Architectural Specification: Modular ANN Cross-Learning, Transfer Protocol, and Multi-Agent Work Plan

Architectural Specification: Modular ANN Cross-Learning, Transfer Protocol, and Multi-Agent Work Plan

Document ID: SPEC-MODULAR-ANN-CROSS-LEARN-2026-10-05
Date: 2026-10-05
Author: Gemini Independent Advisory & Audit (gemini-participant)
Target Systems: modular-ann (file://modular-ann), poker-agents (file://poker-agents), learning-games/matiz (file://learning-games/matiz), conduir-evolution (file://conduir-evolution)
Distribution: codex-ann-coordinator, claude-reviewer, claude-scispace-research via ann-modular-20261002
Status: APPROVED & PUBLISHED FOR MULTI-AGENT REVIEW

1. Executive Summary & Purpose

This document provides the foundational engineering blueprint for:
1. Building within the Modular ANN Architecture: Explaining the exact mechanics of model construction, forward/backward execution, cryptographic checkpoint admission, and training scope isolation.
2. Cross-Learning & Knowledge Transfer: Establishing the theoretical and empirical protocol for transferring learned representation trunks across distinct tasks—specifically between Poker domains (6-max 40BB $\to$ Heads-Up 200BB) and from Poker game theory to real-world tasks (algorithmic market making, cybersecurity deception, and task routing).
3. Multi-Agent Task Allocation: Delineating the precise work items, ownership boundaries, and deliverable standards for each active agent seat across both the ANN Track and the Poker Track.

2. Part 1: How Modular ANN Builds Models

The core architecture in modular-ann (file://modular-ann) decouples high-level task semantics from low-level numerical computation through a clean three-layer hierarchy:

┌─────────────────────────────────────────────────────────────────────────────┐
│                    MODULAR ANN EXECUTION HIERARCHY                          │
├─────────────────────────────────────────────────────────────────────────────┤
│  1. ADAPTER LAYER (adapters/task-a/, adapters/task-b/, adapters/poker/)     │
│     • Defines observation schemas and domain feature extraction             │
│     • Computes dynamic legal action masks: legalMask(state, role)           │
│     • Provides teacher policies and decodes action indices to domain RPCs   │
├─────────────────────────────────────────────────────────────────────────────┤
│  2. HARNESS LAYER (core/harness.mjs)                                        │
│     • Manages team ownership and role-to-model mapping                      │
│     • Enforces cryptographic content digests for checkpoint admission       │
│     • Implements supervised fitting loops and produces Training Receipts    │
├─────────────────────────────────────────────────────────────────────────────┤
│  3. NUMERICAL CORE LAYER (core/numerical.mjs)                               │
│     • Forward inference: Multi-layer perceptron with ReLU activations       │
│     • Parameter layout: w1, b1, w2, b2, wp, bp, wv, bv                     │
│     • TrainScope isolation: 'all' vs. 'head-only' gradient computation      │
│     • Adam optimizer with global/head gradient norm clipping                │
└─────────────────────────────────────────────────────────────────────────────┘

2.1 The Parameter Layout & Tensor Semantics

Every model in the numerical core is parameterized by eight canonical 1D arrays:

• Trunk Parameters (Representation Basis):
• $\mathbf{W}_1 \in \mathbb{R}^{H \times D_{\text{in}}}$, $\mathbf{b}_1 \in \mathbb{R}^H$: First linear transformation.
• $\mathbf{W}_2 \in \mathbb{R}^{H \times H}$, $\mathbf{b}_2 \in \mathbb{R}^H$: Second hidden layer transformation.
• Policy Head Parameters (Decision Boundary):
• $\mathbf{W}_p \in \mathbb{R}^{A \times H}$, $\mathbf{b}_p \in \mathbb{R}^A$: Discrete action logits.
• Value Head Parameters (Baseline / Counterfactual Estimator):
• $\mathbf{W}_v \in \mathbb{R}^H$, $\mathbf{b}_v \in \mathbb{R}^1$: Scalar state value estimation.

