Global Workspace Competition + Broadcast + Working Memory with Ignition Dynamics

Purpose

This skill standardizes the attention + limited-capacity + global broadcast control layer:
taking multi-modal token streams, running capacity-limited competition with deterministic
tie-breaking, stabilizing winners via iterative rounds, triggering "ignition" when a coalition
is stable/confident, and broadcasting selected content to specialist modules—while maintaining
persistent working-memory state across timesteps via CfC/LTC (or GRU fallback).

Key Files

Target Module	Template Asset	Purpose
brain_ai/workspace/competition.py	assets/competition_template.py	Token staging, 4-term scoring, deterministic top-K, slot construction
brain_ai/workspace/global_workspace.py	assets/global_workspace_template.py	Main orchestrator: iterative rounds, convergence, full pipeline
brain_ai/workspace/ignition.py	assets/ignition_template.py	Ignition score, gate, lock-in prevention
brain_ai/workspace/broadcast.py	assets/broadcast_adapters_template.py	Broadcast adapters (temporal, symbolic, decision), feedback
brain_ai/workspace/working_memory.py	assets/working_memory_template.py	CfC/LTC/GRU backends, state persistence, dt handling
brain_ai/config.py (extend)	assets/workspace_config_template.py	CompetitionConfig, RoundsConfig, IgnitionConfig, BroadcastConfig, WMConfig

Public Contract

All workspace interactions use a single interface:

forward(encoder_outputs: Dict[str, EncoderOutput],
        state: Optional[WorkspaceState] = None,
        return_details: bool = False) -> WorkspaceOutput
reset_state(batch_size: int, device: torch.device, dtype=torch.float32) -> WorkspaceState
detach_state(state: WorkspaceState) -> WorkspaceState

Input encoder_outputs maps modality names to EncoderOutput(feats, mask, salience, time).
D == workspace_dim always.

WorkspaceOutput Contract

Every forward call returns:

Field	Shape	Dtype	Description
slots	(B, K, D)	float	Winner token embeddings (workspace slots)
slot_mask	(B, K)	bool	Valid slots per batch item
winners	WinnersMetadata	—	modality_ids, local_ids, scores per slot
ignition_score	(B,)	float	Ignition scalar in [0, 1]
ignited	(B,)	bool	Whether ignition fired
broadcast_packets	Dict[str, BroadcastPacket]	—	Per-specialist adapted tensors
workspace_state	WorkspaceState	—	Persistent WM state + ignition state
telemetry	Optional[Dict]	—	Round history, attention, convergence metrics

Competition Pipeline

Capacity-limited selection with fixed modality ordering and deterministic tie-breaking.

Step	Operation	Output
Stage	Concatenate modality tokens in fixed order	(B, T_total, D) token table
Score	w_content*f + w_salience*s + w_novelty*n + w_task*t	(B, T_total) scores
Gate	Top-K with epsilon tie-breaking	(B, K) winner indices
Construct	Gather winners + optional slot mixer	(B, K, D) slots

Key rules:

• Fixed modality order: ["vision", "text", "audio", "sensors", "engram"]
• Deterministic tie-breaking: score_adj = score + eps1*(-mod_id) + eps2*(-local_id)
• All scoring in fp32 (tie-breaking epsilon requires precision)
• Masked tokens receive -inf score before top-k

Iterative Rounds

Selection → broadcast → re-competition until stable:

Metric	Formula	Threshold
Winner set stability	Jaccard(winners_r, winners_{r-1})	>= 0.9
Slot embedding stability	mean_cosine(slots_r, slots_{r-1})	>= 0.95

Early-stop when both metrics exceed thresholds for M consecutive rounds.
Cap at max_rounds (default 4, range 2–6).

Ignition Dynamics

Non-linear gated broadcast triggered by stable, confident competition.

Ignition Score

ignition = w_stability * stability + w_confidence * confidence
         + w_margin * margin + w_coherence * coherence

Ignition Gate

Condition	Broadcast Gain	Memory Write
ignition >= threshold (committed)	1.0 (full)	Full persistence
ignition < threshold (weak)	weak_gain (e.g., 0.3)	Partial / conservative

Training: smooth sigmoid gate. Inference: hard threshold.

Lock-In Prevention

• Novelty term: 1 - cosine(token, wm_summary) prevents same winners
• Slot dropout: training-only random slot masking
• Winner decay: previous winners get score penalty
• Cooldown: threshold temporarily raised after ignition

Broadcast Adapters

Explicit modules converting slots to specialist-specific formats:

Adapter	Output Shape	Consumer
BroadcastToTemporal	(B, K, D)	HTM temporal memory
BroadcastToPooled	(B, D)	General downstream
BroadcastToSymbolic	(B, N_pred, D_pred)	Reasoning module
BroadcastToDecision	(B, D_decision)	Active inference

Adapters emit aligned masks and support feedback collection from specialists.

Working Memory

Maintains persistent state across timesteps with automatic backend selection.

