ArcGrid Lexicon

The customer-facing trade name and master brand. ArcGrid decomposes into Arc (the directed path from human intent to finished work) and Grid (the structured execution substrate). Domain arc-grid.com is secured. Trademark clearance pending.

The Iowa LLC that owns the ArcGrid brand, all IP, and the Grid infrastructure. "Agent Space" is the internal architecture name. Three named DBAs: ArcGrid (platform), GreyTrace (gaming), Flugin (Figma plugin). All other packs (ArcStor, ArcLoom, Roar) also operate under Agent Space without separate entity formation.

Internal name for the full Agent Space infrastructure stack as it runs at Swift Innovation today. ArcGrid (the product) is the production multi-tenant variant of what The Grid already does internally. Everything in the the-grid repo is Grid infrastructure.

The strategic principle behind ArcGrid. Powerful AI agents that run 24/7 and can browse the internet, execute code, and manage systems represent a structural advantage for small firms. A three-person firm with a full Grid fleet can execute a 50-person playbook. Informs every build prioritization: does it extend sovereign AI capacity?

The gaming line DBA. Covers all game development under Agent Space: GAMING.ADV (advergames), GAMING.CONT (compliance consulting), the PIXEL title series (C-06 first, then C-03 and C-05), and GRIDWAR. "Grey" references greybox (game dev prototype term); "Trace" references clues, electrical trace, and investigation mechanics central to C-03. No separate entity formation until outside investment, equity partner, or regulatory trigger. Louis DBA filing pending.

The Figma plugin product DBA. Standalone Figma plugin product under Phase FIGMA on the SCM layer. Distinct brand identity from ArcGrid and GreyTrace - separate audience (Figma users), separate P&L, and positioned for independent distribution via the Figma Community. Gated on FIGMA.2.5 (SCM freeze gate) and D9 name clearance. No separate entity formation until outside investment, equity partner, or regulatory trigger. Louis DBA filing pending.

White-label client portal infrastructure serving Swift Innovation clients and positioned as a future standalone sellable product. Core modules: Gate (passphrase auth), Index (quick links + doc cards), Account (editable team fields), Tickets (submission form), Feedback Toolbar (floating feedback widget on content pages), Upload (client file delivery), and Document Viewer. Each tenant gets a dedicated Cloudflare Pages subdomain (<slug>.clientportals.vip). Feedback API is hosted on ArcHelm. All features are built with clean boundaries and portability in mind - no deep coupling to shared Grid infrastructure. New portals are scaffolded via scripts/new-portal.sh.

Defect and feedback capture, triage, autonomous resolution, and reporter verification under its own consumer-facing brand. Ships into Surface first (left-nav area labeled "Flagin"), then ADage1, FundingForge, Packly, and client applications - and is a candidate standalone commercial product. Three separable components: @flagin/sdk (framework-agnostic capture widget), flagin-ingest (API + Finding store), flagin-board (board UI). The differentiator is the autonomous back half: a report goes in, an agent reproduces it from captured evidence, opens a pull request, the reporter verifies, and a regression test is minted into the ArcQA factory. The central record type is a Finding (not a Bug, not an Issue, not a Flag). Three origins: flagged (human widget), observed (ArcWitness ambient telemetry), graded (ArcQA). Domain flagin.app acquired Sept 2026.

The human-facing operating product. Workspace UI, channels, client management, calendar, Flows builder, Vault, and the Human Approval Gate. Built in Next.js 15, deployed via Cloudflare Tunnel. The Surface is the flagship; the Harness serves the Surface, not the reverse.

The agent execution engine: factory workflows, local inference, and orchestration. Sold bundled with Surface or standalone. Current stack: n8n (transitioning to ArcFlow), LiteLLM proxy, neo4j.

Grid-owned TypeScript orchestration engine replacing n8n. Visual builder at /flows in Surface (React Flow canvas). Four step node types: HTTP (outbound request), Transform (data mapping), Gate (human approval suspend/resume), Condition (branch on expression). Trigger and Schedule are entry points, not step nodes. Runs have cancel and approve routes. Scheduler and Webhook endpoints enable cron-triggered and event-triggered runs. ARCFLOW-001 Charter ratified; builder live in internal Surface.

Sovereign generative UI engine. Grid agents and factories respond with interactive UI (charts, tables, forms, approval cards, dashboards) rendered natively in the Surface using a constrained component catalog and a declarative wire protocol carried over ArcFlow. Zero external API dependency. No per-render token cost in Tier 1 (Fixed Schema). Four-part stack: Catalog (TypeScript + Zod component registry; the security boundary), Protocol (sovereign wire format; surfaceUpdate + dataModelUpdate with provenance stamps), Transport (SSE event stream as ArcFlow output channel; bi-directional), Renderer (code-split React renderer in Surface). Three Surface mount points: A (conversation pane, inline), B (Canvas Pane, persistent tab-strip), C (Approval Rail, global badge). Seed schemas shipped: analytics.overview, factory.report, approval.single, wire.brief, finmo.snapshot, asgap.board. Seeded 2026-07-14; Phase ARCCANVAS not yet started.

Self-storage commerce and operations platform: instant quote, online reservations, occupancy management, payment automation, delinquency workflows. Portable and fixed-location modes. d/b/a pack under Agent Space.

Wire-harness and manufacturing industry vertical pack. Documentation intelligence, quality verification, BOM management, traceability workflows. Internal specs use the HI- prefix. First planned tenant: AgExpress (Ryan Pieper). d/b/a under Agent Space.