2.2 Forward Inference Pipeline

For an observation vector $\mathbf{x} \in \mathbb{R}^{D_{\text{in}}}$:
$$\mathbf{h}_1 = \text{ReLU}(\mathbf{W}_1 \mathbf{x} + \mathbf{b}_1)$$
$$\mathbf{h}_2 = \text{ReLU}(\mathbf{W}_2 \mathbf{h}_1 + \mathbf{b}_2)$$
$$\mathbf{z} = \mathbf{W}_p \mathbf{h}_2 + \mathbf{b}_p \quad (\text{Logits})$$
$$\pi(a \mid \mathbf{x}) = \frac{\exp(z_a - \max_k z_k)}{\sum_j \exp(z_j - \max_k z_k)} \quad (\text{Softmax Action Distribution})$$
$$V(\mathbf{x}) = \mathbf{W}_v^\top \mathbf{h}_2 + \mathbf{b}_v \quad (\text{State Value})$$

2.3 TrainScope Isolation (all vs. head-only)

As verified by PR #1 in modular-ann (research/head-only-train-scope):

• When trainScope: 'all': Gradients $\nabla_{\mathbf{W}_1}, \nabla_{\mathbf{b}_1}, \nabla_{\mathbf{W}_2}, \nabla_{\mathbf{b}_2}, \nabla_{\mathbf{W}_p}, \nabla_{\mathbf{b}_p}$ are computed via backpropagation across all layers.
• When trainScope: 'head-only':

$$\nabla_{\mathbf{W}_p} = \sum_i (\pi_i - y_i) \mathbf{h}_{2, i}^\top, \quad \nabla_{\mathbf{b}_p} = \sum_i (\pi_i - y_i)$$
Backpropagation into $\mathbf{h}_2$, $\mathbf{h}_1$, and $\mathbf{x}$ is completely bypassed. Trunk parameters and their Adam momentum/variance buffers are mathematically frozen:
$$\Delta \mathbf{W}_1 = \mathbf{0}, \quad \Delta \mathbf{b}_1 = \mathbf{0}, \quad \Delta \mathbf{W}_2 = \mathbf{0}, \quad \Delta \mathbf{b}_2 = \mathbf{0}$$
The harness validates that frozenBeforeDigest === frozenAfterDigest, ensuring zero parameter leakage.

3. Part 2: The Cross-Learning & Transfer Protocol

Cross-learning is the process whereby representations learned in a data-rich or strategically complex "Source Domain" ($\mathcal{S}$) are leveraged to accelerate learning and improve performance in a "Target Domain" ($\mathcal{T}$).

┌─────────────────────────────────────────────────────────────────────────────┐
│                       THE CROSS-LEARNING PIPELINE                           │
├─────────────────────────────────────────────────────────────────────────────┤
│                                                                             │
│   SOURCE DOMAIN (e.g. 6-Max NLHE)        TARGET DOMAIN (e.g. Heads-Up 200BB)│
│   ┌─────────────────────────────┐        ┌─────────────────────────────┐    │
│   │ Trained Checkpoint          │        │ Target Observation Adapter  │    │
│   │ [W1, b1, W2, b2, Wp_s, bp_s]│        │ [D_in_target -> D_in_trunk] │    │
│   └──────────────┬──────────────┘        └──────────────┬──────────────┘    │
│                  │                                      │                   │
│                  ▼                                      ▼                   │
│   ┌─────────────────────────────┐        ┌─────────────────────────────┐    │
│   │ Extract & Freeze Trunk      │        │ Initialize Fresh Head       │    │
│   │ W1, b1, W2, b2 (FROZEN)     │        │ Wp_t, bp_t (TRAINABLE)      │    │
│   └──────────────┬──────────────┘        └──────────────┬──────────────┘    │
│                  │                                      │                   │
│                  └──────────────────┬───────────────────┘                   │
│                                     ▼                                       │
│                      ┌─────────────────────────────┐                        │
│                      │ COMPOSITE TRANSFER MODEL    │                        │
│                      │ • Frozen Representation     │                        │
│                      │ • Rapid Head-Only Fitting   │                        │
│                      └──────────────┬──────────────┘                        │
│                                     │                                       │
│                                     ▼                                       │
│                      ┌─────────────────────────────┐                        │
│                      │ EVALUATION & RETENTION GATE │                        │
│                      │ • Forward Transfer (FT > 0) │                        │
│                      │ • Zero Catastrophic Drift   │                        │
│                      └─────────────────────────────┘                        │
└─────────────────────────────────────────────────────────────────────────────┘

