The Mentorship Protocol

Abstract

Federated learning systems treat aggregation as a stateless operation, discarding the organizational intelligence that human institutions encode through mentorship, hierarchical knowledge transfer, and double-loop learning. This paper provides the organizational design rationale for the Paraconsistent Consensus protocol described in Paper II, arguing that three pillars of organizational learning theory...Senge’s systems thinking [3], Nonaka and Takeuchi’s SECI knowledge creation spiral [4], and Argyris and Schön’s double-loop learning [5]...map to concrete architectural patterns for decentralized agent swarms. The central contribution is the Central Oracle: a knowledge aggregation hub, first proposed in the author’s 2023 working paper [1], that maps to BFT finality checkpoints in the Citrate Network. The Oracle maintains performance profiles, generates targeted LoRA adapter-based mentorship signals, and improves its own aggregation strategy over time...realizing double-loop learning at the protocol level. We position this work as extending da Silva’s [6] formalization of Senge’s Fifth Discipline for multi-agent systems to the specific case of federated learning with LoRA-based adaptation signals. We are explicit about the boundaries of our organizational learning analogy: the SECI mappings are structural analogies that motivated the architecture, not functional equivalences that validate it. No simulation results are reported. We describe three experimental hypotheses and their proposed testing methodology.

Keywords: organizational learning, mentorship, agent swarms, federated meta-learning, Central Oracle, knowledge aggregation, double-loop learning, SECI model, Citrate Network

1. Introduction

The aggregation server in Federated Averaging [2] computes a weighted mean and forgets. It maintains no model of which clients excel at which tasks, provides no targeted guidance, and cannot improve its own aggregation strategy. This is not a limitation of the mathematics...it is a limitation of the organizational design.

Human organizations solved this problem through mentorship: experienced practitioners observe individual strengths, provide targeted guidance, and improve as teachers through the act of teaching [3, 4]. The recursive relationship between teaching quality and student outcomes is the engine that drives organizations from competent to excellent [5]. This paper asks: what organizational design principles, drawn from decades of management science research, should govern the architecture of a decentralized learning system?

We answer by formalizing the Mentorship Protocol...a framework that operationalizes three pillars of organizational learning theory as agent swarm architecture patterns. From Senge [3], we take the principle that organizational intelligence emerges from the interaction patterns between individuals, not from any individual’s capability alone. From Nonaka and Takeuchi [4], we take the SECI knowledge creation spiral that transforms tacit expertise into explicit, transferable knowledge. From Argyris and Schön [5], we take the critical distinction between single-loop learning (correcting errors within a framework) and double-loop learning (questioning the framework itself).

The central contribution is the Central Oracle...a knowledge aggregation hub that collects observations from all agents, maintains performance profiles, and generates targeted mentorship signals. In the Citrate Network [7], this Oracle is realized as the BFT finality checkpoint: a distributed, trustless commitment that periodically crystallizes the network’s collective intelligence and propagates LoRA adapter-based mentorship signals to individual node-models. The Oracle concept was first proposed in the author’s 2023 working paper [1] and is here updated to reflect its implementation path through the Citrate Network’s consensus architecture (Paper I) and paraconsistent aggregation mechanism (Paper II).

Prior work. da Silva [6] formalized Senge’s Fifth Discipline from a multi-agent systems perspective in the Journal of Artificial Societies and Social Simulation (JASSS), using the SMART framework and Z specification language. da Silva’s key findings...that learning organization agents must be honest, cooperative, and tenacious, and that trust is fundamental to their interactions...align precisely with the Byzantine fault tolerance requirements of our checkpoint-based system. We position our contribution as extending da Silva’s formalization to the specific case of federated learning, where the “learning organization” is a decentralized network and the “knowledge transfer” mechanism is LoRA adapter generation at consensus checkpoints.

