ArcGraft — GRID-HANDOFF-GRAFT-001 Rev A
ArcGraft Composition Matrix
How the Grid builds, combines, and owns AI competence. Each department model is a home base plus QLoRA adapters, served through ArcRouter, traceable through ArcLedger. 5 phases from a zero-cost first adapter to a Noema MoE. Doctrine: ArcTide, the Reversible Mind.
5
Phases
25
Techniques
15
Packages amended
14
Decisions
7
ArcTide invariants
8
Frontier bets
Phase 0 — Foundation (2–4 weeks, zero cost)
| Task | Scope | Status | Gate |
|---|---|---|---|
| P0-01 | Home base bake-off Finish the Noema bake-off with GRAFT criteria: commercial license clear for derivative training; size family sharing one tokenizer; strong tool calling at 4-bit; healthy adapter/merge ecosystem; fits Rig 3 with KV cache room. Ratify D-GR-01. | Blocked | ESCALATE: D-GR-01 |
| P0-02 | Capture verification Confirm ArcRouter port 4000 capture is intact and DATACLASS M0 labels exist for the corpus slice. | Next | DATACLASS M0 |
| P0-03 | llama-server multi-LoRA on Rig 3 Stand up llama-server with multi-LoRA as a new ArcRouter destination. Per D-GR-06 (proposed: llama-server now, vLLM at next hardware tier). | Next | Unblocked |
| P0-04 | Department model sort worksheet List every model each department uses today; classify as adapter candidate / sense / tool / frontier escalation. | Next | Unblocked |
| P0-05 | First QLoRA adapter Train first QLoRA adapter (r=16) on one task class. 2–6 hours on Rig 3 RTX 3070s. | Future | P0-01 + P0-02 |
| P0-06 | Capacity benchmark Measure tokens/sec, concurrency, memory on Rig 3 and a GTX 1660 for the home base at 4-bit. Replaces estimates with real numbers. | Future | P0-03 |
| P0-07 | ArcHub lineage schema for adapters Extend lineage record: base hash, data ref, DATACLASS, recipe, eval, owner, license. Write first records. | Future | P0-01 |
| P0-08 | Structured decoding Enable structured/constrained decoding for every local route that feeds a Surface object (R-23). Nearly free — fixes most format failures. | Future | P0-03 |
| P0-09 | QA-006 evaluation — exit gate Run QA-006 on adapter route vs current route. Adapter must match or beat current local route on its task class at equal or lower cost. Lineage and ledger receipts must exist. | Future | P0-05 |
Phase 1 — First Attachments (1–2 months)
| Task | Scope | Status | Gate |
|---|---|---|---|
| P1-01 | Department adapter v1 First full department on home base + adapters. Proposed: Marketing (highest volume, clearest accept/reject history). | Future | P0 exit + D-GR-02 |
| P1-02 | Skill adapters 3–5 skill adapters mapped to that department's Curate tags; hot-swapped per request. | Future | P1-01 |
| P1-03 | Client brand-voice adapters Brand-voice adapters for 2–3 clients. Ownership: client owns adapter files and lineage; Swift retains base, cores, pipeline, and cross-client learnings. Contract terms per A-11. | Future | P0 exit + D-GR-03 + legal |
| P1-04 | ArcRouter adapter-aware routing Route tuple expands: (base, adapter stack with strengths, steers, server, tier). Curate tags resolve to adapter stacks. Adapter IDs written to ArcLedger metadata. | Future | P0 exit |
| P1-09 | ArcTrace training events ArcTrace emits: tenant, DATACLASS, task class, department, route, input context ref, output, human edits (diff), final accepted version, outcome, reason. This data is the moat (FB-01). | Future | D-GR-13 (ESCALATE) |
| P1-10 | Proof point report Cost per task and quality vs old multi-model setup for the first department. Investor-grade evidence. | Future | P1-01 |
| P1-11 | Technique radar Monthly intake of new methods (file 08 §8.4): Wire feed watching arXiv cs.CL/cs.LG, lab blogs, HF trending. Agent scores on Value/Cost/Fit/Phase. Winners get time-boxed experiments. | Next | Unblocked |
| P1-12 | ArcTide Experiment E-1 (round trip) Write client style rules as graph nodes, train adapter, probe candidate rules back, compare with source. First test of the reversible-mind hypothesis. | Future | P1-03 |
Phases 2–4 summary
| Phase | Goal | Key new capability | Exit gate |
|---|---|---|---|
| P2 | Merging and Judgment (2–4 mo) Specialist fine-tunes per sub-skill; TIES/DARE merge into department cores; evolutionary merge search; DPO on ArcTrace approve/reject pairs; Noema Check v1; learned routing in shadow mode; all departments migrated. | Merged cores + DPO judgment | Merged cores beat adapters on QA-006; DPO measurably reduces revisions |