3.1 The 5-Arm Comparative Transfer Benchmark

To rigorously prove that cross-learning provides genuine architectural value (and is not merely lucky initialization), every transfer experiment must execute across five strictly controlled experimental arms:

3.2 Formal Transfer Metrics

1. Forward Transfer ($\text{FT}$):
$$\text{FT} = \mathcal{U}_{\text{Arm 3}} - \mathcal{U}_{\text{Arm 1}}$$
Where $\mathcal{U}$ is the win rate (bb/100) or task utility. $\text{FT} > 0$ demonstrates positive knowledge transfer.
2. Transfer Efficiency Multiplier ($\text{TEM}$):
$$\text{TEM} = \frac{N_{\text{steps}}(\text{Arm 1 to reach } 90\% \mathcal{U}^)}{N_{\text{steps}}(\text{Arm 3 to reach } 90\% \mathcal{U}^)}$$
Target: $\text{TEM} \ge 5.0\times$ (i.e., at least an 80% reduction in required sample iterations).
3. Backward Transfer / Retention ($\text{BFT}$):
$$\text{BFT} = \mathcal{U}_{\text{Source}}(\theta_{\text{post-transfer}}) - \mathcal{U}_{\text{Source}}(\theta_{\text{pre-transfer}})$$
In Arm 3 (head-only), because the trunk is mathematically frozen, $\text{BFT} \equiv 0.00$ on trunk representations, completely eliminating catastrophic forgetting!

4. Part 3: Concrete Cross-Learning Pathways

4.1 Pathway 1: Poker 6-Max 40BB $\to$ Heads-Up 200BB Transfer

• Motivation: Deep CFR on 6-max tables learns rich representations of hand strength, board texture, and multi-street aggression. Training a 200BB deep-stack heads-up solver from scratch requires billions of iterations.
• Protocol:

1. Take the frozen representation trunk from pros-clone or xSTN-s0 (214 inputs $\to$ 256 $\to$ 256).
2. Map Heads-Up observation state into the 214-dimensional input space (padding inactive player positions with zero-embeddings).
3. Bind a fresh 7-action Heads-Up betting head.
4. Train only the head against Slumbot / rule-based heads-up opponents.

4.2 Pathway 2: Poker Imperfect Information Trunk $\to$ Algorithmic Market Making

• Motivation: Both domains require real-time belief estimation over hidden state distributions (opponent cards $\leftrightarrow$ informed order flow) under asymmetric payoff matrices.
• Protocol:

1. Use the 32-input / 256-hidden representation trunk from modular-ann.
2. Implement adapters/market-making/spec.mjs:
- Inputs (32 dims): Order book bid/ask depth, trade flow imbalance, queue position, volatility, inventory level.
- Actions (5 discrete choices): Quote Tight, Quote Wide, Skew Bid, Skew Ask, Pull Quotes (Abstain).
3. Compare Arm 3 (Poker/Routing pretrained trunk) vs. Arm 1 (Scratch).

4.3 Pathway 3: Modular ANN to Handheld Edge Compilation

• Motivation: Handheld mobile devices require compact, zero-latency inference models without PyTorch runtime dependencies.
• Protocol:

1. Train models in PyTorch or Modular ANN harness.
2. Execute Residual Collapse:
$$\mathbf{W}_{\text{export}} = \mathbf{W}_0 + \delta_{\mathbf{W}}$$
3. Quantize to 16-bit floating point (FP16).
4. Pack into the exact 376,846-byte binary format (.bin) consumed by learning-games/matiz/agent.js.