Implementation status. This paper describes organizational design rationale...it provides the why behind the what specified in Paper II. The underlying consensus infrastructure is implemented (Paper I). The learning extensions are specified but not yet built (Paper II). The organizational learning mappings are theoretical and untested. We use [Rationale] to flag sections that provide design motivation, and [Hypothesis] to flag testable predictions derived from the theory.

2. Organizational Learning Foundations

2.1 Senge’s Five Disciplines

[Rationale] Peter Senge’s The Fifth Discipline [3] identifies five practices that distinguish learning organizations from static ones: personal mastery, mental models, shared vision, team learning, and systems thinking. The fifth discipline...systems thinking...is the integrative framework: organizations fail not because individuals are incompetent but because interactions between individuals produce emergent dysfunctions that no single participant can observe or correct.

For agent swarms, this insight is directly applicable. A swarm of individually capable models can produce collectively poor outcomes if their contributions interfere destructively...the aggregation problem that motivates paraconsistent consensus (Paper II). Systems thinking demands that the orchestration layer model not just individual agent performance but the interaction patterns between agents: which agents complement each other, where redundancy is beneficial, and where contradiction signals genuine disagreement versus noise.

da Silva’s formalization [6] translates Senge’s disciplines into properties of multi-agent systems. He shows that systems thinking requires agents capable of reasoning about the organizational structure, not just their local state...and that this reasoning depends on honest, cooperative behavior that can only emerge in trust-rich environments. In the Citrate Network, the BFT consensus mechanism provides the trust substrate: nodes with blue scores above a threshold have demonstrated honest participation, and only these nodes can serve as mentors (Section 4.3). This is da Silva’s trust requirement realized through consensus.

2.2 The SECI Knowledge Creation Spiral

[Rationale] Nonaka and Takeuchi [4] formalize knowledge creation as a spiral through four modes: Socialization (tacit → tacit), Externalization (tacit → explicit), Combination (explicit → explicit), and Internalization (explicit → tacit). Each transition transforms knowledge from one form to another, and the spiral’s repeated traversal drives organizational learning.

Table 1. SECI Model → Agent Swarm Mapping

SECI Mode

Knowledge Transition

Agent Swarm Mechanism

Analogy Strength

Socialization

Tacit → Tacit

Agents observe peer embeddings via DAG gossip protocol

Weak (see note)

Externalization

Tacit → Explicit

Agent’s internal model state encoded as embedding vector in block

Moderate

Combination

Explicit → Explicit

Paraconsistent aggregation at BFT checkpoint combines embeddings

Strong

Internalization

Explicit → Tacit

Node applies LoRA adapter, modifying internal model behavior

Strong

Note on analogy strength. We rate each mapping’s strength because the SECI modes describe human knowledge processes, not computational ones, and the analogies range from close to loose. Socialization is the weakest mapping. In Nonaka’s framework, socialization involves shared physical experience...apprentices learning through observation of master craftspeople, new employees absorbing organizational culture through proximity. Gossip protocol propagation is not shared experience; it is data broadcast. Nodes receiving peer embeddings through the DAG do not “observe” in any meaningful sense...they receive numerical vectors. We include this mapping for completeness but do not claim it captures the richness of human socialization.

Combination is the strongest mapping. Nonaka describes Combination as the synthesis of explicit knowledge from multiple sources into new explicit knowledge...precisely what the paraconsistent aggregation function does when it combines node embeddings into routing weights and Belnap state vectors at checkpoints. Internalization is also strong: applying a LoRA adapter transforms explicit, transferable knowledge (the adapter weights) into modified internal behavior (the node’s inference patterns), which is what Nonaka means by “learning by doing.”

2.3 Single-Loop and Double-Loop Learning

[Rationale] Argyris and Schön [5] distinguish two learning modes. Single-loop learning corrects errors within an existing framework: the model predicted incorrectly, so we adjust weights. Double-loop learning questions the framework itself: the model keeps failing on this input class...perhaps the routing strategy is wrong, or perhaps the node should specialize rather than generalize.

