Raman Marozau
CTO & Founder of Target Insight Function
Principal Engineering Architect
Raman Marozau · 2026-03-31
Bitcoin has operated for more than fifteen years without AI in consensus-critical validation. That remains true, and this research does not propose changing it.
At the same time, adaptive and ML-based techniques are increasingly applied to operational domains: fee estimation, peer selection, relay policy, Lightning Network routing, and anomaly detection. These integrations are happening incrementally, often without shared acceptance criteria for safety, trust boundaries, failure handling, or systemic effects.
This creates a review gap. Proposals can appear locally reasonable while introducing unclear trust transitions, unstable degradation behavior, or network-level side effects that no single node controls.
This R&D project defines the problem space, proposes a protocol-level enforcement substrate stdio_bus, and provides formal models for trust boundaries, policy enforcement, fallback behavior, latency discipline, and isolation. The research is phased: problem framing first, implementation proposals after community discussion of the problem space.
This research covers non-consensus operational behavior only.
In scope:
Non-goals:
Reader contract: this document presents problems and research directions. The community decides what matters and what to prioritize.
This is timely not because of a single "AI feature," but because review pressure is already visible in active Bitcoin Core development:
Fee estimation accuracy and testability are under active discussion (#27995, #18243). As fee logic becomes more sophisticated, the question of what constraints and evidence reviewers should expect becomes pressing.
Relay and mining policy separation is being discussed (#32401), alongside growing mempool policy composition complexity (#29319). These are policy-layer changes where composition conflicts and precedence ambiguity matter.
Process isolation and boundary hardening are actively tracked in the multiprocess effort (#28722). As components are separated into distinct processes, the question of what isolation guarantees are sufficient becomes concrete.
Network topology resilience is a live concern: net split risks (#33958), peer selection randomization (#34019), and connection diversity (#28462, #16599) are all open threads where adaptive behavior could help or harm depending on how it is constrained.
These are current review surfaces where composition, boundaries, and failure behavior already matter — and where adaptive logic will increasingly be proposed.
Adaptive components and critical node software optimize for fundamentally different properties:
| Adaptive behavior tends toward | Critical node behavior requires |
|---|---|
| Non-deterministic outputs | Reproducible decision paths |
| Opacity in reasoning | Auditable evidence for decisions |
| Diverse failure modes (drift, hallucination) | Graceful, deterministic degradation |
| Variable latency with tail risk | Bounded completion behavior |
| Training-dependent behavior changes | Stable, reviewable invariants |
The formal problem: given a deterministic system S with safety invariants R, and an AI agent A that produces probabilistic recommendations, find an integration architecture that provides utility from A while guaranteeing R holds regardless of A's output quality, maintaining bounded response time, and degrading gracefully when A fails.
This is a structural conflict, not an implementation bug. Without explicit architecture and constraints, these property sets collide in every system that attempts to combine them.
This research identifies six problem areas. Each is independent but connected through a shared review gap: missing acceptance criteria for adaptive components in non-consensus paths.
| ID | Problem Area | Focus | Boundary |
|---|---|---|---|
| R1 | Non-consensus systemic risks | Network/market emergent effects | Not trust semantics or failure handling |
| R2 | Trust boundary and anti-laundering | Data trust states, provenance tracking | Not policy decision logic |
| R3 | Deterministic fallback | Failure handling, recovery protocol | Not timing/budget discipline |
| R4 | Policy enforcement semantics | Decision logic, composition, audit | Not trust provenance |
| R5 | Latency budget discipline | Timing, budget reservation, load | Not failure state machines |
| R6 | Risk-based isolation | Containment, leakage, data minimization | Not policy or trust semantics |
Problem. When many nodes independently deploy similar adaptive strategies, local optimizations can couple into emergent network-level dynamics. Some behaviors remain consensus-safe while becoming network/market-unsafe: transaction propagation fragmentation, convergent fee behavior creating market distortion, peer-selection concentration reducing graph diversity, mining centralization pressure from asymmetric AI advantages, and herd behavior eroding behavioral diversity that itself serves as a resilience property.