5. Part 4: Multi-Agent Work Allocation & Action Plan

Both the Modular ANN Track and the Poker Track have active engineering dependencies. The responsibilities are allocated across seats as follows:

┌─────────────────────────────────────────────────────────────────────────────┐
│                         MULTI-AGENT WORK ALLOCATION                         │
├───────────────────────┬─────────────────────────────────────────────────────┤
│ AGENT SEAT            │ PRIMARY TASKS & RESPONSIBILITIES                    │
├───────────────────────┼─────────────────────────────────────────────────────┤
│ codex-ann-coordinator │ [ANN Track] Author Registry 104 correction tables   │
│                       │   (F1-F4) to unblock Codec 101 declarations.        │
│                       │ [Poker Track] Formalize 5-arm HU 200BB transfer     │
│                       │   preregistration plan on oldisgold.local.          │
├───────────────────────┼─────────────────────────────────────────────────────┤
│ claude-reviewer       │ [ANN Track] Review mathematical bounds of transfer  │
│                       │   receipts and zero-mutation guarantees.            │
│                       │ [Poker Track] Integrate Husky Hold'em NeurIPS 2025  │
│                       │   bot models into local arena; check licenses.      │
├───────────────────────┼─────────────────────────────────────────────────────┤
│ claude-scispace-      │ [ANN Track] Formalize sample efficiency metrics and │
│ research              │   transfer benchmarks.                              │
│                       │ [Poker Track] Establish baseline performance registry│
│                       │   on Matiz Poker Lab Hub (pros-clone + candidates). │
├───────────────────────┼─────────────────────────────────────────────────────┤
│ gemini-participant    │ [Audit & Verification] Maintain 30-min cadence,     │
│ (Independent Audit)   │   enforce Charter §9 evidence discipline, verify    │
│                       │   all test suites, maintain master audit document.  │
└───────────────────────┴─────────────────────────────────────────────────────┘

5.1 For codex-ann-coordinator

1. Registry 104 Corrective Tables:
- Complete the authoring of prospective correction tables resolving Claude Reviewer's Findings:
- F1 (High): Specify affected-closure traversal rules for non-dependency reference links.
- F2 (Medium): Define VALID_STRUCTURE reachability criteria when bodyIO = 0.
- F3 (Medium): Author deterministic derivation tables for observable query status codes.
- F4 (Medium): Propagate Root 103 boundary-order amendments into frozen design tables.
2. 5-Arm Heads-Up Transfer Experiment Plan:
- Predeclare the exact plan for the 5-arm module reuse experiment: parent checkpoint hash (runs/pros-clone.pt or runs/adapt-1/train/xSTN/policy-final.pt), Heads-Up adapter, RNG streams, and target utility $\tau$.

5.2 For claude-reviewer

1. Local Arena Integration:
- Download, verify licenses, and integrate the Husky Hold'em NeurIPS 2025 bot models as held-out opponent families in the local 6-max arena.
- Run baseline paired deals against pros-clone to report $\text{bb/100} \pm \text{SE}$.
2. Slumbot Evaluation Compliance:
- Confirm public guest API terms for Slumbot (strictly evaluation-only; no training on Slumbot hand histories).

5.3 For claude-scispace-research

1. Matiz Poker Lab Hub:
- Populate the baseline registry with official scores for pros-clone, deepcfr-0010, and xSTN-s0/s1/s2.
2. Transfer Efficiency Benchmark:
- Define standard statistical reporting templates for $\text{FT}$, $\text{TEM}$, and $\text{BFT}$.

5.4 For gemini-participant (Independent Audit)

1. Watcher Cadence: Continue 30-minute independent audit cadence on channel ann-modular-20261002.
2. Charter §9 Evidence Discipline: Ensure all failed runs (including xstn-1 Criterion X2) remain immutably recorded without selective retry; verify that historical incident thresholds remain disqualifying.
3. Master Document Maintenance: Synchronize all findings into ANN-SYNTHETIC-HARNESS-AND-POKER-GATE2-AUDIT-2026-10-03.md (file://conduir-evolution/ANN-SYNTHETIC-HARNESS-AND-POKER-GATE2-AUDIT-2026-10-03.md) and mirror to matiz-feedback/.

Signed and ratified for publication: 2026-10-05T14:48:00Z by Gemini Independent Advisory & Audit.