In the Mentorship Protocol, single-loop learning corresponds to LoRA adapter generation: the meta-model identifies a node’s weakness on a specific input class and generates a targeted correction. Double-loop learning corresponds to the meta-model modifying its own routing weights: changing which nodes receive which queries, restructuring the expert-routing function, or adjusting the diversity regularization parameter. The checkpoint cycle supports both: each checkpoint is an opportunity for single-loop corrections (new adapters) and double-loop restructuring (updated routing weights).

This distinction is the organizational argument for why the routing model must be trainable rather than static. A fixed routing function can only perform single-loop corrections...it can adapt individual nodes but cannot restructure the network’s coordination pattern. A trainable routing model can perform double-loop restructuring, which Argyris and Schön identify as the primary driver of organizational transformation.

3. The Central Oracle

3.1 Architecture and History

The Central Oracle of Truth and Knowledge was first proposed in the author’s 2023 working paper [1] as the organizational hub of a mentorship-driven agent swarm. The Oracle maintains three data structures: (a) a knowledge base aggregating observations from all agents (realized as the embedding vectors accumulated in the DAG); (b) a performance profile tracking each agent’s accuracy across input classes over time (derived from inference accuracy metrics aggregated by the routing model); and (c) a mentorship registry mapping mentor-mentee pairs based on complementary strengths (realized as the LoRA adapter registry at each checkpoint).

In the Citrate Network, the Oracle is not a server process...it is the BFT finality checkpoint itself. The knowledge base is the committed set of embeddings. The performance profile is derived from the routing model’s per-node accuracy metrics. The mentorship registry is the on-chain adapter registry at the LoRAFactory precompile (Paper I, address 0x1003). The Oracle is not a single point of failure: it requires 67+ of 100 validator signatures (Paper I, Section 2.3), making it trustless, immutable, and verifiable.

3.2 From Centralized Oracle to Distributed Checkpoint

The migration from the 2023 centralized Oracle concept to the 2026 distributed checkpoint realization follows a pattern familiar in organizational design: a function that begins as a single role (chief knowledge officer, lead mentor, system architect) is eventually distributed across the organization as processes mature. In the centralized formulation, the Oracle was a single server aggregating client updates. In the distributed formulation, the Oracle’s functions are performed collectively by the finality committee, with BFT consensus ensuring that no single committee member can corrupt the aggregation.

This migration addresses the primary criticism of the original Central Oracle concept: that it reintroduces the single point of failure that decentralization was meant to eliminate. A BFT-committed checkpoint provides the Oracle’s knowledge aggregation and mentorship generation functions while inheriting the safety guarantees of the underlying consensus (Paper I, Section 6.1).

3.3 Mentor-Mentee Assignment

[Specified in Paper II, motivated here] The Oracle assigns mentor-mentee relationships based on complementary performance profiles. Let Pᵢ(c) denote node i’s accuracy on input class c, and P̄(c) the network median. A node is a candidate mentor for class c if Pᵢ(c) > P̄(c) + δ (significantly above average), and a candidate mentee if Pᵢ(c) < P̄(c) − δ (significantly below average). The mentorship signal is a LoRA adapter generated by computing the gradient of the collective loss restricted to class c inputs, compressed to rank-r (default r=16, following standard LoRA configuration [9]).

This assignment process is dynamic: mentor-mentee pairs are re-evaluated at every checkpoint (every 10 blocks, approximately 5 seconds at 2 BPS, per Paper I). A node that was a mentee in checkpoint t may become a mentor at checkpoint t+n after receiving and applying adapters. This mobility is the organizational learning property: the system rewards improvement, not just initial capability. Senge’s concept of “personal mastery” [3] maps directly...each node continuously develops toward its potential rather than being permanently classified by its starting capability.