Backend	Library	Strengths	Fallback
CfC	ncps	Continuous-time, irregular dt	Primary
LTC	ncps	More expressive, long-range	Secondary
GRU	torch.nn	Always available	Fallback

Key requirements:

• reset_state(batch_size, device, dtype) creates fresh state
• detach_state(state) for truncated BPTT
• Ignition-gated write: next_state = gain * new + (1-gain) * old
• dt passed as timespans to CfC/LTC; concat/gate_scale for GRU
• Downstream code never branches on backend type

Configuration Surface

CompetitionConfig

Field	Default	Purpose
capacity_limit	7	K workspace slots (Miller's Law)
num_heads	16	Multi-head attention heads
w_content	1.0	Content scoring weight
w_salience	0.5	Salience scoring weight
w_novelty	0.3	Novelty scoring weight
w_task	0.0	Task bias weight
eps_tie_modality	1e-6	Tie-breaking epsilon (modality)
eps_tie_token	1e-8	Tie-breaking epsilon (token)
use_slot_mixer	True	Diversity-encouraging MLP

IgnitionConfig

Field	Default	Purpose
ignition_threshold	0.3	Ignition firing threshold
w_stability	0.3	Stability weight in ignition
w_confidence	0.3	Confidence weight
weak_gain	0.3	Broadcast gain when not ignited
slot_dropout	0.1	Training slot dropout rate
winner_decay	0.1	Decay for previous winners
cooldown_steps	0	Post-ignition cooldown
w_margin	0.2	Margin weight in ignition score
w_coherence	0.2	Cross-modal coherence weight

See references/ignition-dynamics.md Section 6 for the complete field reference
including gate_temperature, use_learned_ignition, and cooldown_boost.

WMConfig

Field	Default	Purpose
mode	"auto"	Backend: cfc, ltc, gru, auto
hidden_dim	4096	Hidden state dimension
buffer_capacity	7	Memory buffer entries
use_dt	True	Pass time deltas to backend
dt_input_mode	"concat"	How GRU handles dt

Presets: WorkspaceConfig.minimal(), WorkspaceConfig.dev(), WorkspaceConfig.production_1b(),
WorkspaceConfig.production_3b(), WorkspaceConfig.production_7b().

Common Failure Modes

Symptom	Cause	Fix
Non-deterministic winners	Unstable tie-breaking in top-k	Use epsilon adjustment with fixed modality order
Workspace "stuck" on same tokens	No lock-in prevention	Enable novelty, winner decay, slot dropout
Ignition fires every step	Threshold too low	Raise ignition_threshold, enable cooldown
Ignition never fires	Threshold too high or unstable rounds	Lower threshold, increase max_rounds
Working memory drift	State in fp16	Force fp32 for all WM state
Broadcast shape mismatch	Missing adapter for consumer	Register appropriate BroadcastAdapter
Iterative rounds never converge	Thresholds too strict	Lower stability_threshold, increase max_rounds
GRU ignoring time deltas	dt not wired through	Set use_dt=True, configure dt_input_mode

Anti-Patterns

• Nondeterministic modality ordering — use fixed MODALITY_ORDER constant
• Relying on GPU sort stability — always use epsilon tie-breaking
• Pooling to (B, D) before broadcast — keep (B, K, D) slots for adapters
• Running > 6 iterative rounds — latency trap; keep max_rounds <= 6
• Using fp16 for competition scores — tie-breaking epsilon needs fp32
• No ignition gate — broadcast same gain regardless of stability
• Branching on WM backend outside working memory module
• Unconditional memory write — use ignition-gated persistence
• Forgetting detach_state in truncated BPTT — gradient explosion

Additional Resources

Reference Files

• references/competition-broadcast.md — Full spec: token staging, 4-term scoring, deterministic top-K, slot construction, iterative rounds, convergence
• references/ignition-dynamics.md — Ignition score, gate, lock-in prevention, GNW theory mapping
• references/working-memory.md — CfC/LTC/GRU backends, state persistence, dt handling, ignition-gated write
• references/testing-matrix.md — All test cases: deterministic top-k, convergence, state persistence, ncps fallback

Asset Templates

• assets/competition_template.py — TokenStager, CompetitionScorer, DeterministicTopK, SlotConstructor, self-test
• assets/global_workspace_template.py — GlobalWorkspace orchestrator, IterativeRoundManager, WorkspaceOutput, self-test
• assets/ignition_template.py — IgnitionScorer, IgnitionGate, LockInPrevention, component computers, self-test
• assets/broadcast_adapters_template.py — BroadcastAdapter base, 4 adapters, registry, FeedbackCollector, self-test
• assets/working_memory_template.py — WorkingMemory, CfC/LTC/GRU backends, MemoryBuffer, state management, self-test
• assets/workspace_config_template.py — CompetitionConfig, RoundsConfig, IgnitionConfig, BroadcastConfig, WMConfig, presets

Scripts

• scripts/validate_workspace.py — Runtime contract validation (competition/rounds/ignition/broadcast/WM checks)
• scripts/gen_workspace_tests.py — Generates tests/test_workspace.py (~80+ test cases)
• scripts/workspace_benchmark.py — Performance benchmarking (throughput, latency, memory, scaling)