1. The Shape of the Value Curve

The returns are sub-linear for three reasons.

Diminishing marginal information. The first hundred observations of a new operation type establish the baseline. You learn the typical pass rate for each compliance domain, the common failure modes, the expected evaluation duration. The next thousand observations refine the baseline, you learn about seasonal variation, edge cases, the distribution of outcomes rather than just the mean. The next ten thousand primarily confirm what you already know. Within a stable regime, the information per observation declines. A regime shift, a statutory amendment, a change in supervisory practice, re-steepens the curve: the observations that follow it are the most informative in the corpus, which is why the system keeps observing at full resolution, and why anomaly detection exists (Section 5).

Jurisdiction diversity bounds. A zone (one deployment of this institutional stack, running a single jurisdiction’s rules or composing several) with a hundred thousand observations from a single jurisdiction knows that jurisdiction better than a zone with ten thousand observations from ten jurisdictions, and knows less of the range of compliance patterns, the domain-verdict associations Section 4 turns into precedent: it has seen one rulebook applied where the other has seen ten. The comparison is one of breadth, not of depth in the home jurisdiction. A new jurisdiction introduces compliance patterns the corpus has not seen, while additional observations from a known jurisdiction mostly confirm patterns it has. The value of breadth grows concavely in jurisdictional diversity, not linearly, because jurisdictions overlap in the patterns they contribute, most heavily within a legal tradition: the tenth jurisdiction of a tradition adds fewer new patterns than the first, which is the overlap Section 6 weights by legal-tradition similarity.

Irreducible noise. Compliance outcomes have stochasticity that no amount of data can eliminate. A misfiled document, a missed deadline, a regulator exercising discretion on the facts of a single case, these are idiosyncratic events, not patterns to be learned. They are noise. The noise floor sets an asymptote, not a reversal: the estimate of a pattern’s rate keeps sharpening as observations accumulate, but the outcome of the next case keeps its irreducible variance, so the improvement a marginal observation makes to predicting that outcome tends to zero once the stable patterns are established and what remains unexplained is idiosyncratic.

The characterization is a heuristic, not a theorem: intelligence quality should exhibit diminishing returns in observation volume and should benefit from jurisdictional diversity under a concave curve. A logarithmic volume term and square-root diversity term are plausible modeling choices, but this paper does not derive them from a statistical model.

2. Why Sub-Linear Compounding Is Valuable

The alternative to sub-linear compounding, in most deployed systems, is no institutional learning at all.

Consider what compliance infrastructure looks like without an observation corpus. Every operation is evaluated against static rules. The rules are updated manually, based on the intuitions of the people who wrote them, informed by whatever cases they happen to remember. There is no systematic retrieval of past outcomes. There is no anomaly detection. There is no evidence-based feedback into rule changes. The institution has no memory.

Now consider what even modest institutional memory provides:

Precedent retrieval. When a new operation of a familiar type begins, the system can retrieve relevant precedent: “this operation type, in this jurisdiction, for this entity type, has historically passed these compliance domains with 98% reliability and failed this domain 15% of the time.” This doesn’t override the compliance evaluation, the compliance tensor still runs, the rules still apply. But it focuses attention. The compliance officer knows where to look.

Anomaly detection. When a compliance domain’s pass rate drops suddenly, from 95% to 40% over a two-week window, the observation corpus detects the shift. Without the corpus, the drop is invisible until someone manually notices a pattern in individual case outcomes. With the corpus, the anomaly surfaces automatically and can be investigated: is it a regulatory change, a rule misconfiguration, a new class of non-compliant operations, or an upstream degradation?

Evidence-based rule updates. When the intelligence layer proposes a change to compliance rules, the proposal comes with evidence: how many observations, from how many jurisdictions, over what time period, showing what pattern. The governance review process can evaluate the proposal on its evidentiary merits rather than on the authority of whoever proposed it.