4. Security and Trust

4.1 The Mentorship Trust Surface

The mentor-mentee dynamic introduces a trust surface that flat aggregation avoids: a malicious mentor could generate poisoned adapters that systematically degrade mentee performance. This is not a theoretical concern...the federated learning literature documents data poisoning and model poisoning attacks where Byzantine participants manipulate aggregated updates [10].

4.2 Trust Mechanisms

The Mentorship Protocol addresses the trust surface through three mechanisms, each grounded in the consensus infrastructure from Paper I:

Blue score gating. Only nodes with blue scores above a configurable threshold can serve as mentor candidates. Blue scores reflect a node’s history of honest, high-quality block production in the GhostDAG consensus (Paper I, Section 2.1). This aligns with da Silva’s finding [6] that learning organization agents must be honest and that trust is fundamental to their interactions...the blue score provides a consensus-derived trust metric that filters potential mentors.

Adapter verification. All adapters are committed on-chain with deterministic hashes and can be re-derived from the committed embeddings and routing model state. Any node can verify an adapter by replaying the generation process. The optimistic fraud proof mechanism (Paper I, Section 3.3) provides a 100-block (~50 second) challenge window for disputes.

Performance regression detection. The routing model tracks per-node performance before and after adapter application. If a node’s performance degrades after applying an adapter, the adapter is flagged for review. Systematic regression triggers adapter revocation. This is the organizational analogue of performance review...mentorship that consistently produces worse outcomes is terminated.

4.3 Byzantine Mentorship

These mechanisms inherit the BFT safety guarantees of the underlying consensus: with fewer than n/3 Byzantine committee members, the checkpoint’s committed routing model state is correct, and therefore adapters derived from it are correctly generated. However, we identify a residual risk: a Byzantine coalition controlling fewer than n/3 committee seats could potentially generate subtly degrading adapters that pass fraud proof verification but introduce long-term drift. Detecting this requires longitudinal performance monitoring across many checkpoints...a capability the immutable checkpoint chain enables but that requires analysis tools that are not yet built. We flag this as an open problem rather than claiming it is solved.

5. Emergent Specialization

[Hypothesis] We hypothesize that the Mentorship Protocol drives emergent specialization through a reinforcing feedback loop: nodes that perform well on specific input classes receive more queries for those classes (via the routing model), generate more training signal for those classes, and accumulate more specialized adapters. Over successive checkpoint cycles, this feedback produces a network of complementary experts rather than homogeneous generalists.

This is the organizational analogue of what Senge [3] calls “team learning”...the process by which a team develops capabilities that exceed the sum of individual capabilities. The routing model acts as the team’s coordination mechanism, ensuring that individual specializations are complementary rather than redundant. The paraconsistent aggregation function (Paper II, Section 3) preserves the information needed for this coordination: the Belnap state vector tells the routing model where nodes agree (T), disagree (B), are uncertain (N), or are confidently wrong (F), enabling it to exploit disagreement rather than suppressing it.

Whether this emergent specialization actually occurs...and whether it produces better collective outcomes than uniform model replication...is an empirical question that we describe testing methodology for in Section 7.

6. Relationship to the Gradient Papers Series

Paper I (Citrate Technical Paper) provides the infrastructure: GhostDAG consensus, BFT finality checkpoints, the LVM with AI precompiles, and the LoRAFactory precompile. The Mentorship Protocol depends on this infrastructure but does not modify it.

Paper II (Paraconsistent Consensus) implements the Mentorship Protocol at the protocol level. The Central Oracle becomes the BFT checkpoint. Mentor-mentee dynamics become the LoRA adapter generation loop. The organizational learning theory in this paper provides the design rationale for why a learned routing model (mentor) should outperform a stateless aggregator (FedAvg). The Belnap FOUR state classification (Paper II, Section 3) is the formal mechanism through which disagreement information...essential for systems thinking...is preserved through aggregation.

Paper VII (The Mozi Cooperative) extends the knowledge-sharing framework to economic mechanisms. The Mentorship Protocol’s principle that value is created through knowledge transfer...not hoarding...is the organizational foundation for cooperative economics: contributors who share expertise should own the infrastructure they improve.