Why now in Core. Open issues on peer diversity (#16599, #34019, #28462), net split risks (#33958), fee estimation (#27995), and relay efficiency (#34542) each address individual symptoms. None frame the systemic risk class connecting them.
Research direction. Enforce network-facing constraints through a trusted mediation layer: bounded relay reachability, bounded fee-policy deviation, bounded peer centralization with diversity floors. Coordination-risk controls: model diversity, randomized tie-breaking, anti-correlation noise, synchrony monitoring, operator-local emergency disable conditions with shared observability signals.
Related open issues. #27995, #16599, #34019, #28462, #33958, #34542.
Problem. Without formal trust-state transitions and provenance continuity, advisory outputs can accumulate de facto authority through processing chains. The boundary between untrusted influence and trusted decision input becomes ambiguous. If advisory-influenced data can be transformed and later appear indistinguishable from core-derived data, trust boundaries collapse silently (trust laundering).
Why now in Core. Bitcoin Core already handles trust distinctions implicitly — peer-provided data vs locally validated data, RPC input vs internal state. But there is no explicit vocabulary or checklist for reviewing trust transitions when adaptive/external advisory inputs are introduced into operational decision paths.
Research direction. Propose a minimal trust-state vocabulary (untrusted / validated / trusted) for review of non-consensus decision inputs. Endorsement checks at the mediation boundary (schema, authorization, policy, binding, capability, budget) before advisory data can influence actions. Provenance continuity so advisory influence cannot be laundered. Authorization bound to the original initiator to prevent confused-deputy escalation.
Related open issues. No directly related open issues found — this is an under-explored area in current review practice.
Problem. If fallback and recovery behavior is not specified deterministically, failure handling becomes inconsistent exactly during stress events — unstable degradation and unpredictable recovery when safety margins are thinnest.
Why now in Core. Bitcoin Core handles some failure modes deterministically (peer disconnection on protocol violation, block download timeout). But there is no shared convention for how adaptive components should degrade and recover. As proposals for adaptive fee estimation, peer scoring, or routing emerge, reviewers need minimum fallback properties.
Research direction. Minimum required properties: monotonic degradation (only stricter modes on failure), bounded recovery gating (eligibility, low-risk probing, success-streak validation before promotion), idempotent retries, explicit anti-flapping controls (hysteresis, minimum dwell, cooldown, transition rate limits). Timeout ownership and state transitions in the mediation layer.
Related open issues. No directly related open issues found — fallback/degradation discipline is under-explored for adaptive components.
Problem. As adaptive logic is integrated, validation checks tend to be added at multiple points in application code, producing fragmented enforcement. Without a formal policy model and single deterministic mediation point, it is hard to prove all advisory outputs are evaluated consistently. Composition conflicts and hidden bypass paths become difficult to detect.
Why now in Core. Policy separation (#32401), mempool policy composition (#29319), fee estimation constraints (#27995), and fee testability (#18243) are all active. Each addresses a specific policy domain; none frame general advisory-to-action mediation semantics.
Research direction. For new adaptive policy paths, require one explicit decision path with deterministic semantics and auditable trace. Define effect semantics (permit, deny, forbid with non-overridable precedence), deterministic combining (forbid-overrides default), fail-closed defaults, hierarchical policy layers, complete decision witness for replay, and pre-deploy static analysis (conflicts, shadowing, gaps, obligation incompatibility).
Related open issues. #27995, #32401, #29319, #18243.
Problem. Adaptive inference adds variable latency to operational paths. Without explicit budget reservation, advisory processing can consume time required for mandatory policy, apply, audit, and fallback steps — turning latency variance into reliability and safety risk.
Why now in Core. Block download stalls, RPC batch overload, and IBD sync delays under resource pressure are symptoms of the same underlying problem: no explicit budget discipline separating mandatory processing time from optional/advisory time. As adaptive components are proposed for latency-sensitive paths, reviewers need a convention for timing constraints.
