcity": [0.45, 0.78]
    "Operational complexity": [0.30, 0.62]
    "Cost uncertainty": [0.40, 0.66]
    "Automation bias": [0.78, 0.72]
    "Graph write throughput": [0.15, 0.40]
    "Vendor / tool immaturity": [0.25, 0.45]
    "Organizational resistance": [0.62, 0.68]
    "Unvalidated framework": [0.88, 0.92]

Figure 56. Limitations positioned by tractability and severity. The upper-right quadrant — fundamental and programme-threatening — contains privacy amplification, unvalidated framework, spurious correlation, and entity resolution ceiling. These cannot be engineered away. An organization must either accept them or decline to adopt.

39.2 Fundamental Limitations #

These are properties of the approach, not defects of a design.

L1 — Correlation is not causation, and no amount of engineering changes this #

The framework detects statistical and temporal relationships in observational data. Confounding, selection effects, and reverse causation are all indistinguishable from causal relationships without intervention. The framework caps causal confidence at 0.70 (Chapter 19) and labels causal edges as hypotheses, but this is disclosure, not solution.

Consequence: every decision informed by a causal correlation carries an irreducible risk that the intervention will not produce the expected effect because the relationship was not causal.

L2 — Combinatorial explosion is structural #

Correlation cost grows super-linearly in domain count, entity cardinality, window width, and modality count simultaneously (Chapter 26). Scope governance manages this by refusing work, which means the framework’s tractability depends on organizational discipline that will erode under commercial pressure.

Consequence: the framework is only economically viable in organizations willing to say no to correlation requests, indefinitely.

L3 — Correlation quality is bounded by entity resolution quality #

Every correlation depends on knowing that entity X in system A is entity X in system B. Entity resolution is imperfect in every real enterprise. Errors compound: a false merge creates spurious correlations across two unrelated entities; a false split hides genuine ones.

Consequence: the framework’s outputs inherit and amplify an existing weakness rather than compensating for it. Organizations with poor master data will build a system that confidently relates the wrong things.

L4 — Privacy risk is amplified, not merely transferred #

The framework manufactures inferences that were not present in any source system (Chapter 24). Purpose binding, k-anonymity thresholds, and prohibited-inference registries reduce this risk substantially but cannot eliminate it, because the capability to infer is inherent to the architecture.

Consequence: an organization deploying this framework accepts responsibility for a class of harm — inferential harm — that its existing platforms could not have caused.

L5 — The framework is entirely unvalidated #

No production deployment. No benchmark. No independent review. No reference implementation. The design is internally coherent and built from mature components, and that is the strongest statement that can honestly be made.

Consequence: any organization adopting it is conducting the first experiment, at its own cost and risk.

39.3 Tractable but Serious Challenges #

#	Challenge	Description	Mitigation	Residual risk
C1	Cold-start problem	Confidence calibration requires outcome feedback; outcomes are sparse and delayed at launch	Launch at L1 shadow mode; use priors; extend calibration window	Early confidence scores are unreliable and may be trusted anyway
C2	Spurious correlation at scale	Testing many hypotheses continuously produces false discoveries	Multiple-comparison correction; scope discipline; expiry; calibration	Base rate of spurious relationships remains non-zero and scales with scope count
C3	Skills scarcity	Requires streaming, graph, ML, governance, and domain expertise simultaneously	Phased build; platform team model; vendor partnership	Very few practitioners hold this combination; hiring is a real constraint
C4	Operational complexity	More moving parts than any incumbent architecture	Managed services; strong observability; runbooks	Failure modes are novel and initially unfamiliar to operators
C5	Cost uncertainty	Streaming state, graph writes, and unsampled lineage are all expensive	Cost per acted-upon insight; scope-level budgets; tiering	Costs may exceed avoided correlation tax; this is unmeasured
C6	Automation bias	Confident-looking Insight Objects induce over-reliance	Confidence display standards; disagreement tracking; periodic manual review	Human oversight degrades over time in every automated system studied
C7	Graph write throughput	Instance-level graph writes are the throughput bottleneck	Hot/cold tiering; batched writes; confidence-gated persistence	Constrains achievable correlation rate
C8	Alert absence-detection false positives	Negative correlations fire on source outages	Source-health gating; degradation-mode suppression	Gating itself can mask genuine absences
C9	Tooling immaturity	No integrated product exists; assembly required	Build on mature components with standard interfaces	Integration burden falls entirely on the adopter
C10	Organizational resistance	Cross-domain correlation exposes failures that cross ownership boundaries	Executive sponsorship; blameless framing; federated governance	Political resistance is the most common cause of platform failure and is not an engineering problem
C11	Scope creep	Each successful scope invites five more	Hard scope budget; explicit retirement; cost attribution	Discipline erodes; this is the mechanism by which L2 materializes
C12	Regulatory uncertainty	EU AI Act, evolving inference jurisprudence, sectoral rules	Conservative design; legal review; documented purpose binding	Rules may change in ways that invalidate deployed scopes

Table 82. Twelve tractable challenges with mitigations and residual risk. C10 (organizational resistance) deserves particular attention: it is the most common cause of data platform failure generally, and this framework is unusually exposed because its outputs frequently implicate more than one department.

