Hierarchical Active Learning (HAL): A Mathematical Formulation and System Architecture for Sublinear Continuous Active Learning

1. Executive Abstract & Theoretical Framework

The foundational protocols of electronic discovery (eDiscovery) are undergoing a structural and mathematical crisis. As corporate data environments expand into multi-terabyte collections of heterogeneous electronically stored information (ESI)—spanning dense financial ledgers, transactional databases, executive email spools, collaborative chat transcripts, and Building Information Modeling (BIM) files—traditional Technology-Assisted Review (TAR) systems fail under the weight of their own algorithmic assumptions. Specifically, Continuous Active Learning (CAL), hailed for a decade as the gold standard of predictive coding, assumes that responsiveness in complex litigation can be modeled via a single, global binary classification hyperplane.

In complex litigation involving multiple intersecting causes of action, regulatory inquiries, and cross-claims, this assumption is demonstrably false. A global classifier trained across disparate conceptual domains suffers from Context Collapse: the global variance between distinct topics (such as patent specifications versus casual administrative banter) overwhelms the subtle micro-variance separating responsive from non-responsive documents within an individual topic.

The Hierarchical Active Learning (HAL) architecture resolves this breakdown through a rigorous two-tier topological machine learning pipeline. In Tier 1, dense static vector representations are partitioned into semantically uniform Voronoi clusters steered by modified, unit-normalized Rocchio centroids. In Tier 2, localized active learning models operate strictly within individual cluster subspaces. Grounded in the statistical learning theory of Léon Bottou and Vladimir Vapnik (1992), the hierarchical sample complexity bounds of Sanjoy Dasgupta and Daniel Hsu (ICML 2008), the clusterable PAC bounds of Ruth Urner et al. (COLT 2013), and the multi-topic triage optimization frameworks of Eugene Yang, David D. Lewis, and Ophir Frieder (CIKM 2021), HAL achieves sublinear label complexity while providing exact mathematical stopping criteria under the landmark “Knee” formulation of Gordon V. Cormack and Maura R. Grossman (CIKM 2016) and Federal Rule of Civil Procedure 26(b)(1).

2. The Evolution and Epistemological Limitations of Legacy TAR

The technological transition from linear manual review to algorithmic document classification was catalyzed by the landmark National Institute of Standards and Technology (NIST) Text REtrieval Conference (TREC) Legal Track (2006–2011). In their seminal 2011 paper, Maura R. Grossman and Gordon V. Cormack demonstrated that supervised machine learning could consistently surpass human reviewers in both recall (the proportion of relevant documents identified) and precision (the proportion of retrieved documents that are truly relevant).

Early algorithmic review, designated TAR 1.0 (Simple Active Learning / SAL), relied on drawing an initial random statistical sample (a “seed set”) to train a static classifier, typically a linear Support Vector Machine (SVM) or Logistic Regression model. The trained model was subsequently executed across the unreviewed corpus, generating rank-ordered review queues. TAR 1.0 proved brittle: if the initial seed set failed to capture a low-prevalence conceptual pocket, the classifier remained permanently blind to that topic.

TAR 2.0, or Continuous Active Learning (CAL), improved upon TAR 1.0 by eliminating the static seed set. Using Cormack and Grossman’s AutoTAR / Baseline Model Implementation (BMI), CAL continually updates the classifier as human coding proceeds, continuously feeding reviewers the highest-scoring unreviewed documents in a dynamic feedback loop. While CAL dramatically improved review velocity over simple active learning, it retained an obsolete statistical architecture: a single, monolithic binary classifier operating over the entire, unpartitioned corpus.

3. The Binary Fallacy & Context Collapse in High-Dimensional ESI

Legacy CAL enforces an artificially constrained binary classification framework: every document is forced into a scalar label y ∈ {0, 1} (“Responsive” versus “Non-Responsive”). In complex commercial litigation and regulatory proceedings, responsiveness is never binary; it is multi-dimensional, intensely localized, and context-dependent.

Consider a typical multi-district litigation corpus containing executive communications, general ledger journal entries, software source code, HR dispute memos, and routine vendor solicitations. A monolithic global classifier minimizes a global empirical risk objective over the full dataset D:

Because the feature space spans vastly disparate semantic distributions, the global covariance across distinct topics completely dominates the subtle, localized variance within any single topic:

The classifier squanders its structural capacity learning to separate lexical macro-differences (e.g., distinguishing a financial spreadsheet from a marketing email) rather than mastering the subtle evidentiary rules separating a responsive financial statement from an unprivileged, non-responsive one. This breakdown is formally known as Context Collapse.