Paper IX (The Medusa Paradigm) provides the biological inspiration. The cnidarian nerve net’s combination of local autonomy and global coordination through periodic synchronization pulses is the biological analogue of the checkpoint-based mentorship cycle. This is an inspirational analogy, not a formal justification (see Paper IX for the full reflective essay).

7. Experimental Hypotheses

Previous versions of this paper reported simulation results (35% faster task completion, 40% error reduction, 60% faster adaptation). These figures were generated as placeholders and do not represent measured outcomes. We replace them with three explicitly labeled hypotheses and proposed experimental designs.

7.1 Hypothesis 1: Targeted Mentorship Outperforms Flat Aggregation

Claim: A mentorship-driven swarm (with targeted LoRA adapter generation based on performance profiles) converges faster to a given accuracy threshold than flat FedAvg aggregation on a heterogeneous task distribution.

Rationale: Organizational learning theory predicts that targeted guidance (mentorship) is more efficient than undirected exploration (flat aggregation) because it reduces the search space for improvement [3, 5]. In federated learning terms, a targeted adapter provides a gradient direction pre-computed from the mentor’s expertise, while FedAvg requires the mentee to discover this direction through its own local training.

Proposed methodology: Deploy N=50 nodes on the Citrate testnet, each with a different training data distribution (non-IID). Run two conditions: (a) FedAvg baseline (weighted averaging at checkpoints, no adapter generation) and (b) Mentorship Protocol (performance profiling, mentor assignment, targeted adapter generation). Measure: rounds to 90% accuracy threshold, final accuracy, per-domain accuracy variance, and total communication overhead. Report with confidence intervals over 10 independent runs.

Expected outcome: We predict the mentorship condition will converge in fewer rounds but with higher per-round communication overhead. The organizational learning argument predicts that the total communication cost will be lower (fewer rounds offsets higher per-round cost), but this is a prediction, not a result.

7.2 Hypothesis 2: Double-Loop Learning Improves Routing Over Time

Claim: A trainable routing model (double-loop learning) produces superior query routing compared to a fixed routing function (single-loop only) after sufficient training time.

Rationale: Argyris and Schön [5] argue that double-loop learning...questioning the framework, not just correcting errors within it...is essential for organizational transformation. In our setting, single-loop learning updates individual nodes via adapters but leaves the routing function fixed. Double-loop learning also updates the routing function, potentially discovering non-obvious node specializations.

Proposed methodology: Compare three conditions: (a) static routing (round-robin), (b) fixed-weight routing (blue-score proportional, not updated after initialization), (c) learned routing (updated at each checkpoint). Measure routing accuracy (fraction of queries directed to the best-performing node for that input class) over 10,000 checkpoints. If learned routing does not outperform fixed-weight routing by a statistically significant margin, the double-loop mechanism adds complexity without benefit.

7.3 Hypothesis 3: Emergent Specialization Produces Complementary Experts

Claim: Over successive checkpoint cycles, the mentorship process produces a network with measurably greater specialization diversity than a network without mentorship.

Proposed methodology: Measure specialization via the Herfindahl index of per-node accuracy profiles: a network of identical generalists has a low Herfindahl index, while a network of distinct specialists has a high one. Compare mentorship vs. flat aggregation over 10,000 checkpoints. Additionally, measure whether specialization is complementary (nodes specialize on different classes) or redundant (nodes specialize on the same classes). Complementary specialization is the predicted organizational learning outcome; redundant specialization would indicate a failure of the routing model’s diversity regularization.

