Forge System End-State

1 System Architecture

▼

End-to-end data flow from user input to product repositories. Every arrow is a live connection in production.

User
Telegram / CLI / Web

➔

OpenClaw Gateway
:18789
WS + Web UI relay

➔

Forge Orchestrator
State machines + dispatch

➔

API Bridge
:3100
Model routing + enrichment

➔

Hermes Brain
:18810
Cross-track coordinator

Research Loops
Swarma + MiroFish
:3300

➔

Sprint Pipeline
12-node state machine

➔

Product Repos
antilles-v2, lottery

←

Memory Router

:3200

13 providers + Event Bus

←

Cost Tracker

:3500

Budget alerts

Service Map

API Bridge

:3100

OpenClaw ↔ Toolkit ↔ Model providers. Intent classification, circuit breaker, rate limiting.

Memory Router

:3200

Unified memory API: 13 providers, MCP + HTTP transports, event bus with watchers.

OpenClaw Gateway

:18789

Always-on chat gateway. WebSocket + Web UI. Telegram relay.

Autoresearch

:3300

L3 research pipeline service. Swarma outer loop, experiment inner loop.

Cost Tracker

:3500

Paperclip L2.5. Cross-track cost tracking, budget alerts, spend visualization.

Ollama

:11434

Local inference: qwen3-coder:30b, qwen3:32b, glm4:9b, qwen2.5-coder:7b.

MLX Server

:8400

Apple Silicon accelerated inference. OpenAI-compatible API.

MiroFish Backend

:5002

Swarm intelligence simulation engine. Flask, Apple Silicon fork.

MiroFish MLX

:8080

Agent inference via mlx-lm. Qwen3-4B-4bit.

MiroFish Frontend

:3000

Vue 3 + Vite web UI for simulation visualization.

Credential Broker

internal

macOS Keychain credential access with ACL enforcement.

Hermes Orchestrator

:18810

Central coordinator: cross-track synthesis, budget, CTO reporting.

Hermes Per-Track Instances

RMT Reputation

:18800

Identity Attestation

:18801

Lottery

:18802

x402 Trust Channels

:18803

Event Bus

Reactive pub/sub system with watchers (Sprint, Forge, Git) feeding handlers (CxdbIngest, WorkingMemoryUpdate).

sprint:transition

sprint:stall

research:experiment

research:convergence

research:review

forge:transition

forge:review

git:commit

eval:complete

2 Memory Architecture

▼

13 memory providers federated through Memory Router (:3200). SOUL/MEMORY/USER scoping. Results scored and ranked.

#	Provider	Type	Size / Entries	Writable	Status	Notes
1	CXDB	sqlite+fts5	57 MB / 51K entries, 57K edges	Yes	Connected	Primary knowledge graph. Attestations, graph edges, FTS5 search.
2	Claude Memory	file-scan	~/.claude/projects	Yes	Connected	Claude project memory files. Primary conversation context.
3	Working Memory	file-scan	11 scan paths	Yes	Connected	Structured sections in work-stream .md files. Auto-trimmed at 2200 chars.
4	Sprint State	json	1 checkpoint file	No	Connected	Read-only sprint checkpoint view.
5	Git History	git-log	3 repos (antilles, forge, lottery)	No	Connected	Searchable git log across all repos.
6	Session Logs	file-scan	~/.openclaw/logs	No	Connected	OpenClaw session log search.
7	RMT Reputation	sqlite	~/.forge/rmt-reputation.sqlite	Yes	Connected	Dedicated RMT reputation data store.
8	Hermes Memory	sqlite	17 entries	Yes	Partial	Hermes bridge memory. Low entry count -- needs enrichment from state.db.
9	Hermes state.db	sqlite	2.4 MB	Yes	Partial	Hermes persistent state. Needs connecting via bridge provider.
10	OpenViking	tiered-api	L0/L1/L2 tiers	No	Partial	83% token reduction via tiered context. Config present, server needs activation.
11	OpenClaw Sessions	file-scan	30-day retention	No	Connected	OpenClaw workspace memory markdown files.
12	Mem0 / Qdrant	vector-db	Not deployed	Yes	Not Connected	Semantic vector search. Planned for Phase 3 deployment.
13	Dispatch Metrics	jsonl	Planned	No	Not Connected	OpenClaw dispatch metrics aggregation. Planned Phase 3.

8 Connected

3 Partial

2 Not Connected

3 Research Pipeline

▼

Autonomous research loop: scientific method enforcement with 5-dimensional eval matrix. Swarma outer loop drives experiment scheduling.

Full Research Loop

Eval Criteria

➔

Data Loading

➔

Swarma

➔

MiroFish Sim

➔

Scoring

➔

Verdict

➔

Development

➔

Verify

↻

Loop

5-Dimensional Eval Matrix

Scenario

Compound Attack, Eclipse, Sybil Farm

Actor

Honest, Colluding, Adversarial

Scale

100 / 1K / 5K / 10K agents

Condition

Normal, Degraded, Byzantine

Metric

Sybil %, Precision, F1, Latency

Scientific Method Enforcement

1. Hypothesis

Falsifiable claim with measurable predictions

2. Experiment

Controlled parameter sweeps across eval matrix

3. Evidence

Quantitative results stored in CXDB

4. Cross-Review

Multi-model verdict with Grok CTO override

5. Iterate

Strategy update feeds next hypothesis cycle

4 State Machines

▼

Two interlocking state machines with guard-enforced transitions. Cross-model review is mandatory at every gate.