In foundational learning theory, Léon Bottou and Vladimir Vapnik (1992) demonstrated that when data distributions are non-stationary or multimodal, training a single global hyperplane forces severe compromise across localized sub-domains, resulting in elevated empirical error. Simultaneously, Rich Caruana (1997) proved that forcing unrelated or contradictory tasks into a shared model representation induces negative transfer and catastrophic gradient interference. In eDiscovery, Context Collapse manifests as plateauing recall curves, escalating false positive rates, reviewer fatigue, and severe budget overruns.

4. Dense Vector Geometry & Hyperspherical Invariance

Resolving Context Collapse requires abandoning sparse lexical vector spaces (such as TF-IDF and BM25) in favor of dense, continuous semantic representations. While sparse vectors represent documents across a vocabulary of hundreds of thousands of orthogonal dimensions, they suffer from acute “vocabulary mismatch”—failing to recognize semantic equivalence between distinct legal terminology (e.g., “off-balance sheet transaction” versus “special purpose vehicle”).

Hierarchical Active Learning embeds atomic document chunks into a dense d-dimensional Euclidean space (d = 256) using static vector distillation. To ensure geometric invariance to document chunk length, every feature vector x ∈ R^d is projected onto the surface of a unit hypersphere S^{d-1} via L_2-normalization:

On the unit hypersphere, the semantic proximity between any document vector x and an issue target vector c_k is directly computed as their inner product, which is identical to Cosine Similarity:

5. Tier 1: Macro-Sorting via Rocchio Centroids & Voronoi Partitioning

Rather than forcing a single model to simultaneously separate all legal issues from all background noise, HAL decouples the review pipeline into two specialized tiers. Tier 1: Macro-Sorting organizes unstructured ESI into semantically coherent, topic-pure geometric partitions before active classification begins.

When an attorney codes seed documents for a specific legal issue k (e.g., “Executive Scienter”), HAL computes an initial semantic centroid c_k^{(0)}. As human review proceeds, the system dynamically updates the centroid location using a dense-space formulation of the classic Rocchio relevance feedback algorithm (1971):

# Dense Hyperspherical Rocchio Centroid Update with Inertial Anchoring
def update_rocchio_centroid(c_prev, D_pos, D_neg, alpha=1.0, beta=0.75, gamma=0.15):
    # c_prev: previous normalized centroid vector (1 x 256)
    # D_pos: matrix of positive document vectors (N_pos x 256)
    # D_neg: matrix of negative document vectors (N_neg x 256)
    v_pos = np.mean(D_pos, axis=0) if len(D_pos) > 0 else np.zeros_like(c_prev)
    v_neg = np.mean(D_neg, axis=0) if len(D_neg) > 0 else np.zeros_like(c_prev)

    c_raw = (alpha * c_prev) + (beta * v_pos) - (gamma * v_neg)
    norm = np.linalg.norm(c_raw)
    return c_raw / norm if norm > 0 else c_prev

The hyperparameters α = 1.0, β = 0.75, and γ = 0.15 govern the system’s geometric dynamics: α maintains the inertia of the attorney’s original issue definition, β regulates the gravitational pull of newly verified positive evidence, and γ provides bounded negative repulsion against irrelevant distractors. Immediate L_2-normalization back to S^{d-1} mathematically prevents query drift, ensuring the centroid remains anchored within meaningful semantic space.

Once issue centroids {c_1, c_2, …, c_K} are established, HAL executes a Bounded Voronoi Partitioning across the entire corpus. In classical computational geometry, a Voronoi tessellation assigns every point in space to its nearest centroid. However, in eDiscovery, a vast percentage of collected ESI represents irrelevant corporate noise. HAL enforces a strict noise-rejection threshold τ:

Documents whose similarity to the closest centroid falls below τ (typically calibrated at 0.65 cosine similarity) are excluded from the specialized training queues. This geometric thresholding guarantees that the resulting clusters possess extreme semantic purity, creating an ideal substrate for Tier 2 classification.