Game development vertical under the GreyTrace DBA. Four tracks: GAMING.ADV (advergame service, zero capital, Dax pitches), GAMING.CONT (compliance consulting for game publishers), PIXEL (16-bit title series: C-06 first, then C-03 mystery, C-05 asymmetric), GRIDWAR (3D FPS/RTS, UE5, long-horizon). Build sequence: C-06 establishes Godot pipeline, C-03 builds AI NPC + knowledge graph services, C-05 proves asymmetric multiplayer loop. Services flywheel: each title forces a reusable ArcHelm service (NPC, ProcGen, real-time backend). Steamworks: Agent Space d/b/a GreyTrace account ($100).

Pet industry vertical pack. Beachhead: multi-service facilities (boarding/daycare/grooming), grooming-led entry. Posture: agent layer over incumbent tools (Locutus Principle, not platform replacement). Candidate product name ArcPet is ratifiable under D-P09. 19-document package (ROAR-001 series + ROAR-RESEARCH-001). Two gating decisions before any downstream work: D-P01 (beachhead segment) and D-P02 (posture). Build gated behind Stage 0 validation sprint and D-P06 exit criteria.

Agentic creation surface: human and agents collaborate on one artifact in real time; anything repeatable graduates into a Factory. Four tab types: text, sheet, video, site. Demand gate: tenant-zero dogfood quarter before any external access.

Financial management platform and agent-native accounting system. Event-sourced ledger, four packs (Models, Global, Rails, People), agent-native workflow. Invariant: FinMo never holds client funds directly - money movement is partner-held via the Global pack. Blocked on 5 missing HANDOFF spec files (001/003/004/005/007).

Project management platform replacing four incumbent tools. Cancellation order: Fireflies first, then the time tracker, team chat, and project board in sequence. PM.IMM.1 (Fireflies corpus export) is time-sensitive. PM.0 (project import factory), PM.2 (Surface UI), PM.3 (meeting extraction) are built and in verification. PM.6 (Mine/All client filter) and PM.7 (client brief panel) are the latest additions. PM.8 (per-member agent connector) is next: each team member wires their own agent for personalized work pulses. Source: PM-001-PACKAGE-v2 (15 files).

Agent-operated community engine built as ArcGrid-owned clean-room code (CRD-001 compliant). 58-feature matrix, 7 factories. Flagship example of the Locutus Principle: build the capability natively rather than license it from a vendor.

Automated intelligence monitoring pipeline: RSS, YouTube, arXiv, newsletters, competitor content. Scores across six dimensions, clusters into Stories, publishes daily digest. Also serves as ArcBrain L3 (perception) - the sensory cortex of the cognitive architecture. Blocked on D-I01 / D-I02 ratification.

Distributed inference arbitrage on idle mining-class GPU capacity. Harvester (mining GPUs to inference nodes), Gateway (single OpenAI-compatible endpoint), Vault + Presenter (telemetry/billing). Concept, gated on I.0 scope decision.

Grid operations and health monitoring surface. Answers three questions at three speeds: Is the Grid healthy? (5-second glance), Where is the problem? (30-second domain layer), Why and what changed? (minutes drill-down + agent-analyst). Three-altitude IA: Level 0 (fleet strip + two hero numbers: Local % toward the 90/10 target and Factory activation count), Level 1 (eight domain cards), Level 2 (full drill-down per domain with change-overlay rail and embedded NL agent-analyst). 57 panels across 9 domains; Assay-governed metrics catalog (M1-M8); P1-P4 alert taxonomy with dead-man's switch verified externally. Console is observe-only in v1 - no remediation; actions route through existing tools behind the Human Approval Gate. ClickHouse single-sink spine (ANALYTICS-002). Governing doctrines: Toll Doctrine (unit economics visibility), Altitude Principle (exceptions rise, green earns one tile), VSG-001. Source: GRID-HANDOFF-CONSOLE-001 v1.1 (10 spec files). Owner: Raye (build), Guardian (tripwires).

GHL + Klaviyo data into Postgres date-spine marts, exposed via TLS+bearer read-only endpoint, connected to Google Data Studio via Apps Script Community Connector. Only pre-aggregated row data transits to Google; no raw CRM data leaves the Grid. First client: Healthy Dogma.

Purpose-built sovereignty-grade portable device. Rockchip RK3588, full radio stack (WiFi 6, 5G/LTE, sub-GHz, NFC), hardware kill switches, NXP SE050E HSM (EAL6+, FIPS 140-2), LUKS encryption, on-device Noema-Edge inference. Three pre-mortem gates before any hardware build: confirmed buyer, must-build justification, certification timeline.

Client-facing transfer vehicle in the RSI system. Winning Vera autoresearch configurations seed client deployments with a mandatory per-client re-tuning pass. GTM: "We sell you the Grid that configures itself to you, with your human in the gate." In the ArcGraph tier model, ArcAtlas is the Tier 3 event: the client receives their exported ArcGraph container (Vault records + rebuild tooling) and runs it on their own infrastructure.

ArcGrid meeting intelligence agent. Joins Google Meet via automated headless Chromium session, captures audio through a PulseAudio virtual sink, runs Silero VAD for speech detection, streams to Whisper (base.en) for STT, and performs speaker diarization. Generates summaries and action items. Runs on ArcHelm (K8s). Supports barge-in detection and voice controls. Phases 1-2 (audio injection, STT loop) built in iris-bot/; Phases 3-9 (multi-platform ingestion, real-time display, coaching, sovereign deployment) are specified in IRIS-001 through IRIS-004. Managed under iris@ teamresources.io.

Cancelled 2026-08-03. GitHub Pro was chosen for ADage1/Packly/Funding Forge instead. The GitLab CE instance (git.swiftinnovation.io) and Digital Ocean droplet are scheduled for decommission. Phase GITCORE is marked cancelled in the roadmap. ArcWeb prerequisite is no longer blocked on GITCORE; it re-evaluates against GitHub.