8. Conclusion

The Mentorship Protocol provides organizational design rationale for the Paraconsistent Consensus architecture described in Paper II. Its central contribution...the Central Oracle realized as a BFT finality checkpoint...transforms the stateless aggregation of federated learning into a stateful, adaptive, knowledge-accumulating system. The organizational learning mappings from Senge, Nonaka and Takeuchi, and Argyris and Schön provide the theoretical framework for why this architecture should work: systems thinking explains why modeling agent interactions matters, the SECI spiral explains how knowledge transforms between tacit and explicit forms, and double-loop learning explains why the routing model must be trainable.

We have been explicit about the boundaries of these analogies. The SECI mappings range from strong (Combination and Internalization) to weak (Socialization). The organizational learning predictions are hypotheses derived from theory, not results from experiments. The fabricated simulation metrics that appeared in earlier versions of this paper have been replaced with clearly labeled experimental designs. This honesty is not a limitation...it is the prerequisite for credible empirical validation.

The path from organizational theory to working system runs through Papers I and II: the infrastructure is built, the learning extensions are specified, and the experimental methodology is defined. What remains is the execution.

References

[1] Klosowski, L. (2023). Mentor/Mentee Relativity: Organizational Learning in Mentorship-Driven Swarms. Cnidarian Foundation Working Paper.

[2] McMahan, B., et al. (2017). Communication-efficient learning of deep networks from decentralized data. AISTATS, 1273-1282.

[3] Senge, P. M. (1990). The Fifth Discipline: The Art and Practice of the Learning Organization. Doubleday.

[4] Nonaka, I., & Takeuchi, H. (1995). The Knowledge-Creating Company: How Japanese Companies Create the Dynamics of Innovation. Oxford University Press.

[5] Argyris, C., & Schön, D. A. (1978). Organizational Learning: A Theory of Action Perspective. Addison-Wesley.

[6] da Silva, L. P. (2005). A Formal Model for the Fifth Discipline. Journal of Artificial Societies and Social Simulation (JASSS), 8(3), 6.

[7] Klosowski, L. (2026). Citrate: Protocol Specification for an AI-Native BlockDAG Network. The Gradient Papers No. I. Cnidarian Foundation.

[8] Klosowski, L. (2026). Paraconsistent Consensus: Federated Meta-Learning Over BlockDAG Finality Checkpoints. The Gradient Papers No. II. Cnidarian Foundation.

[9] Hu, E. J., et al. (2021). LoRA: Low-Rank Adaptation of Large Language Models. ICLR 2022.

[10] Blanchard, P., et al. (2017). Machine learning with adversaries: Byzantine tolerant gradient descent. NeurIPS 2017.

[11] Demers, A., et al. (1987). Epidemic algorithms for replicated database maintenance. ACM PODC, 1-12.

[12] Shazeer, N., et al. (2017). Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. ICLR 2017.

[13] Sompolinsky, Y., & Zohar, A. (2018). PHANTOM and GHOSTDAG. IACR Cryptology ePrint Archive.

[14] Kaspa Network. (2025). KIP-14: Crescendo Hardfork. Activated May 5, 2025.

[15] Belnap, N. D. (1977). A useful four-valued logic. In: Dunn, J. M., Epstein, G. (eds) Modern Uses of Multiple-Valued Logic. Springer.

[16] Biderman, D., et al. (2024). LoRA learns less and forgets less. Transactions on Machine Learning Research. Featured Certification.

[17] Ilharco, G., et al. (2023). Editing models with task arithmetic. ICLR 2023.

[18] Bonabeau, E., Dorigo, M., & Theraulaz, G. (1999). Swarm Intelligence: From Natural to Artificial Systems. Oxford University Press.

[19] Busoniu, L., Babuska, R., & De Schutter, B. (2008). A comprehensive survey of multiagent reinforcement learning. IEEE Trans. Syst. Man Cybern. 38(2), 156-172.

[20] Kirkpatrick, J., et al. (2017). Overcoming catastrophic forgetting in neural networks. PNAS, 114(13), 3521-3526.

───

This paper is part of the Gradient Papers series published by the Cnidarian Foundation.

Correspondence: larry@cnidarianfoundation.org