Research direction. For latency-sensitive paths, require explicit timeout budget split, fallback-time reservation (advisory cannot consume fallback budget), and logged timeout decisions for replay and debug. Mode-aware adaptation (normal/degraded/shedding/emergency) with replay-compatible controller state. Under load: admission thresholds, priority scheduling, fairness policy, shedding behavior enforced by the trusted boundary.
Related open issues. No directly related open issues frame this as formal budget discipline for adaptive paths.
Problem. Advisory components are increasingly isolated with process/container mechanisms, but these are often treated as binary ("sandboxed or not") rather than quantified risk controls. Without risk-based guarantees, isolation cannot answer: how much can leak, through which channels, what blast radius remains after compromise.
Why now in Core. The multiprocess effort (#28722) confirms process isolation is an active engineering priority. This research extends the isolation question to include formal leakage guarantees, data classification, and risk-proportional containment.
Research direction (Phase 1 — minimum baseline). Define IPC allowlist, privilege boundaries, fail-closed error handling, and audit hooks. Enforce explicit invariants: no direct core writes, restricted egress via trusted IPC boundary, bounded resources, fail-closed sandbox errors, no cross-tenant flow.
Research direction (Phase 2 — advanced track). Extend to covert-channel modeling (timing/size/order/contention/error), measurable leakage limits, class-based leakage budgets with dual limits on rate and per-session volume, and data minimization architecture where sensitive raw inputs stay outside advisory reach.
Related open issues. #28722.
This risk class deserves expanded treatment because it is the most underestimated and least covered by existing review practice.
A common assumption is that if consensus remains safe, the system is safe. This is false at the network and market level. Certain behaviors remain fully consensus-correct while being systemically harmful:
Propagation fragmentation. Different adaptive relay policies can reduce effective transaction reachability. Some transactions propagate to subsets of the network only — without violating any consensus rule. This degrades the reliability property that users depend on.
Convergent fee behavior. If a significant fraction of nodes use similar adaptive fee strategies, coordinated market effects emerge without explicit coordination. Synchronized fee spikes or drops distort the natural fee market. No individual node is "wrong," but the aggregate effect is harmful.
Mining centralization pressure. Miners with access to superior adaptive optimization gain systematic advantages in fee revenue and block construction efficiency. This creates centralization pressure that undermines the decentralization property Bitcoin depends on — while every block remains consensus-valid.
Peer-selection concentration. Local optimization of peer metrics (latency, reliability) can reduce network graph diversity, increasing vulnerability to partitioning attacks. Each node appears to make a good local decision; the network-level effect is harmful.
Herd behavior and diversity loss. Convergent adaptive strategies reduce behavioral diversity across the network. This diversity is itself a resilience property — it makes coordinated attacks harder. Convergent strategies erode it silently, and the erosion is invisible to consensus-level review.
These effects are particularly dangerous because current review practice does not have a standard way to evaluate them. If we only ask "is consensus safe?", we miss "is the network still healthy and decentralized under these dynamics?"
The formal models in this research require an enforcement mechanism. The research identifies that application-level checks are insufficient — guarantees must be enforced at the protocol/mediation boundary.
Why protocol-level. If trust transitions, policy evaluation, fallback control, timeout ownership, and isolation invariants are implemented as scattered application-level checks, they can be bypassed, forgotten, or composed incorrectly. Protocol-level enforcement means these guarantees are structural — they hold by construction of the communication architecture, not by developer discipline.
stdio_bus as substrate. The stdio_bus transport layer provides the concrete enforcement plane for this research:
What stdio_bus does not solve alone. Network-level systemic effects (R1) require governance, metrics, and operational controls beyond the transport layer. stdio_bus provides the enforcement hooks (policy gates, audit, kill-switch path), but systemic risk management requires deployment-level coordination.
Relationship to phases. stdio_bus is presented here as R&D context. In community-facing discussion (bitcoin-dev, GitHub issues), the problem framing is presented first without implementation specifics. stdio_bus enters the discussion only after problem statements are accepted.