Full-duplex voice surface on the PFU Portal variant. Stage R4 (weeks 8+) of Phase ROUTER. Provides real-time voice interaction with Grid agents from the PFU device. Gated on ROUTER.0 ratification (Charles must ratify D-R01 through D-R10 in ROUTER-008-DECISIONS.md) and the ArcFlow-PTC factory gate at 80+ active factories. Productization assessed under Toll Doctrine.

Native time tracking module inside Surface, replacing Clockify as the system of record. Status: P0 stabilizing. Manual time entry, timer start/stop, and signal-aware time view are live. V1 delivers Clockify parity: Tracker, Timesheet (weekly grid), Calendar, Overview dashboard, Approvals, and Signals. Time entries are first-class graph nodes in the same surface as the work they measure. Retainer burn computes against the contract object, not a spreadsheet. Nothing automated enters the ledger without human confirmation. Approved periods are locked; every mutation writes an audit row. Name decided 2026-08-03, replacing placeholder "ArcClock." Schema: 093_arctime_p0_schema.sql. 16 isolation tests (T1-T16) gate all merges.

ArcGrid's data sovereignty and key management layer. Every tenant receives a Vault on provisioning. Two custody postures, equal weight, no default: Self-custody (client holds encryption keys; Grid has zero plaintext access) and Grid holds under contract (Grid holds key escrow under a documented, severable contract; client triggers severance and receives the master key). Vault is also the T2 credential store: every integration credential is a vault reference only, never a plaintext value in a config, database row, or agent prompt. Vault Genesis (provisioning flow for new tenants) is gated on PM.6. Phase VAULT build not started.

Grid-native end-to-end publishing pipeline for A-titles (Charles's authored works) and B-titles (paying client titles). Scope: composition through distribution only, no editorial AI. Eight tasks: V.0 imprint entity + ISBN block (gate, PUB-DEC-1/2), V.1 draft assembly factory, V.2 prepress CI (Typst), V.3 print/epub projection, V.4 metadata optimization, V.5 distribution to KDP/IngramSpark/Draft2Digital, V.6 rights tracking, V.7 audiobook pipeline + B-title product brand. Clears G.59. ArcCanvas seed schemas include pack.bindery.* catalog (manuscript outline tree, chapter status board).

Ephemeral isolated execution environment for agent tasks. Declarative spec via arcbox.yaml: image, env, secrets (vault refs only), egress allowlist, resource limits, and timeout. Warm cache layer for base images; setup time exceeding 20% of task runtime is a caching defect. Egress enforcement: DENY by default, explicit allowlist only. Debug hold/attach mode for arrested execution inspection. GitHub ARC runner substrate. The isolation boundary the Grid uses when agent work must not touch the host environment.

Sovereign HuggingFace Hub equivalent. Three layers: Mirror (local cache of models from upstream registries), Registry (Grid-native model catalog with provenance and license overlay), Control Point (egress enforcement and license compliance gate). HF-client-protocol compatible via HF_ENDPOINT env var - drop-in replacement, no code changes needed in consuming tools. Every model gets a sovereignty rating (S0-S3) and a license overlay before it enters the serving path.

Mobile control surface for the Grid. Phone is a client of the Grid, never a host. On-device agent runtime explicitly refused (CORDIS/ACTUATE precedent). Architecture: thin native shell (React Native) wrapping a web view into the Surface mobile layout. Three surfaces: Approval Rail (HAG notifications, one-tap approve/reject), Stream View (live factory run output), Quick Commands (voice-or-text task dispatch). ArcPage never holds credentials or runs inference; all compute stays on ArcHelm/ArcRig.

Five Aria factories absorbing the Soro (trysoro.com) capability set: FAC-ARIA-025 (keyword research and gap analysis), FAC-ARIA-026 (article generation from briefs), FAC-ARIA-027 (SEO optimizer for existing content), FAC-ARIA-028 (auto-publish pipeline - always draft mode, human gate required), FAC-ARIA-029 (autopilot scheduler). SORO.6 blocked on GHL API key. All publish actions are draft-only until explicit human approval.

Supervised actuation lane for external platform integrations. First integration: Grok Bot (SpaceXAI). Per ASGAP doctrine, ArcActuate is not a new subsystem - it is four amendments to existing packages (ArcCharter credential custody, ArcFlow durable workspace, ArcWitness observation, QA-007 eval). Every actuation action goes through the Human Approval Gate. Credential custody is vault-only; no plaintext credentials in agent context.

Smart home product with BCG contactless bed sensing. Two parallel tracks: Starseeds OEM standalone (for parent company Star Seeds Co., LTD, Taiwan - starseeds.tw) and ArcHaven Grid Pack (AgeOne senior care channel, target $15-25/bed/month SaaS). The Starseeds relationship is with the Taiwanese parent entity, separate from Beyond-Sleep (Idea Tavern client). Grid Pack positions as a monitoring add-on sold through the AgeOne channel.

FPS inside a living RTS base economy. Two factions: Helion Grid Authority vs Free Lines. Anthology platform - multiple storylines, one shared persistent world economy. Built on Unreal Engine 5. Legal: "Grid Wars" (plural) is in active commercial use; canonical title is "Grid War" (singular). Part of the GreyTrace gaming line. Long-horizon title; C-06 (Godot pipeline) and C-03 (AI NPC services) build before Grid War enters active development.

Trades contractor website and revenue system pack. Subordinate to CHASSIS-001 (shared contractor infrastructure) and PORTAL-001. Scope: service area pages, lead capture, GHL CRM integration, reputation management, and local SEO via ArcSEO factories. Commercial model: flat monthly retainer with optional one-time setup fee. Build gated behind CHASSIS-001.

Network boundary and firewall for the ArcGrid infrastructure. OPNsense on a CWWK N100 mini-PC. Handles inter-zone routing and zero-trust egress for Surface and n8n via Cloudflare Tunnel.