39.4 Failure Modes Specific to This Architecture #

Failure mode	Symptom	Root cause	Detection
Correlation sprawl	Hundreds of low-value scopes, rising cost, falling precision	Absent scope budget and retirement discipline	Cost per acted-upon insight rising; reuse rate falling
Confidence inflation	Everything scores high; discrimination lost	Calibration not performed against realized outcomes	Reliability curve diverges from diagonal
Graph rot	Traversals return stale, irrelevant relationships	Expiry not enforced; decay not applied	Median edge age rising; traversal precision falling
Insight theatre	Insights generated, none acted upon	Poor routing; no ownership; low trust	Action rate per insight near zero
Shadow correlation	Teams rebuild correlations outside the platform	Platform too slow, restrictive, or untrusted	Duplicate logic discovered in application code
Purpose drift	Correlations used beyond their declared purpose	Weak runtime enforcement of purpose binding	Access audit shows out-of-purpose consumption
Autonomy ratchet	Actions promoted to higher autonomy without evidence	Promotion criteria bypassed under pressure	Promotion audit; precision at time of promotion
Silent degradation	Detection quality declines without alarm	Upstream schema or semantic drift	Contract validation; correlation rate anomaly detection

Table 83. Architecture-specific failure modes. Insight theatre — generating insights nobody acts on — is the most likely outcome of a poorly governed deployment and is also the hardest to admit, because the platform appears to be working.

39.5 When Not to Adopt #

Illustrative only

Do not adopt this framework if any of the following is true.

Your data quality and entity resolution are not already good. The framework amplifies both.
Your cross-domain questions are few, stable, and already answered adequately.
You cannot sustain a multi-year investment horizon.
You cannot staff streaming, graph, ML, and governance expertise simultaneously.
You are not prepared to refuse correlation requests indefinitely.
Your regulatory environment prohibits the inferences that would create value.
Your organizational culture punishes the surfacing of cross-departmental failures.
You need a supported product with a vendor to call. This is a design, not a product.

Any one of these is sufficient reason to extend an existing architecture instead.

39.6 What Would Change the Author’s Assessment #

Intellectual honesty requires stating what evidence would revise the position:

Against the framework: a controlled comparison showing well-instrumented incumbent architectures achieve equivalent cross-domain detection at materially lower cost; or evidence that persisted correlations achieve reuse rates below approximately 2 consumers per correlation type; or demonstration that spurious-correlation rates remain unacceptable under realistic scope counts despite correction.
For the framework: a reference implementation achieving the proposed SLOs at realistic scale; sustained calibration convergence in a domain with delayed ground truth; alert compression above 5:1 with recall preserved.

The author’s honest expectation is that the framework’s concepts — persisted scored correlations, correlation scopes as governance objects, the Autonomy Ladder, correlation-grounded RAG — will prove more durable than the framework as a whole. Being absorbed into existing platforms would be a satisfactory outcome.

Key Takeaways #

Five limitations are fundamental and cannot be engineered away: correlation is not causation, combinatorial explosion is structural, quality is bounded by entity resolution, privacy risk is amplified, and the framework is entirely unvalidated.
Organizations with poor master data will build a system that confidently relates the wrong things. The framework amplifies existing weaknesses rather than compensating for them.
Deploying this framework means accepting responsibility for inferential harm — a class of harm existing platforms could not have caused.
Tractability depends on organizational discipline that erodes under commercial pressure, which makes scope creep the mechanism by which combinatorial explosion actually materializes.
Organizational resistance, not engineering difficulty, is the most likely cause of failure, because correlation outputs routinely implicate more than one department.
Insight theatre is the most probable failure of a poorly governed deployment and the hardest to detect, because the platform appears to be working.
Eight conditions each independently disqualify adoption, and any organization meeting one should extend an existing architecture instead.
The author expects the concepts to outlast the framework. Absorption of persisted correlations, correlation scopes, the Autonomy Ladder, and correlation-grounded retrieval into existing platforms would be a satisfactory outcome.

Next: 40. Future Roadmap →