Research State Machine (6 nodes)

STRATEGY_DRAFT

➔

STRATEGY_REVIEW

➔

EXPERIMENT_RUNNING

➔

EXPERIMENT_SCORING

➔

VERDICT

➔

STRATEGY_UPDATE

↻

Bridge: VERDICT → BUILD_NEEDED → Sprint KICKOFF — Research findings trigger product sprints when implementation is warranted.

Sprint State Machine (12 nodes)

CARRY_CHECK

➔

SEED

➔

PLAN

➔

DESIGN

➔

BUILD

➔

VALIDATE

➔

CODE_REVIEW

➔

AUDIT

➔

SIMPLIFY

➔

RETRO

➔

COMMIT

➔

COMPLETE

Guard Enforcement

No Self-Review

assertCrossModelReview() enforced at every gate. Writing model cannot review.

Grok Mandatory

CTO oversight required in every review panel. Final binding decision on Round 3.

Min 2 Reviewers

At least two distinct models must review each diff before gate proceeds.

3-Round Deliberation

R1: Independent review. R2: Shared findings debate. R3: Grok decides.

5 Product Tracks

▼

Four active product tracks. Each has a dedicated Hermes instance, research loop, and sprint pipeline.

RMT Reputation

antilles-v2/feature/rmt-core-porting

Production Ready

0.9397

Composite

93.7%

Sybil Detection

0.96

Precision

Research Complete94%

Identity Attestation

antilles-v2/feature/identity-foundation

Building

304

Commits

Portal

UI Built

ADR-2

Revocation

Identity Foundation72%

x402 Trust Channels

antilles-v2/agent/x402

Research Phase

x402

Protocol

80/20

Fee Split

3

Scenarios

Research + Strategy35%

Lottery

lottery/main

Independent Repo

VRF

Randomness

ZK

Proofs

Fair

Distribution

Research + Params40%

6 Competitive Landscape

▼

Threat map of competitive and adjacent projects in decentralized reputation / identity space.

Helixa

132K agents

HIGH

Large-scale agent reputation network. Direct competitor in decentralized identity scoring. First-mover advantage in agent-to-agent trust.

ERC-8004

130K deployed

HIGH

ERC standard for on-chain agent identity. Gaining adoption rapidly. Threatens to become the default identity primitive.

Virtuals Protocol

18K agents

MEDIUM

Virtual agent marketplace. Different focus (entertainment) but expanding into reputation and trust scoring.

Cred Protocol

Growing

MEDIUM

Credential protocol for decentralized reputation. Overlapping trust model but weaker on sybil resistance.

FairScale

Inactive

DEAD

Previously attempted fair reputation scoring. Project abandoned. Code open-sourced -- review for useful primitives.

7 Execution Status

▼

Four-phase execution plan. Currently in Phase 1 with infrastructure buildout nearing completion.

1

Phase 1: Forge Infrastructure

In Progress

Core infrastructure, state machines, memory federation, API bridge, model dispatch, cost tracking.

State machines API Bridge Memory Router CXDB Cost Tracker Event Bus Credential Broker Cross-model review Eval framework Hermes integration OpenViking activation LaunchAgent hardening

2

Phase 2: Research Loops

Swarma harness MiroFish engine Scenario library Continuous orchestrator Eureka integration Dashboard generator Track loop automation Convergence alerts

3

Phase 3: Products

After Research

Product shipping: RMT core porting, identity portal launch, x402 prototype, lottery contract deployment.

RMT core port Identity portal x402 prototype Lottery contracts Mem0/Qdrant deploy Dispatch metrics

4

Phase 4: Advanced

Future

Advanced capabilities: full autonomous operation, cross-product synthesis, on-chain deployment, governance.

Autonomous sprints Cross-product graph On-chain deploy Governance module External API

8 Economic Model

▼

Revenue model combining FICO-style credit scoring, staking mechanics, and payment channel fees.

★

FICO Model

Reputation scores sold to relying parties. Charge per-query or subscription for trust assessments. Higher accuracy = higher price tier.

⚖

ve-Staking for TAs

Trust Anchors stake tokens with vote-escrow mechanism. Longer lock = higher multiplier. Slashing for misbehavior enforces honest attestation.

♦

80/20 Channel Fees

x402 payment channels: 80% to service provider, 20% protocol fee. Micro-payments for trust verification. Volume scales with agent adoption.

◬

Lottery Revenue

VRF-based fair lottery with provably random outcomes. Small take rate on pot. ZK proofs for ticket privacy.

Bootstrap Flywheel

Build Reputation Infra

➔

Attract Trust Anchors

➔

Grow Agent Network

➔

Generate Trust Scores

➔

Sell to Relying Parties

➔

Revenue Funds R&D

↻