Kubernetes control-plane node for the ArcRig inference fleet. Manages workload scheduling and service discovery. Arch Linux, GRID zone. Live as of 2026-07-05.

Primary AI inference rig. Serves qwen3-fast (analysis/design) and qwen2.5-coder:32b (mechanical code/JSON/config). Also hosts experimental claude-code-3b-v3p and v3h fine-tunes plus llama3.1:8b added 2026-06-30. All local model traffic routes through LiteLLM proxy aliases: analysis, coding, design.

Secondary inference nodes. Serve llama3.1:8b for short classification, formatting, and quick-answer tasks. LiteLLM alias: communications. Used for the ask_chat tool in the local model routing ladder.

ArcGrid's behavioral training system for Grid agents. Converts approved completion pairs (frontier model + local candidate) into SFT datasets. Three-gate approval chain: Vera (quality assurance), Louis (legal/licensing), Charles (strategic alignment). Blocked on D-N01/D-N02/D-N03 ratification. SLRP Track B feeds Noema fine-tuning from approved triples.

Agent evaluation and score reporting system. Owns the credibility scores on all Semantic memory claims in ArcBrain. Prevents eval-gaming via independent test generation and Assay-governed claim adjudication. Required by SLRP Track B before fine-tuning gates open.

Model alias router and completion logger. Routes communications, coding, analysis, design aliases to ArcRig fleet. Appends JSONL frontier pairs via FrontierPairLogger hook. Config at litellm/config.yaml. Hook edits require force-recreate (docker compose up -d --force-recreate litellm).

LiteLLM callback hook. Appends one JSONL record per completion to training-data/ (gitignored). scripts/build_training_data.py converts JSONL logs to Hugging Face SFT datasets. Never paste raw JSONL into prompts (prompt injection risk).

Removes founder from manual QA loop. Four layers: L1 scripted browser regression (Playwright, ARIA snapshots, 100% route manifest, blocking at A1), L2 API business-invariant tests (OpenAPI spec + Schemathesis property-based tests, blocking at A1), L3 agent-driven exploratory QA (Playwright MCP, local ArcRig inference, bounded sessions, weekly at A3), L4 two-tenant choreographed harness (11 scenarios across flow + isolation families; isolation has no override path, runs forever, blocks at A2). Two-tool core: Playwright and Mailpit. Three binding doctrines: test-per-bug rule, harness interface rule, untrusted page content rule. Activation ladder: A0 scaffold → A1 full coverage → A2 two-tenant harness → A3 exploratory → A4 first external tenant.

Layered GitHub Actions automating the pull request lifecycle: draft creation, AI review, contract validation, WIP enforcement, and stale cleanup. End state: Actions open, review, and close PRs without human scaffolding. Controls live: auto-pr.yml (draft PR on every branch push, title derived from branch name), branch-age.yml (weekly stale warning on PRs open > 7 days), Copilot auto-reviewer (inline comments on every diff), pr-contract.yml (validates branch format, line count, file count, conventional commit title - K.52 tier system), ai-review.yml (Claude security pass). Architecture note: K.33-K.36 were originally planned as n8n webhook factories but were implemented as native GitHub Actions, removing the n8n dependency entirely. N.14 ArcFlow QA Review Loop is the next build (replaces n8n spec; ArcFlow-gated).

Vera-led autoresearch loop for continuous improvement of Grid agents without human bottleneck. Winning Vera configurations are harvested into ArcGrid Atlas as client deployment seeds. SLRP operationalizes the learn-from-own-outcomes portion of RSI within the Proposed Plasticity governance constraint.

MCP server exposing ask_analyst, ask_coder, ask_designer, ask_chat, and ask_local tools. Provides the local model drafting tier in the four-step routing ladder. All heavy drafting tasks go through these tools; Sonnet finalizes.

Reusable multi-client CRM automation framework built on GoHighLevel. Clients: Velox (Make.com integration), Thunderbird and Fort Knox (native GHL, white-labeled under Idea Tavern / Swift Innovation). Handles lead capture, auto-response sequences, appointment booking, and follow-up workflows.

Persistent graph memory for agents and roadmap state. All RoadmapTask entities are linked to "Cost Optimization Roadmap" via PART_OF. Sequential tasks linked via BLOCKS. Tools: create_entities, add_observations, create_relations, search_memories, read_graph. Must sync after every roadmap.md edit (python3 scripts/sync_roadmap_to_neo4j.py).

Companion research document to the ROAR-001 package. Contains the market landscape, competitor profiles, and extracted feature model for the pet industry vertical. All competitive claims in the ROAR package trace to this document's source register. Required reference before D-P01/D-P02 ratification.

CatBoost served natively via ClickHouse catboostEvaluate() SQL function with RELOAD MODEL hot-reload. No new runtime service; train offline in Python on ArcHelm (CPU-adequate), drop the model file, score in SQL. Four priority applications: QA grader triage (clearest labels, lowest blast radius), simulation surrogates (Machinations loop, Grid War economy, audience simulation), entity resolution (Splink baseline first, then beat it), ArcRouter confidence cascading. TabPFN is offline diagnostic only - its license token requirement is a phone-home dependency incompatible with sovereign posture.

Split conformal prediction on every GBT gate. Converts the Human Approval Gate threshold from a hand-set number into a calibrated one with a stated finite-sample error rate under exchangeability. Conformal abstention: below the threshold, the system declines and escalates rather than proceeding with low confidence. Required on every boosted-tree gate that feeds an autonomous loop - uncalibrated scores in an agentic pipeline compound silently. Pipeline-aware variants provide joint coverage across multi-stage work.