| P3 | Absorption (4–8 mo) Governed teacher roster with license clearance; FuseChat-style cross-family pipeline; vision projector for Design/Dev; RL with verifiable rewards; home-base upgrade drill (portability proof). | Cross-family fusion + vision | Fused model beats unfused home base; base-swap rehearsed end to end |
| P4 | Consolidation (8+ mo) Branch-Train-MiX: departments as experts in one Noema MoE; Noema Pocket (1–4B distill for Rigs 1+2 and edge); ArcSwarm packaging; ArcTide Experiments E-3 and E-4. | Noema MoE + Pocket | MoE vs routed cores decision; Pocket shipped |
Technique Radar — R-01 to R-25
| ID | Technique | Value | Cost | When | Notes |
|---|---|---|---|---|---|
| R-01 | Quantization (GGUF, AWQ, GPTQ; 4-8 bit) | 5 | 1 | Now | Phase 0. GTX 1660: ≤4B comfortable, 7-9B tight. RTX 3070: ≤9B comfortable, 14B borderline. |
| R-02 | Multi-LoRA serving | 5 | 2 | Now | Phase 0. llama-server on Rig 3. 500 adapters stored, served copies scale with traffic. |
| R-03 | QLoRA training | 5 | 2 | Now | Phase 0-1. Rig 3 for ≤9B (2-6h). GTX 1660 marginal for 7-9B with Unsloth. |
| R-04 | Speculative decoding | 4 | 2 | Now | Phase 1. Same-family small draft. Pocket serves double duty. |
| R-05 | Prompt/prefix caching | 4 | 1 | Now | Phase 1. System prompt and brand context. Cheap latency and cost reduction. |
| R-06 | Steering vectors | 3 | 1 | Now | Phase 1. Tone/trait dials in Surface. Coarse but near-zero cost. |
| R-07 | Embedding and reranker fine-tuning | 5 | 2 | Now | Phase 1. Tiny models, train in minutes on any GPU. Measure with QA-004. |
| R-08 | Merging (TIES, DARE, SLERP, soups) | 4 | 1 | Now | Phase 2. CPU merge on ArcHelm. 32 GB RAM handles models up to ~7B. mergekit initially. |
| R-09 | DPO / preference tuning | 5 | 3 | Now | Phase 2. Small models on current rigs. ArcTrace accept/reject pairs are the data source. |
| R-10 | Learned routing (RouteLLM-style) | 4 | 2 | Now | Phase 2. Shadow mode behind ArcRouter Selector per DR-3. Reduces frontier escalations. |
| R-11 | Evolutionary merge search | 4 | 3 | Now | Phase 2. Overnight on ArcHelm + Rigs. QA-006 as fitness function. |
| R-12 | Test-time compute + verifier | 4 | 2 | Now | Phase 2. Noema Check as verifier in the loop. |
| R-13 | Synthetic data pipelines | 4 | 3 | Now | Phase 2. Augment sparse task classes. |
| R-14 | RL with verifiable rewards | 5 | 4 | Next tier | Phase 3. Tests-pass for Dev; schema/process checks for Ops. Needs ArcProof. |
| R-15 | Knowledge fusion (FuseLLM/FuseChat) | 5 | 4 | Next tier | Phase 3. Cross-family absorption. Requires teacher roster with license clearance (D-GR-10). |
| R-16 | Vision projector | 4 | 3 | Next tier | Phase 3. Noema sees screenshots, mockups, PDFs, brand boards. Design + Dev departments. |
| R-17 | Prune + distill (Minitron-style) | 4 | 4 | Next tier | Phase 4. Noema Pocket 1–4B from department cores. Rigs 1+2 and edge use. |
| R-18 | Branch-Train-MiX / upcycling | 4 | 5 | Next tier | Phase 4. Departments as MoE experts. Requires next hardware tier. |
| R-19 | Sparse autoencoders (SAE) / interpretability | 5 | 4 | Now | ArcTide research lane from Phase 1. Strongest on small models. Supports tide-out (surfacing). |
| R-20 | Machine unlearning | 4 | 4 | Later | Research lane. TIDE-I5 retraction. Immature; use steering/SAE suppression first. |
| R-21 | Continual learning (avoiding forgetting) | 4 | 3 | Now | Phase 2 onward. Retrain cadence, replay sets, rollback rules. Adapters make this cheaper. |
| R-22 | Long-context / KV-cache compression | 3 | 2 | Now | Phase 1. KV cache grows with context and concurrent users. Important at scale. |
| R-23 | Structured / constrained decoding | 5 | 1 | Now | Phase 0. Nearly free. Fixes most Surface object format failures immediately. |
| R-24 | Federated fine-tuning | 3 | 4 | Later | Research lane. Relevant for ArcNode on-prem client deployments. |
| R-25 | Watermarking / fingerprinting | 3 | 3 | Now | Phase 3. Client and Swift adapter provenance. AuthenLoRA / dual-watermark approach. |
ArcTide Doctrine — GRID-DOCTRINE-TIDE-001
Every piece of competence in the Grid exists in up to three states and can move between them under a ledger. Explicit (ArcGraph: rules, rubrics, exemplars, facts). Tacit (adapters and cores in ArcHub: fast, fluent, opaque). Title (ArcLedger receipt chain: provenance, contribution, ownership). Ratification pending D-GR-11.