6. Tier 2: Micro-Sorting and the VC Dimension Advantage

In Tier 2: Micro-Sorting, independent, localized active learning classifiers are initialized strictly inside each individual Voronoi cluster C_k. The objective function for a Tier 2 classifier is restricted to the local partition:

The profound mathematical superiority of localized classification is established through the statistical learning theory of Vapnik-Chervonenkis (VC) dimension and sample complexity bounds. In standard Probably Approximately Correct (PAC) learning models, the worst-case sample complexity required to learn a concept class to an error bound of ε with confidence 1 - δ scales directly with the VC dimension h of the hypothesis space:

When a global classifier attempts to separate complex legal evidence across a raw, heterogeneous corpus, the effective VC dimension approaches the total ambient dimension of the vocabulary, resulting in bloated sample complexity and requiring thousands of human coding decisions before the model stabilizes.

HAL radically alters this sample complexity regime by leveraging the mathematical principles of Hierarchical Active Learning established by Sanjoy Dasgupta and Daniel Hsu (ICML 2008). Dasgupta and Hsu proved that when active learning is conducted over a hierarchical spatial cluster tree T, the label complexity required to achieve an ε-optimal hypothesis is bounded not by the global VC dimension, but by the tree’s cluster diameter and localized boundary margins:

Because Voronoi partitions outside the margin boundary are pruned in bulk without requesting individual document labels, the algorithm converges with exponentially fewer human queries. Furthermore, as proven by Ruth Urner, Sharon Wulff, and Shai Ben-David (COLT 2013) in their analysis of PL-Clusterable Active Learning, when data distributions exhibit Probabilistically Lipschitz (PL) cluster structure in dense embedding space, localized active learning achieves sublinear sample complexity:

By restricting classification to a homogeneous subspace, the local intrinsic dimensionality d_{local} collapses. The Tier 2 classifier (implemented in Catalyst via non-linear Gradient Boosted Decision Trees / LightGBM) rapidly masters fine-grained legal distinctions—such as disqualifying date ranges or jurisdictional exemptions—without suffering interference from unrelated corporate noise.

# Localized Tier 2 Micro-Sorting within an isolated Voronoi partition
# Prevents global context collapse; sublinear sample complexity
clf = LinearSVC(dual="auto", random_state=42)
clf.fit(x_train, y_train)

# Compute signed decision scores and uncertainty margin distances
predictions = clf.predict(x_unreviewed)
decision_scores = clf.decision_function(x_unreviewed)
margin_distances = np.abs(decision_scores)

# Stream zero-copy columnar predictions to Parquet for client triage
out_table = pa.Table.from_arrays(
    [
        pa.array(chunk_ids_unreviewed, type=pa.string()),
        pa.array(predictions, type=pa.int32()),
        pa.array(margin_distances, type=pa.float32()),
    ],
    names=["chunk_id", "predicted_label", "margin_distance"],
)
pq.write_table(out_table, self.visual_analytics_dir / "tier2_predictions.parquet")

7. Multi-Topic Review Triage & Cost Optimization

In complex litigation involving government subpoenas, antitrust investigations, and commercial fraud claims, legal teams rarely review documents for a single binary issue. Instead, they must evaluate documents across K distinct, correlated legal issues {T_1, T_2, …, T_K} simultaneously.

In traditional eDiscovery, law firms handle multi-topic investigations through one of two deeply flawed workflows: (1) forcing a single reviewer to read every document and manually check 15 different issue boxes (inducing severe cognitive fatigue and high error rates), or (2) running K separate, parallel CAL reviews across the entire corpus (multiplying review costs by K and repeatedly presenting the same document to different reviewers).

HAL incorporates the groundbreaking research of Eugene Yang, David D. Lewis, and Ophir Frieder (ACM CIKM 2021) on Minimizing Cost in Legal Document Review Workflows and Lewis and Yang (DESI VIII / ICAIL 2021). Yang, Lewis, and Frieder formalized the review triage problem as an optimization policy π that routes each document x_i through a sequential decision hurdle to minimize the expected total review cost:

In the HAL architecture, the Tier 1 Gatekeeper filters general corporate noise at minimal inspection cost c_{gate}. Only documents demonstrating substantial responsiveness are cascaded into Tier 2 specialized queues, where senior subject-matter attorneys evaluate specific claims at cost c_{spec, k}. To eliminate false-positive leakage caused by lexical overlap across claims, Catalyst enforces the Bayesian Joint Calibration Formula:

A document is elevated into a high-cost specialized queue (e.g., “Executive Scienter”) only when general litigation responsiveness has been established, mathematically suppressing irrelevant boilerplate containing isolated specialist keywords.