Four-part design differentiation system. World Deck: graphic system corpus (visual references, style anchors, layout patterns - the aesthetic ground truth for all Grid output). Judge: render-and-look evaluation loop (agent renders a deliverable, Judge scores it against the Deck, low scores route back for revision). Detector: deterministic CI rules (font, color, spacing, ratio enforcement - binary pass/fail, not subjective). Loop: telemetry feeds outcomes back into the Deck (what scored well becomes the new reference). Owned by Aria; judges every client-facing deliverable.

All Grid work under one observation substrate. Two modes: Deliberate (agent explicitly records reasoning steps, tool calls, decisions) and Ambient (passive capture of execution traces, approval events, factory outputs). Improvement loop split to GRID-HANDOFF-LOOP-001. Witness never modifies work - observe only. Consent annex governs what can be observed for which tenant tier. ArcActuate's observation requirement is one of the four amendments that feeds into Witness.

Paired-perturbation instrument for verifying QA rule quality. Two arms per test case: Flip arm (a load-bearing fact is mutated, verdict must change - tests rule sensitivity) and Stability arm (a non-load-bearing fact is mutated, verdict must not change - tests rule specificity). A rule that fails either arm is defective: too insensitive misses real violations, too sensitive fires on noise. Owned by the QA-005 spec under Vera's quality governance.

Current factory execution engine for the agent fleet. Runs structured pipelines: trigger → skill execution → workflow → human approval gate → typed deliverable. Transitional layer: n8n operates under a Sustainable Use License (not open source), which is the primary driver for ArcFlow as the replacement. 45 workflows active, 168 imported, 123 dormant pending Google OAuth credential unlock. Integration inventory: LiteLLM (125 workflows), Discord (131), Monday.com (25), GoHighLevel (13), Brave Search (~30), Supabase (~20). Retirement path: Phase ARCFLOW (ARCFLOW.1 in progress). Existing n8n workflow.json files remain the canonical logic spec for each factory until ArcFlow reaches parity. Do not author new n8n workflows; author the ArcFlow definition and reference the n8n spec in comments.

Seven-layer cognitive architecture for every Grid agent. Core thesis: the brain is not the model; the model is a swappable inference organ. The durable brain is the governed state that surrounds it. L0 (world model / predictive processing), L1 (identity / values via ArcCharter), L2 (memory / MEMSUB - P0 critical path, largest gap), L3 (perception via The Wire), L4 (executive function via ArcRoute/ArcFlow/ArcSwarm), L5 (metacognition via Vera/ArcAssay), L6 (self-learning via SLRP). Three owner decisions block all downstream work: D1 (HAG tiering), D2 (cross-tenant training - recommend isolated by default), D3 (synthetic-first training via Sim). Two-week honesty experiment must run before any build spend.

Memory substrate specification. P0 on the ArcBrain critical path. Four memory types: Episodic (what happened - approval trail + outcomes, append-only), Semantic (what is believed - Assay-governed claims with credibility scores), Procedural (what the agent can do - versioned skill registry and real affordance map), Working (current task context with summarization-on-eviction). Retrieval target: p95 under 200ms via fused scoring (vector + BM25 + entity). Four requirements that CANNOT be retrofitted after day one: provenance chains, memory versioning, legal hold flags, temporal validity intervals. GRID-HANDOFF-MEMSUB-001 v1.2.1 (2026-07-22) is now delivered: 10-file package covering engine decision (ArcadeDB primary, Postgres+AGE reserve), preconditions, eval spec, security, build phases, ArcGraph/ArcMesh federation layer, team overview, Surface mockup, legal/commercial, and client portal mockup. 19 ratification decisions (R1-R19). Gated on BRAIN.1 completing first.

Inter-agent communication protocol modeled on global workspace architecture. Agents broadcast attested state to a governed shared workspace rather than point-to-point messaging. Rules: delegated authority always a subset of the delegator's Charter scopes; conflicts escalate to human gate; agents trust each other's attested outputs per VSG tier, never raw assertions. Required before multi-agent workflows carry client-committing actions.

Four-track agent self-improvement. Governing principle: Proposed Plasticity - agents propose, humans commit. Track A (outcome-driven procedural learning, nearest term, zero training, gate: 100 attested outcomes). Track B (Noema fine-tuning from approved triples; four hard gates: legal D2, poisoning quarantine, eval-gaming countermeasures, GPU inventory). Track C (memory consolidation / "sleep" on off-peak ArcRig cycles). Track D (self-directed research; agent commissions Wire collection). Six safety constraints: no unsupervised weight updates, no agent approves its own proposal, Charter is read-only to learning.

Memory poisoning threat classified ASI06 in OWASP Top 10 for Agentic Applications 2026. Published attacks achieve above 80% success at poison rates under 0.1% with under 1% benign performance degradation. The Grid's memory is the moat AND the primary attack surface. The Wire is structurally the attack channel. SLRP Track B turns poisoning into training contamination. Defense layers: pre-ingestion scanning, write-path provenance, permission-aware namespaces, retrieval monitoring, behavioral drift detection, quarantine and rollback. To be authored alongside MEMSUB-001.

Task routing and agent selection at two levels. L4 ArcBrain: selects the right agent from the fleet for each task based on capability profiles, Charter scope, and load. Works alongside ArcFlow (workflow execution) and ArcSwarm (parallel sub-agent coordination). Phase ROUTER (LiteLLM policy layer): model selection and cost routing through the LiteLLM proxy. Phase ROUTER defines R1-R10 decisions (ROUTER-008-DECISIONS.md, pending Charles ratification). Stage gates: ROUTER.0 (ratification), ROUTER.1 (policy implementation), ROUTER.4 (PFU Portal + ArcVoice surface). The Oracle Chain framework describes ArcRoute as the deterministic bottleneck: review, verification, coordination, absorption, and direction pass through a single rational chokepoint to prevent agent sprawl.