None of these capabilities requires the corpus to be enormous. A few thousand well-structured observations already provide meaningful precedent retrieval, since a precedent is certified from the zone’s own record (Section 4). A few months of history already enable anomaly detection. The value starts early and grows steadily, even though it grows slowly.

3. The Feedback Loop

The observation corpus enables a feedback loop that is the mechanism of compounding:

Operations produce observations. Observations accumulate into the corpus. The intelligence layer retrieves precedent, detects anomalies, and constructs rule update proposals. Approved proposals change the compliance rules. Changed rules affect future evaluations. Future evaluations produce new observations. The cycle repeats.

Each iteration is small. A single rule update might change a threshold in a single compliance domain for a single jurisdiction. But each iteration is evidence-based: it is grounded in actual observations of actual outcomes, not in abstract reasoning about what the rules should be.

This is how the common law works. A court decides a case. The decision becomes precedent. Future courts apply the precedent. If the precedent produces bad outcomes, a court distinguishes or overrules it. Each step is incremental. The system learns slowly. But it learns, and the learning is durable because it’s grounded in evidence.

The key architectural constraint is that intelligence influences the rules, not the evaluation. Rule update proposals change the inputs to the compliance tensor. They do not modify the tensor’s evaluation logic. The tensor remains a deterministic function from rules, attestations, and jurisdiction configuration to verdicts. This separation is what makes the feedback loop correctable: every regression is attributable to a specific rule input, and reversible without touching evaluation logic. If a rule update produces worse outcomes, subsequent observations will show the regression, and a follow-up proposal can refine or revert the change. A revert is itself a rule change and travels the path of the update it undoes, so the window between a bad update and its revert is the detection latency of Section 5 plus the latency of the amendment path: the follow-up proposal, governance review, and adoption by the regulator whose rule it is (Section 7). The entities evaluated inside that window bear the cost, a point Section 7 returns to. The evaluation logic itself is never corrupted by the learning process.

4. The Precedent Construction Pipeline

A precedent, in this system, is a reusable compliance pattern: a stable set of domain-verdict associations for a specific operation type, jurisdiction, and entity type, backed by a minimum observation count and stable across the sub-windows of the observation period. Both conditions are local, so a zone certifies precedents from its own record. Zone count is a confidence grade on a certified precedent, not a condition of certification, and it rises above the single-zone grade only where federation (Section 6) supplies the digests that make another zone’s pattern visible; a zone outside federation holds every precedent it certifies at the single-zone grade.

The discovery pipeline runs whenever the zone recomputes its per-jurisdiction operational profiles (Section 5). For each operation-jurisdiction-entity triple with enough observations, it computes domain outcome summaries. Domains with very high or very low pass rates are candidate stable domains. Stability is assessed by partitioning the observation window into sub-windows and checking consistency: a pattern confined to one month fails precedent status. The observation-count threshold and the stability check are the conditions of certification. The confidence grade is then set by zone count: the single-zone grade for a pattern the local record alone supports, rising with each federated partner whose digests show the same stable pattern for the same triple, because diversity of evidence matters as much as volume. The grade is stored with the precedent and reported with every retrieval. The triple that keys this pipeline is itself mutable — an operation family re-cut, an entity-type schema revised — and when a category changes, apparent instability at the boundary is an artifact of the partition rather than a fact about the world, a distinction no volume of observations can recover; so reclassifications are recorded as events, keyed into the comparison windows, and stability is assessed only within a consistent partition.

Precedents serve as prior expectations, not overrides. When a new operation begins, applicable precedents predict which compliance domains will need attention and estimate expected evaluation duration. The compliance tensor still runs. The precedent informs; it does not decide.

The legal analogue is persuasive authority, not binding precedent. The ruling of a coordinate court, or of a foreign court in the comparative reading of Section 6, informs the current case: it creates an expectation, and the current facts are still evaluated on their own merits under the deciding court’s own law. A superior court’s ruling binds the courts below it as law, and nothing this pipeline produces has that force. A precedent here is evidence about outcomes, and it becomes law only when a rule update it supports is adopted through the governance and regulatory process of Sections 3 and 7.