This research is based on:
Not used: closed issues, speculation about future Bitcoin Core direction, claims about specific AI model capabilities.
Comparative evaluation (documented in research) compared 7 architectures across safety, performance, reliability, and cost:
| Architecture | Unsafe Rate (per 10K) | Policy Bypass | Escape Rate | p95 Latency Overhead | Fallback Correctness |
|---|---|---|---|---|---|
| B0 (no AI) | 0.5 (0.1–1.2) | N/A | N/A | 0% | N/A |
| B1 (rule-based) | 0.3 (0.1–0.8) | N/A | N/A | +2% | N/A |
| C0 (direct AI) | 5.2 (3.1–8.4) | N/A | 0.1% | +5% | N/A |
| C4 (full stack) | 0.2 (0.0–0.6) | 0% | 0% | +15% | 100% |
Values show mean with 95% confidence intervals where applicable. Table shows representative subset (B0, B1, C0, C4); full comparison including C1–C3 is documented in the evaluation research (Doc 09).
C4 (full advisory pattern with trust + policy + fallback + isolation) achieves lowest unsafe rate, zero policy bypass, zero escapes, and 100% fallback correctness at 15% latency overhead. It sits on the Pareto frontier for safety-performance trade-off.
C4 is recommended for safety-critical applications with mature governance. For non-critical applications or immature governance, simpler approaches (C3 or B1) may be sufficient.
| Phase | Timing | Action | Goal |
|---|---|---|---|
| A | Week 1 | bitcoin-dev framing post + Issue R1 (systemic risks) | Establish problem framing, get initial reaction |
| B | Week 2 | Issue R4 (policy enforcement) | Most practical issue for Core reviewers |
| C | Week 3 | Umbrella issue (index) + Issue R6 (isolation) | Connect to multiprocess effort |
| D | Weeks 4-6 | Issues R2, R3, R5 | Complete problem surface |
| E | After problem acceptance | Research corpus summary + stdio_bus as substrate | Solution-phase, only after explicit community signal |
| Horizon | Timeframe | Focus |
|---|---|---|
| H1 (Hardening) | 0-18 months | Mechanized proofs of core invariants, runtime covert-channel monitoring, benchmark standardization |
| H2 (Expansion) | 18-36 months | Hardware-backed isolation (TEE), distributed trust boundaries, human factors modeling |
| H3 (Foundational) | 36+ months | Multi-agent safety algebra, unified risk calculus, post-quantum audit integrity |
| Problem Area | Related Open Core Issues | Research Docs | Proposed Discussion Output |
|---|---|---|---|
| R1: Systemic risks | #27995, #16599, #34019, #28462, #33958, #34542 | 07 (Bitcoin case study) | Network stability constraints, coordination-risk controls |
| R2: Trust boundary | (gap — no open issues) | 02 (Trust boundary model) | Minimal trust-state vocabulary for review |
| R3: Fallback | (gap — no open issues) | 04 (Fallback state machine) | Minimum fallback properties for adaptive components |
| R4: Policy enforcement | #27995, #32401, #29319, #18243 | 03 (Policy algebra) | Decision semantics, composition rules, audit requirements |
| R5: Latency budget | (gap — no open issues) | 05 (Latency budget theory) | Budget split convention, fallback-time reservation |
| R6: Isolation | #28722 | 06 (Isolation experiments) | Phase-1 isolation baseline, Phase-2 leakage quantification |
Bitcoin does not need AI to remain a valid protocol. But adaptive operational logic is increasingly relevant in non-consensus domains, and ad hoc integration creates genuine risk — not to consensus, but to network health, decentralization, and operational safety.
This R&D project identifies the review gap, maps six concrete problem areas with verified anchors in current Bitcoin Core development, proposes protocol-level enforcement as the implementation substrate, and provides formal models with comparative evaluation.
The practical question is not "AI yes or no." It is whether adaptive components, where they are introduced, will be governed by shared acceptance criteria and verifiable enforcement — or integrated ad hoc, with safety left to convention.
This research provides the engineering foundation for the first option.