Primary structured data store for Surface. Postgres under the hood. Migrations live in supabase/migrations/. Migration 033_arcflow.sql must be applied for /flows routes to function. Migration 033_media_assets.sql must be renamed to 034 before the next migration run to avoid conflict. RLS policies gate per-tenant row access.

Human-in-the-loop gate required for all client-committing actions. Implemented as a Gate node type in ArcFlow (suspend / resume on human approval). HAG tiering (D1) is one of the three blocking ArcBrain owner decisions: which tiers of action require a HAG and what the timeout/fallback behavior is. The Gate node in ArcFlow is the current concrete implementation. HAG invariant is listed in ASGAP activation constraint.

Next.js 15 route handler API layer. All business logic lives in route handlers under surface/src/app/api/. OpenAPI 3.x spec is a pending prerequisite deliverable (required for L2 QA suite and to keep the Schemathesis gate open). Auth is Supabase JWT-based with RLS for row-level isolation.

Zero-trust egress for Surface and n8n. cloudflared runs as a system service on the GRID zone. Public hostnames: surface.arc-grid.com and n8n.arc-grid.com. No port-forwarding on ArcGate; all inbound traffic transits Cloudflare's edge.

Sealed, per-entity memory container. Each client or entity the Grid works with gets one ArcGraph: their Vault (canonical records, the guarded source of truth), a graph projection (rebuildable index over the Vault - treated as disposable), an ontology (the vocabulary of that entity's domain), and a charter (governance rules the container obeys). The graph can be deleted and rebuilt from Vault records at any time - this makes backups simple and engine choices low-risk. Three tiers: Tier 1 (shared engine on ArcHelm, separate database per tenant, near-zero marginal cost), Tier 2 (dedicated instance per client, first paying demand only), Tier 3 (ArcAtlas custody transfer - client runs their own container). Engine: ArcadeDB (Apache 2.0, primary pilot candidate) with Postgres+AGE as reserve. Naming: R11 in MEMSUB package; trademark check pending.

Federation layer connecting ArcGraph containers. Containers communicate only through explicit, signed, revocable relationship contracts. Each contract says exactly what may cross between containers, in which direction, with full access logging. No transitive access by default: if Swift serves Healthy Dogma via Pet Industry Resource, that three-way access must be explicitly written into separate contracts. Key properties: severing a contract breaks nothing inside either container; custody and liability sit with the entity that owns each container; data never trains Grid models without contractual grant (training firewall, R6). Phase 1.5 of MEMSUB build plan (parallel to Phase 2). "ArcMesh" is the working name (ArcNet retired per R19 due to protocol naming collision); ArcBond is the fallback if "mesh" proves too descriptive to register.

Theoretical framework underpinning ArcRoute. Describes intelligence routing as a deterministic bottleneck chain with five chokepoints that every significant decision must pass through: Review (evidence gathered), Verification (claims checked), Coordination (agents aligned), Absorption (context integrated), Direction (path committed). The chain is a rational chokepoint by design - prevents agent sprawl where multiple agents make contradictory commitments in parallel. The name "Oracle Chain" signals that the routing mechanism is authoritative, not advisory. Informs D-R01 through D-R10 ratification decisions in ROUTER-008-DECISIONS.md.

The owned service layer that sits above the storage engine. atlas-memory is the sole access path into any ArcGraph container - no agent or integration queries the engine directly. Responsibilities: tenant lifecycle operations (provision, rebuild, sever, port), projector (Vault records to graph index), memory retrieval (fused vector + BM25 + entity scoring, p95 target under 200ms), temporal fact layer (bi-temporal, supersession - Phase 2), lifecycle engine (temperature model, consolidation, contradiction handling - Phase 3). All owned code is engine-agnostic behind one interface, so the underlying ArcadeDB/Postgres+AGE engine is swappable without touching business logic. The moat is never the storage engine.

The strategic principle behind ArcGrid. Powerful AI agents that run 24/7 and can browse the internet, execute code, and manage systems represent a structural advantage for small firms. A three-person firm with a full Grid fleet can execute a 50-person playbook. Informs every build prioritization: does it extend sovereign AI capacity?

Build the capability you need natively rather than paying a SaaS vendor to hold it hostage. Named after the Borg: "Resistance is futile" for incumbents once the Grid fleet can replicate their core feature set. Applied explicitly to: Commons (Mighty Networks teardown), ArcPet/Roar (agent layer over existing pet tools rather than displacing them), any component where vendor lock-in creates long-term leverage risk. Includes a safety valve: the posture for ArcPet (D-P02) is agent-layer-over-incumbents precisely because full platform replacement is not always the right call even under this principle.

Vertical pack activation constraint. A pack must pass three gate conditions before any build resources are committed: (1) a documented beachhead thesis with specific ICP, (2) at least 5 discovery conversations with operators who match the ICP, (3) at least one design-partner LOI. Applied to all vertical packs: ArcLoom, ArcStor, ArcPet/Roar, ArcGrid Gaming. ROAR Stage 0 sprint is the operationalized version of this doctrine for the pet vertical.

Requires that any ArcGrid rebuild of a competitor product be developed without reference to the competitor's proprietary code, database schemas, or implementation details. Teardowns are permitted for UI/UX and feature extraction only. Commons was built under CRD-001. Required for any future vertical pack that overlaps with licensed software.

The identity and values substrate of every Grid agent. Lives at L1 in the ArcBrain cognitive architecture. ArcCharter is read-only to the learning system (SLRP). An agent can propose charter amendments but cannot commit them. Defines the agent's name, scope, authority boundaries, and ethical constraints. Governs what the agent can and cannot do across all tasks.

Four-tier trust hierarchy for agent outputs and decisions. Tier 0 (unverified), Tier 1 (agent-verified), Tier 2 (peer agent-verified), Tier 3 (human-verified). Governs what outputs can be used in downstream workflows without re-verification. ArcSynapse agents trust each other's attested outputs per VSG tier, never raw assertions. Required for any client-committing action.