5. Anomaly Detection

The zone maintains an operational profile for each jurisdiction: a probability distribution over operation types, compliance outcomes, evaluation durations, and failure modes, updated incrementally with exponentially weighted moving averages, a control-chart statistic of statistical process control (Shewhart 1931), introduced by Roberts (1959).

Four classes of anomaly merit attention:

A domain failure spike, a historically reliable compliance domain suddenly failing at an unusual rate, indicates regulatory change, rule misconfiguration, or a new class of non-compliant operations. A duration anomaly, operations taking significantly longer than historical norms, indicates external service degradation, new compliance requirements, or a governance bottleneck. A volume anomaly, observation rates deviating from seasonal patterns, indicates market events, regulatory changes, or system issues. A cross-domain correlation anomaly, two historically independent compliance domains beginning to fail together, indicates a common upstream cause.

Anomaly detection is deliberately conservative. It requires sustained deviation, not single-event outliers, the trade between detection delay and false-alarm run length that Lucas and Saccucci (1990) work out for this statistic. This is important because the alternative, a system that alerts on every blip, quickly teaches its operators to ignore it. A compliance officer who receives a hundred false alarms a month stops reading the alerts. Conservative detection preserves the signal.

6. Federated Aggregation

Here is the hard problem: institutional learning is more valuable when institutions can learn from each other, but institutional data is sovereign. A zone’s observations are its own. Raw observations never leave the zone.

The right model is federated aggregation: zones share aggregated, anonymized digests with corridor partners, without pooling raw evidence. The sharing is voluntary (defaults to nobody, a zone must opt in), gated on minimum population (small zones don’t share, because aggregates over small populations leak too much), and constrained by per-entity budgets (no single entity’s observations dominate any shared aggregate).

What a receiving zone gets is typed summaries: per-jurisdiction operational profiles, compliance domain outcome summaries, and detected innovations carrying an evidence grade. Raw data stays behind the producing boundary.

The grade is the warrant for acting on an observation, and it has to be graded rather than binary because the observations differ in what supports them. At the weakest, an anomaly is a deviation from the fleet, the set of comparable zones whose digests the grading zone receives: a rule whose flag-to-confirmed-violation ratio sits far from the fleet median, judged against the fleet’s own dispersion rather than an absolute cut, since the right ratio for a domain is not knowable a priori. That is enough to look, and not enough to tell a regulator, the more so because the fleet is self-selected under the opt-in above, so the fleet median is a biased null model, not a neutral baseline. A fleet-relative statistic also has a floor and a blind spot, both the price of having no absolute reference: below a handful of comparable peers there is no dispersion to judge against, and the grade is unavailable rather than weakly available; and a deviation the whole fleet shares produces no deviation from the fleet, so a fleet-wide error cannot surface at this grade. Above that grade, an observation is supported when the deviation survives a change the zone did not choose — a statutory amendment elsewhere, a corridor opening — which is the ordinary quasi-experimental warrant (Campbell and Stanley 1963; Shadish, Cook, and Campbell 2002). At the top, the zone ran a deliberate variation and measured the outcome. The higher grades are anchored in changes and outcomes rather than in peers, and do not inherit the blind spot.

The distinction matters because the response ladder is keyed to it, lightest touch first: the weakest grade justifies internal tracking only, the middle grade justifies proposing an amendment with the evidence attached, and only the strongest justifies asserting that the amendment will have the predicted effect. Publishing a weak-grade signal as though it were a strong one is the failure this grading exists to prevent, and it is the failure a receiving zone cannot detect for itself, because it sees the summary and not what produced it. The receiving zone incorporates these summaries into its own operational profiles with three weights: a confidence discount, a legal tradition similarity weight, and a recency weight.