8. Burndown Modeling & Stopping Rules: The Cormack-Grossman Knee

A central challenge in active learning for legal discovery is determining when to stop. Because reviewers evaluate documents in descending order of predicted relevance, the cumulative yield of responsive documents follows a steep concave trajectory. At the start of review, the yield is exceptionally high; as review progresses, the remaining unreviewed responsive documents become increasingly scarce.

In their landmark CIKM 2016 study, Gordon V. Cormack and Maura R. Grossman formulated the “Knee” stopping method for Continuous Active Learning. The instantaneous marginal yield of a review queue is the derivative of cumulative responsive documents found Y(n) with respect to the total documents examined n:

The “Knee” marks the mathematical inflection point where the slope dY / dn drops precipitously, asymptoting toward zero. When the marginal yield falls below an agreed-upon threshold λ_{stop} (typically calibrated at 0.05, representing less than 1 responsive document per 20 reviewed), the review enters the zone of exhaustion.

HAL couples the Cormack-Grossman Knee method with the rigorous statistical stopping rules established by Eugene Yang, David D. Lewis, and Ophir Frieder (ACM Transactions on Information Systems / TOIS 2021): Quant and QuantCI. Quant and QuantCI dynamically certify that a target recall threshold (e.g., Recall ≥ 80%) has been achieved directly from the active learning review stream, constructing exact statistical confidence intervals without requiring a wasteful, pre-review random control set.

Furthermore, HAL integrates the Sequential Probability Ratio Test (SPRT) formulated for eDiscovery by William Webber, Jeremy Pickens, and Victor Zheng (ACM SIGIR 2013). The SPRT tests the null hypothesis H_0: Recall ≥ R_{target} against the alternative H_1: Recall < R_{target} by continuously updating the log-likelihood ratio S_n = ∑_{i=1}^n z_i across the review sequence:

# Sequential Probability Ratio Test (SPRT) Stopping Criterion (Webber et al., SIGIR 2013)
def sprt_evaluation(log_likelihood_ratio, alpha=0.05, beta=0.10):
    # alpha: Type I error (falsely stopping when true recall < target)
    # beta: Type II error (falsely continuing review when target is met)
    A_bound = np.log((1 - beta) / alpha)
    B_bound = np.log(beta / (1 - alpha))

    if log_likelihood_ratio >= A_bound:
        return "CERTIFIED_STOP" # Statistical proof that target recall is achieved
    elif log_likelihood_ratio <= B_bound:
        return "REVISE_QUEUE"   # Queue requires further training
    else:
        return "CONTINUE_REVIEW"

When the stopping criterion is reached, Vantage draws a statistical sample from the Elusion Queue (predicted P < 0.10) and applies exact Clopper-Pearson binomial confidence intervals (Biometrika 1934) to mathematically prove the maximum upper bound of missed documents to opposing counsel and the court.

9. Computational Architecture: Static Distillation & Lakehouse Engine

The economic and physical feasibility of Hierarchical Active Learning depends entirely on vectorization throughput and memory latency. While the legal technology industry has recently attempted to apply massive autoregressive Generative AI models (LLMs) to document review, empirical benchmarks prove that LLM inference is computationally and environmentally unsustainable.

As documented in the TokenPowerBench benchmark (AAAI 2024), a single generative LLM query consumes approximately 2.9 watt-hours of electrical energy. Performing continuous active learning iterations across a 10-million-document litigation corpus using cloud-hosted LLMs incurs astronomical compute bills, introduces multi-minute network latency per batch, and violates client data sovereignty.

HAL circumvents the LLM energy bottleneck by utilizing static vector distillation via Model2Vec. Model2Vec distills contextual transformer networks (such as bge-small-en-v1.5) into compact, static token embeddings through a three-stage distillation pipeline:

• Vocabulary Inference: Passing the entire 30,000-token dictionary of the parent transformer through the network to capture static token vector representations.
• Principal Component Analysis (PCA): Compressing the 768-dimensional transformer embeddings down to 256 dimensions, stripping away high-dimensional statistical noise.
• Zipf Weighting (Intrinsic IDF): Applying Zipf’s law to automatically down-weight ubiquitous corporate stopwords while heavily up-weighting rare, substantive legal terms.

The resulting distilled model occupies a mere 31 MB of memory (a 50x reduction) and processes up to 500,000 sentences per second on standard CPU hardware. At inference time, vectorization requires zero neural network forward passes, executing entirely via contiguous vector lookups in CPU L3 cache.