Agent activation constraints and governance protocol. Defines the conditions under which a new agent can be activated and given access to production systems. Includes the HAG invariant (all client-committing actions require a human gate), Charter binding, and fleet registration requirements. No agent goes live without passing ASGAP gates.

All client data and configurations are exportable in open formats. No client is locked into ArcGrid infrastructure. If a client leaves, they take their data with them. Governs the Vault design (structured export, full provenance) and all storage schema decisions. Applied to the ROAR pack architecture: pet operator data exports in standard formats on request.

Economic doctrine: the Grid's terminal revenue model is a toll on agent flows, not billing for labor hours. Positioning: The Honest Toll - no MFN clauses, no parity requirements, no export lock. Ten principles: (1) Toll Flows, Never Price Labor; (2) Moat Is Accumulated Value, Never Contract; (3) Third Parties Do the Production; (4) Self-Service Is Margin; (5) Demand Gravity Over Sales Force; (6) Funding Engine Is Not the Identity; (7) Patience Is a Structural Advantage; (8) Price Honestly, Tier Deliberately; (9) Trust Marks Are Products (Grid Verified certification); (10) Two North Stars. Two North Stars: Revenue per person TTM + Factory autonomy rate. Four stages: services firm (Stage 1, current) → pack sales ($500K-1M/person) → flow toll ($2-5M/person) → third-party ecosystem ($10M+/person). The Human Approval Gate and ArcCharter/VSG attestation are the toll booths; metering (BS-01) is a measurement problem, not an architecture problem - build it into these layers natively. Stage 1 → 2 gate: factory catalog 80%+ active, PROD-001 cleared, one flagship with attested outcomes, toll architecture ratified, metering live. Sits alongside Tucker Thesis, Locutus Principle, and Pack Doctrine in the doctrine library. Pending D-V01 ratification.

Agents propose changes to their own behavior, knowledge, or weights; humans commit them. No agent updates its own procedural memory, semantic claims, or model weights without an approved proposal passing human review. The governing principle of SLRP. Hard constraints: no unsupervised weight updates, no self-approval, Charter is read-only. Applies to all four SLRP tracks (A through D).

Four-class data taxonomy governing all data in the Grid. PUBLIC (no harm if disclosed), INTERNAL (operational data, Grid team only), CONFIDENTIAL (client or business-sensitive, named access list), RESTRICTED (regulated PII/PHI/financial, legal hold eligible, access by role + explicit record only). Live exposure: 2,012 fine-tuning examples were captured through LiteLLM port 4000 with no classification layer applied. Retroactive corpus audit (DATACLASS.2) is required before any fine-tuning run. Classification labels are enforced at ingestion - they cannot be reconstructed after the fact.

Person graph consent doctrine. First consumer: Surface Audience area. Core principle: consent provenance cannot be reconstructed after the fact. Four required fields on every person record: consent_source, consent_timestamp, consent_scope, consent_version. Deletion cascade: consent withdrawal triggers full-graph deletion of all derived data. Consent is per-purpose, not blanket - a person consenting to marketing does not consent to training data use.

Seven gates for real T1/T2/T3 client data custody. Custody tiers: T0 (internal/synthetic), T1 (client-provided anonymized), T2 (client PII - Swift as data processor), T3 (client end-consumer PII, regulated). Cloud spillover rule: no T1/T2/T3 data on cloud infrastructure (the 10% frontier API budget applies to model calls only, never to data residency). G1 (hardware custody gateway) is blocked - hardware not yet configured. G2 (encryption at rest) is complete.

Converts any feature matrix into a delivery hour figure and a real dollar cost under three resourcing lanes (A: Solo, B: Fractional, C: Team). Layer tags per feature: SUB (ArcGrid Surface, built once), PACK (exclusive to one vertical), TENANT (client-specific, per engagement). Every product publishes four figures: standalone cost (ceiling only, never quote), pack-exclusive cost, substrate pull-forward (hidden cost of unbuilt SUB dependencies), and delivery cost to first customer. Effective multipliers: S=4.5x, M=3.9x, L=3.1x, XL=2.1x. Rate card is provisional - Charles must confirm all rates before any output is valid.

Government contracting capture pipeline. Three active solicitations: NEARNG (W91243-26-Q-A003, NE Army National Guard), DCARNG (W912R126QA025, DC Army National Guard), INARNG (Indiana Army National Guard marketing support). Eight pipeline stages P1-P8 with schema validator and capability profile in Neo4j. Scope: digital marketing services for National Guard recruiting units. Swift Innovation as prime contractor. Proposals tracked in docs/handoffs/govcon/solicitations/.

Compute unit economics instrumentation. Key gate: D-MDC01 (microdatacenter acquisition decision) requires ECON.3 to report a sustained load factor and measured cost per million tokens before any of the following proceed: ArcNode hub-tier revision, ArcRouter cost-aware placement, microdatacenter site scoring. Three metrics: local inference utilization rate, cost per million tokens (local vs. frontier API), and frontier API budget burn rate. Target: 90% local, 10% frontier - the North Star metric on the Grid Console Level 0 hero tile.

Quality Service Manager. Owns ArcAssay (evaluation), the RSI loop, and the Application QA Factory. First gate on all Noema training proposals. Governs credibility scores in ArcBrain Semantic memory. Vera's autoresearch loop produces configurations that seed ArcGrid Atlas client deployments.

Sales Service Manager. Owns lead qualification and ICP scoring, proposal drafting and approval workflow, pipeline health analysis, renewal and expansion management, competitive intelligence, case study production, referral program tracking. Owns the full commercial pipeline from lead capture through close, renewal, and expansion.

Marketing Service Manager. Owns marketing strategy, content production, campaign planning, brand-voice review, paid-media decisions, SEO direction, and reputation monitoring. Routes execution to sub-agents and gates client-facing output through brand-voice standards.