Below the ladder sits a forecast rather than an observation. Because evaluation is a deterministic function of rules, recorded attestations, and configuration (Section 3), a proposed rule change can be replayed over the recorded operation corpus before enactment and its mechanical effect counted: which entities’ compliance states change, which violations are created and resolved, which actions become newly blocked or unblocked, which corridor partners’ traffic moves. The forecast has three limits. Behavior is held fixed: the replay cannot observe how parties would have restructured their conduct under a different rule, so it forecasts mechanical effect, never behavioral outcome. Facts recorded from external, time-varying sources are replayed as recorded, not re-queried, so the difference it reports reflects only what the rule change itself can affect. And the forecast is faithful only if the replay computes the same deterministic evaluation function that governs live operations. On the response ladder, a replay attaches to a proposal to scope its impact; it sits below fleet deviation and never justifies asserting that an amendment will have the predicted effect.

The sharing model is right for the same reason that comparative law is useful without being binding. A court in Singapore can learn from how a court in London handled a similar case, but the Singapore court applies its own law, in its own jurisdiction, under its own authority. The London precedent informs; it does not bind. Federated aggregation implements this pattern computationally: sovereign local memory, voluntary sharing of typed summaries, no pooling of raw evidence, no binding authority of foreign intelligence.

7. Auditing the Layer That Does the Scoring

The intelligence layer never decides a case (Sections 3 and 4). Its outputs still reach the governed: through the rule amendments it proposes, which once adopted bind every subsequent evaluation, and through the attention it directs, which determines what gets investigated. A layer with those two effects is a classifier in everything but name, and it will be wrong about some of the entities it touches. Nothing above examines the layer’s own outputs: the architecture holds every other component to a stated error discipline and exempts the component whose errors land on the governed.

Three obligations are owed. The first is an audit on the outputs themselves, and it has to separate three quantities that are usually collapsed. The flag rate by class of entity is not the same as the rate at which flags against that class turn out to be wrong, and neither is the rate at which real violations by that class are missed. A system can show no skew on the first and be badly skewed on the second and third, and it is the second and third that a wrongly flagged entity experiences. The three are tied by counting confirmed violations two ways, flag rate × (1 − false-flag rate) = base rate × (1 − missed-violation rate), where the base rate is the share of a class’s operations that are in fact violations; so where base rates differ between classes, a system that catches any violation at all cannot equalize all three at once, which is the decomposition and the trade-off of the fairness literature (Hardt, Price, and Srebro 2016; Chouldechova 2017; Kleinberg, Mullainathan, and Raghavan 2017). The classes to test against are institutional rather than personal: entity size, age, harbor-set composition, legal tradition of the home jurisdiction. Small and newly incorporated entities are the ones with the thinnest evidence trail, so they are the ones a thin-evidence heuristic will over-flag.

The first two rates are measurable from the zone’s own record: flags are counted directly, and a wrong flag is exposed when the investigation it triggers fails to confirm a violation. The missed-violation rate is not. A wrongly cleared operation is recorded like any other, and nothing in the zone’s own record marks the error. The labels arrive retrospectively, from three sources: a supervisor later flags a specific cleared operation; a counterparty in a cleared operation is later added to a sanctions list, which labels the operations that cleared against it; and a corridor peer, re-evaluating at its border, refuses an operation or entity the local zone cleared. The first two exist for any screening system. The peer refusal is the label only the network supplies, and it is evidence about the local decision only where the corridor’s recognition parameters — which domains the receiver re-evaluates, and how domain names and recognized grades map between the two rulebooks, the corridor parameterization of the companion Sovereign Jurisdiction Network paper — say the two rules are comparable. Refusal under a stricter or unmapped rule is evidence of jurisdictional difference, not of local error. And a candidate missed violation is treated as strong only when independent label sources corroborate it, the graded-warrant discipline of Section 6 applied to the audit’s own evidence.

The second is an audit on the process, and it turns on a distinction the pipeline above does not draw. Amending a rule and amending the system that scores rules are different acts with different authority. A rule change is a legal act and goes to the regulator whose rule it is. A change to the detection model is not a legal act and does not, but it changes every subsequent rule recommendation, so it cannot simply be an engineering decision either. It needs its own reviewer, its own record, and disclosure to the zones downstream of it, because a zone that receives summaries has no way to see that the thing producing them has changed.