The data substrate operates on top of immutable Apache Parquet tables and LanceDB, fully integrated with DuckDB. When a reviewer codes a batch of documents, the UI dispatches a lightweight JSON micro-payload. A background Python orchestrator awakes, performs zero-copy vector matrix multiplication against LanceDB memory buffers, updates Rocchio centroids, and writes updated cluster assignments to Parquet. The desktop client polls the JSON manifest and instantly redraws the visual topology via DuckDB with zero UI thread locking.

10. Quality Control, Empirical Flaws & The Ground Truth Fallacy

Commercial eDiscovery software vendors frequently claim predictive models with “98% F1-Scores” by benchmarking against artificial, single-consensus datasets. In actual legal discovery, this level of certainty is an empirical impossibility.

Decades of legal information retrieval research, spearheaded by Grossman, Cormack, and the TREC Legal Track, have demonstrated that human inter-annotator agreement on complex legal evidence rarely exceeds the “moderate” range. When measured using Cohen’s Kappa (κ), attorney-to-attorney agreement consistently falls between 0.45 and 0.65 (representing 60% to 75% raw overlap). Human reviewers suffer from cognitive fatigue, varying interpretations of subjective legal guidelines, and confirmation bias.

Training a monolithic global model on data injected with human disagreement severely degrades the classifier’s decision boundary. By enforcing Tier 1 Bounded Voronoi Partitioning, HAL constrains human coding decisions to semantically uniform document buckets. Reviewers evaluate financial records against financial criteria and employment emails against employment criteria, reducing cognitive friction, substantially increasing intra-cluster coding consistency, and mitigating the “Ground Truth Fallacy.”

11. Judicial Defensibility under FRCP 26(b)(1) & FRE 702

In federal and state litigation, technological innovation is legally useless if it cannot withstand evidentiary scrutiny under the Federal Rules of Civil Procedure and the Federal Rules of Evidence.

Under Federal Rule of Civil Procedure 26(b)(1), parties are entitled to obtain discovery only to the extent that it is proportional to the needs of the case, considering:

• The importance of the issues at stake in the action;
• The amount in controversy;
• The parties’ relative access to relevant information;
• The parties’ resources;
• The importance of the discovery in resolving the issues; and
• Whether the burden or expense of the proposed discovery outweighs its likely benefit.

The Cormack-Grossman Knee method provides mathematical proof of proportionality. When review passes the Knee, the marginal cost per responsive document found escalates hyperbolically:

When dY / dn < 0.02, locating a single additional responsive document requires reviewing over 50 non-responsive files, driving the marginal cost per responsive document to thousands of dollars. Continued linear review beyond the Knee conclusively violates FRCP 26(b)(1) as an undue financial burden yielding negligible evidentiary benefit.

Furthermore, under Federal Rule of Evidence 702 and the Daubert standard (Daubert v. Merrell Dow Pharmaceuticals, Inc., 509 U.S. 579), scientific or technical evidence must be grounded in reliable, testable, and reproducible methodologies with known error rates. HAL’s static vector algebra and LightGBM decision trees are 100% deterministic and reproducible. Unlike non-deterministic Generative AI models that generate varying responsiveness scores across identical runs, HAL outputs an immutable, mathematically verifiable audit log documenting exact centroid coordinates, cosine distances, and decision paths.

12. Regional Application: Litigation Support & Forensics in Greater Houston

The practical execution of Hierarchical Active Learning requires robust on-the-ground digital forensics, litigation support, and trial consulting infrastructure. For corporate counsel, litigation boutiques, and enterprise energy firms managing complex commercial disputes, toxic torts, and regulatory enforcement actions in Brazoria County (Pearland, Lake Jackson, Angleton) and the Greater Houston metropolitan corridor, specialized regional vendors play a vital role in ESI preservation, processing, and defensible court productions.