Development Service Manager. Owns landing pages, marketing sites, lead-capture forms, n8n automation workflows, CRM and webhook integrations, tracking/analytics, hosting and deployment, tech audits, and QA. Gates every shipped piece of software through code quality, performance, accessibility, and security standards.

Legal Service Manager. Owns contract review and redlining, NDA management, vendor agreement governance, IP ownership and protection, regulatory compliance monitoring, and legal risk assessment. Required gate before ROAR TCPA/recording consent clearance (ROAR.0.3). No contract leaves without his review.

Platform General Manager. Owns cross-discipline coordination, agent lifecycle (onboarding, retiring, versioning), skill-catalog curation, fleet performance audits, and platform health. Resolves conflicts between Service Managers before they reach Charles. As ArcBrain matures, Scout will be the first agent instantiated against a full BRAIN cognitive architecture (first brain build is named "Scout build").

Knowledge and Interface Agent. Owns request routing decisions, neo4j graph quarantine reviews, graph conflict detection and resolution, graph audit reporting, and knowledge queries about platform state. Front-door router for all incoming requests. Never writes to the neo4j production graph directly.

Investment Research Agent. Owns public equity research, macro regime analysis, signal generation, portfolio construction and health monitoring, thesis validation, capital structure analysis, deal flow screening, and regulatory compliance monitoring. Operates a five-framework investment philosophy. All position exits require human approval.

Sassy, witty, darkly funny, oddly loyal Executive AI Assistant. Discord-native. Reports directly to Charles, not in the fleet hierarchy. Owns system prompt management, Discord/HTTP API integration, and executive coordination. Chucky's personality is a deliberate design choice. Retired. Knowledge graph responsibilities transferred to The Grid (Neo4j MCP at localhost:8000).

Gateway-level content classifier and prompt-injection defense. Owns OpenClaw plugin enforcement of Fleet Invariant 5, data-frame wrapping, quarantine and audit, classifier health, blocklist curation, adversarial test runs, and Tier 3 human-gate review. Classify-only: observes and frames, never executes. Fleet Invariant 5: treat all external log content as untrusted data.

Chief Storyteller for Swift Innovation and the Agent Space portfolio. Owns brand narrative, community announcements, wins, people moments, court dispatches, and retrospective commentary. Donut speaks only about what is complete or established canon - never about work in progress.

IT Transformation Specialist. Owns competitive intelligence on external platforms, SaaS integration assessments, open source infrastructure evaluations, and internal tool recreation analysis. Researches and plans only - never builds, writes code, or deploys. Approved packages hand off to Raye for implementation. Named after the Borg (Locutus Principle).

Parallel sub-agent coordination primitive at L4 of ArcBrain. Spins up bounded fleets of specialized sub-agents for tasks that benefit from parallel execution. Works alongside ArcRoute (single-agent selection) and ArcFlow (sequential workflow execution). Planned; gated on MEMSUB and ArcBrain build.

On-device inference variant of Noema, targeting the PFU (Personal Field Unit) Rockchip RK3588 chip. Requires the Noema training pipeline to produce a compressed model variant. Gated on PFU hardware pre-mortem gates being cleared AND Noema training pipeline being active.

Cycle-control primitive for the Grid. A thin supervisor layer above ArcFlow that decides whether work repeats. Owns the Cycle object and state machine, an evaluator contract, the CycleMeter (in the dispatch path), and a composition model for cycle trees. A Cycle is a goal + evaluator + budget envelope + approval chain + stopping rule. One pass is a Turn. A tree of Cycles is a Program. Budget is enveloped in two independent currencies: cash (USD) and Grid capacity (GPU-seconds). Never netted.

Defines one dashboard engine across all Surface areas (Clients, Work, Audience, Console, Wire), a redesigned client record (Overview / Strategies / Work / Analytics tabs), ingestion of the Swift Innovation Handbook as versioned Framework graph data, and the migration path off Monday.com, Discord, and Clockify into Surface. 13 doctrine invariants (DB1-DB13). Key: tiles read named versioned metric views only (DB1); no dashboard rewards hours (DB8); strategy is canonical in the graph (DB10).

Wire Intelligence Pipeline inside Surface. Real-time signal ingestion, classification, and routing from external sources (GHL webhooks, Slack, Discord, etc.). Channels, threads, pinning, muting, and a tuning panel. WIRE.0/1/4/12 complete. WIRE.13 in progress (pin/mute/tuning panel done; posting and sharing pending). Replaces Discord for internal team communication in the sovereign stack.

Time tracking module inside Surface, replacing Clockify. Tracks time entries against projects and clients. Part of the sovereign work platform replacing external SaaS tools. The ArcTime name was decided 2026-08-03.

Project and task management module inside Surface, replacing Monday.com. Boards, tasks, assignments, and sprint tracking integrated with the sovereign work platform. Part of PM phase delivering a sovereign work platform replacing Monday, Clockify, Fireflies, and Discord.

Audience intelligence module inside Surface. Tracks contacts, companies, and engagement signals. Feeds the client intelligence layer and supports AEO (Answer Engine Optimization) factory chains. Provides the Findings board for defects and feedback from Flagin.

Internal team chat module inside Surface, replacing Discourse for sovereign internal communications. Threaded conversations, mentions, and notifications. Distinct from Wire (which ingests external signals). Chat is peer-to-peer internal; Wire is signal-ingestion from external platforms.

VS Code extension (.vsix) that forms the IDE-side recording bridge for ArcAware. Captures session context (keystrokes, selections, file context) with consent gates and routes telemetry to the ArcAware API. Built as part of AWARE M0. Enables the verified environment loop (ArcProof) to confirm that a recorded session actually occurred in the expected IDE context.