The third is a duty on the output side, where the exposure is of the same kind. A supervisor’s attention is finite, so anything that ranks what reaches them is deciding what gets looked at — and a ranking is an attack surface in a way a raw feed is not. An operator who understands the ranking can bury one item by making others look urgent, and no individual signal has to be falsified for it to work. Two properties limit the damage without solving it: the complete stream must be retained and independently queryable, so that what was ranked low was not thereby lost; and whatever ranks it should be tested adversarially against exactly this maneuver — can a party with knowledge of the ranking push a chosen item far enough down the list that it is never read — rather than only against accuracy on ordinary traffic.

None of the three is developed here. What is claimed is that they are owed, and that until they exist the layer’s error rates on the governed are unmeasured and its ranking is untested against the maneuver above.

8. Why Structured Memory Comes First

The objection is direct: why not train a statistical model?

The institutional learning problem has properties that make statistical training a poor foundation.

Small populations. A zone might have a few thousand entities. A jurisdiction within that zone might have a few hundred. An operation type within that jurisdiction might have a few dozen observations per month. These are not internet-scale datasets. They are small, structured, domain-specific populations where every observation matters and where aggregate statistics are meaningful without curve-fitting.

Sovereignty constraints. Training a shared model requires pooling data, or at least pooling gradients. Both leak information: shared gradients can be inverted to recover the records that produced them (Zhu, Liu, and Han 2019), and a trained model can be inverted for its training data or queried for membership in it (Fredrikson, Jha, and Ristenpart 2015; Shokri, Stronati, Song, and Shmatikov 2017). The observation corpus pools neither, it shares typed summaries of aggregated statistics, so the surface that makes federated learning vulnerable, a shared model that can be inverted, is absent. It does not follow that the remaining surface is safe: aggregate releases admit reconstruction without any model to invert (Dinur and Nissim 2003), which is what the population floor and per-entity budgets of Section 6 exist to resist. And those guards bound each digest, not what a partner accumulates over many: the construction carries no privacy budget, the accounting by which differential privacy (Dwork, McSherry, Nissim, and Smith 2006) bounds what a sequence of releases discloses in total (Dwork and Roth 2014), so quantified privacy remains an open obligation rather than a feature.

Regulatory provenance. A statistical model trained on regulatory data faces serious questions about what it was trained on and who authorized the use of that data. The observation corpus contains only observations of operations the zone itself evaluated, under its own authority, so provenance is internal and recorded: which operation, which entity, which rules, and under whose authority are all in the record.

Inspectability. When the observation corpus says “this domain fails 30% of the time,” the basis for that claim is a specific observation count, from specific zones, over a specific time window. A zone operator can inspect the evidence. When an opaque statistical classifier says “this entity is non-compliant,” the basis for that claim is a parameter matrix that no human can interpret. In a domain where decisions must be defensible, where a regulator can ask “why did you flag this entity?” and the answer cannot be “the classifier said so”, inspectability is mandatory.

The foundational layer, the institutional memory from which all intelligence derives, should be structured, inspectable, and sovereign. Any statistical capability layered on top should treat the observation corpus as its substrate, not its replacement.

9. The Case for Institutional Memory

Institutional memory does not make institutions brilliant.

Most institutional infrastructure has no memory at all. Each operation is evaluated in isolation, against static rules, with no reference to what happened last time. The compliance officer’s experience, her personal observation corpus, accumulated over years, is what provides the institutional memory, and it walks out the door when she retires.

The observation corpus makes institutional memory durable and retrievable. It doesn’t replace the compliance officer’s judgment. It gives her better materials to work with: relevant precedent, anomaly alerts, evidence-based proposals for rule changes. And when she retires, the memory stays.