13. Table of Authorities & Academic References

The mathematical models, empirical principles, and legal defensibility standards incorporated into the Hierarchical Active Learning architecture are grounded in the following academic publications, statutory rules, and judicial precedents:

1. Dasgupta, S., & Hsu, D. (2008). “Hierarchical Sampling for Active Learning.” Proceedings of the 25th International Conference on Machine Learning (ICML 2008), pp. 208–215. doi:10.1145/1390156.1390183.
2. Urner, R., Wulff, S., & Ben-David, S. (2013). “PL-Clusterable Active Learning.” Proceedings of the 26th Conference on Learning Theory (COLT 2013), JMLR W&CP 30:876–897. proceedings.mlr.press/v30/Urner13.html.
3. Bottou, L., & Vapnik, V. (1992). “Local Learning Algorithms.” Neural Computation, 4(6), pp. 888–900. doi:10.1162/neco.1992.4.6.888.
4. Caruana, R. (1997). “Multitask Learning.” Machine Learning, 28(1), pp. 41–75. doi:10.1023/A:1007379606734.
5. Yang, E., Lewis, D. D., & Frieder, O. (2021). “On Minimizing Cost in Legal Document Review Workflows.” Proceedings of the 30th ACM International Conference on Information and Knowledge Management (CIKM 2021), pp. 2331–2340. doi:10.1145/3459637.3482436.
6. Lewis, D. D., & Yang, E. (2021). “Active Learning with Multiple Topics in Legal Discovery.” DESI VIII Workshop at the 18th International Conference on Artificial Intelligence and Law (ICAIL 2021). arXiv:2106.01504.
7. Cormack, G. V., & Grossman, M. R. (2016). “Scalability of Continuous Active Learning for Reliable High-Recall Text Retrieval.” Proceedings of the 25th ACM International on Conference on Information and Knowledge Management (CIKM 2016), pp. 1039–1048. doi:10.1145/2983323.2983776.
8. Yang, E., Lewis, D. D., & Frieder, O. (2021). “Certifying One-Phase Technology-Assisted Reviews.” ACM Transactions on Information Systems (TOIS), 40(1), pp. 1–35. doi:10.1145/3451161.
9. Webber, W., Pickens, J., & Zheng, V. (2013). “Sequential Testing in Classifier Evaluation for eDiscovery.” Proceedings of the 36th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2013), pp. 1057–1060. doi:10.1145/2484028.2484050.
10. Grossman, M. R., & Cormack, G. V. (2011). “Technology-Assisted Review in E-Discovery Can Be More Effective and More Efficient Than Exhaustive Manual Review.” Richmond Journal of Law & Technology, 17(3), Article 5. scholarship.richmond.edu/jolt/vol17/iss3/5.
11. Grimm, P. W., Grossman, M. R., & Cormack, G. V. (2021). “Artificial Intelligence as Evidence.” Northwestern Journal of Technology and Intellectual Property, 19(1), pp. 9–106. scholarlycommons.law.northwestern.edu/njtip/vol19/iss1/2.
12. Rocchio, J. J. (1971). “Relevance Feedback in Information Retrieval.” The SMART Retrieval System: Experiments in Automatic Document Processing, Prentice-Hall, Inc., pp. 313–323. doi:10.1145/361219.361220.
13. Clopper, C. J., & Pearson, E. S. (1934). “The Use of Confidence or Fiducial Limits Illustrated in the Case of the Binomial.” Biometrika, 26(4), pp. 404–413. doi:10.1093/biomet/26.4.404.
14. Minixhofer, B., Pais, V. C., & Naveed, H. (2024). “Model2Vec: Fast State-of-the-Art Static Embeddings.” arXiv preprint. arXiv:2407.13524.
15. The Sedona Conference (2021). “The Sedona Conference Commentary on Defense of Process: Principles and Best Practices for Electronic Discovery.” The Sedona Conference Journal, Vol. 22, pp. 329–418. thesedonaconference.org.
16. Federal Rules of Civil Procedure. Rule 26(b)(1) (Discovery Scope and Limits; Proportionality); Rule 37(e) (Failure to Preserve Electronically Stored Information). Legal Information Institute.
17. Federal Rules of Evidence. Rule 702 (Testimony by Expert Witnesses); Rule 901 (Authenticating or Identifying Evidence). Legal Information Institute.

@techreport{deepthought2026hal,
  author      = {Epistemon and {Deep Thought Labs}},
  title       = {{Hierarchical Active Learning (HAL): A Mathematical Formulation and System Architecture for Sublinear Continuous Active Learning: Overcoming Context Collapse, Multi-Topic Review Triage, and Burndown Optimization via Dense Manifolds and FRCP 26(b)(1) Defensibility}},
  institution = {Deep Thought Labs},
  type        = {Architectural White Paper},
  series      = {Sublate Catalyst Technical Reports},
  year        = {2026},
  month       = {10},
  url         = {https://deepthoughtlabs.ai/papers/hal/},
  note        = {Permanent canonical archive}
}