The value compounds sub-linearly. The first year of observations is worth more than the second, which is worth more than the third. But the third year is still worth something. And the tenth year, with precedent spanning regulatory cycles, market crises, and jurisdictional evolution, is worth quite a lot: the years that cross regime shifts carry the observations no volume within a stable regime can supply.

10. Conclusion

Compounding intelligence is the observation that structured institutional memory improves institutional decisions over time. The value curve is sub-linear under the paper’s heuristic model: diminishing returns in observation volume within a stable regime, additional value from jurisdictional diversity, and a return to steep at each regime shift. The mechanism is a feedback loop: operations produce observations, observations surface patterns, patterns generate evidence-based proposals, approved proposals refine rules, refined rules produce better operations. Federated aggregation enables cross-zone learning without pooling sovereign data, following the model of comparative law rather than shared databases. The layer that scores the evidence owes three things: an audit of its outputs, an audit of its own amendments, and a ranking of what it surfaces that keeps the full stream and is tested against an adversary who understands it; Section 7 states them as obligations.

Under the heuristic of Section 1, institutional intelligence follows a concave curve: steep at first, flattening over time, re-steepened by the regime shifts anomaly detection exists to catch.

References

Campbell, D. T. and Stanley, J. C. (1963). Experimental and Quasi-Experimental Designs for Research. Rand McNally.

Chouldechova, A. (2017). “Fair Prediction with Disparate Impact: A Study of Bias in Recidivism Prediction Instruments.” Big Data, 5(2), 153–163.

Dinur, I. and Nissim, K. (2003). “Revealing Information While Preserving Privacy.” In Proceedings of the ACM Symposium on Principles of Database Systems (PODS), 202–210.

Dwork, C., McSherry, F., Nissim, K., and Smith, A. (2006). “Calibrating Noise to Sensitivity in Private Data Analysis.” In Theory of Cryptography (TCC), Lecture Notes in Computer Science 3876, 265–284.

Dwork, C. and Roth, A. (2014). The Algorithmic Foundations of Differential Privacy. Foundations and Trends in Theoretical Computer Science, 9(3–4), 211–407.

Fredrikson, M., Jha, S., and Ristenpart, T. (2015). “Model Inversion Attacks that Exploit Confidence Information and Basic Countermeasures.” In Proceedings of the ACM SIGSAC Conference on Computer and Communications Security (CCS), 1322–1333.

Hardt, M., Price, E., and Srebro, N. (2016). “Equality of Opportunity in Supervised Learning.” In Advances in Neural Information Processing Systems 29 (NeurIPS 2016), 3315–3323.

Kleinberg, J., Mullainathan, S., and Raghavan, M. (2017). “Inherent Trade-Offs in the Fair Determination of Risk Scores.” In 8th Innovations in Theoretical Computer Science Conference (ITCS 2017), LIPIcs 67, 43:1–43:23.

Lorgat, R. The Observation Corpus. Companion paper in this series.

Lorgat, R. The Sovereign Jurisdiction Network. Companion paper in this series.

Lucas, J. M. and Saccucci, M. S. (1990). “Exponentially Weighted Moving Average Control Schemes: Properties and Enhancements.” Technometrics, 32(1), 1–12.

Roberts, S. W. (1959). “Control Chart Tests Based on Geometric Moving Averages.” Technometrics, 1(3), 239–250.

Shadish, W. R., Cook, T. D., and Campbell, D. T. (2002). Experimental and Quasi-Experimental Designs for Generalized Causal Inference. Houghton Mifflin.

Shewhart, W. A. (1931). Economic Control of Quality of Manufactured Product. D. Van Nostrand.

Shokri, R., Stronati, M., Song, C., and Shmatikov, V. (2017). “Membership Inference Attacks Against Machine Learning Models.” In IEEE Symposium on Security and Privacy (S&P), 3–18.

Zhu, L., Liu, Z., and Han, S. (2019). “Deep Leakage from Gradients.” In Advances in Neural Information Processing Systems 32 (NeurIPS 2019).