TIDE-I1 — No orphan weights
Every tacit artifact links to its explicit sources and trace data. An adapter without lineage cannot be served.
TIDE-I2 — Graph is canonical for facts
Weights may hold skill; they are never the only holder of a fact. Facts used at answer time come from ArcGraph and records.
TIDE-I3 — Every crossing receipted
Consolidation, surfacing, retraction, transfer, and export each write ArcLedger receipts. No silent transitions.
TIDE-I4 — Surfaced knowledge is a candidate
Anything extracted from weights enters ArcGraph as a candidate and passes the Human Approval Gate before becoming canonical.
TIDE-I5 — Retraction propagates
Removing or correcting a graph node marks every dependent artifact stale. Interim: suppress at answer time. Final: retrain from corrected sources.
TIDE-I6 — Family neutrality
Every production skill must pass a portability test: recompiled from its explicit sources into a second model family, it reaches an agreed share of its original eval score.
TIDE-I7 — Title follows contribution
Receipts attribute contributions (client data, Swift process, teacher models, human reviewers) so ownership and licensing can be computed, not argued.
Rig capacity quick-reference
What fits where (4-bit quantization)
- GTX 1660 (6 GB, Rigs 1+2, 24 cards): comfortable ≤4B; tight 7-9B (short context, single user, Unsloth); no 12B+. No bf16 (fp16 only, slower training). Embeddings, rerankers, evals, speculative drafts, Pocket inference.
- RTX 3070 (8 GB, Rig 3, 5 cards): comfortable ≤9B; borderline 14B with Unsloth; no 30B+. QLoRA training (2-6h for 7B adapter). Multi-LoRA serving home base. Merge evaluation runs.
- ArcHelm (CPU, 32 GB RAM, Rig 3 node): CPU merge jobs (any method ≤7B fits in RAM). Pipeline orchestration. Evolutionary search coordinator.
- ArcNode (planned): 14B adapter training; local dev inference; leaf routing capture sink (D-AN-08 upgraded to required).
- Cloud burst (D-GR-05): fusion training runs, full fine-tunes, MoE serving, large teachers. Per-run spend cap required.
- D-GR-04 (ESCALATE): one 16-24 GB GPU unlocks 14B adapters comfortably, real concurrency, and vLLM. Timing aligned with ArcNode M0 pilot.
Frontier bets
| ID | Bet | First experiment | Kill criterion |
|---|---|---|---|
| FB-01 | Process-trace training Train on how agency work actually happens (drafts, edits, rejections, approvals) not just finished text. The moat is the data, which only Surface captures. Start immediately. | DPO adapter from ArcTrace accept/reject pairs on one task class (P2-04) | No measurable reduction in revisions after two training cycles |
| FB-02 | Self-compounding client adapters Each client's adapter retrains automatically from their approved work with eval gates and one-click rollback. | Weekly automatic retrain for one client adapter in shadow mode | Quality flat or declining over 6 cycles, or forgetting incidents |
| FB-03 | Governed merge-on-demand Compose a model per request from an adapter library with policy controls on which adapters may combine (tenant, DATACLASS, license). | Router-selected adapter blends for one department with a policy table | Blends never beat the best single stack |
| FB-04 | Business-outcome evolution Evolutionary merging scored on business results (approval rate, revisions avoided, client KPIs) instead of benchmarks. | One evolution run using QA-006 net value as fitness | Fitness signal too noisy to rank candidates |
| FB-05 | Verifiable lineage and title Receipt chains for every dataset, teacher, merge and promotion; title computed from contribution. Patent search required before external claims. | Attribution report for one production adapter (ArcTide E-5) | Attribution not reproducible across runs |
| FB-06 | Cross-family Rosetta layer One learned translator between model families' internal spaces so adapters transfer with little retraining. | Probe alignment of two small models on Swift tasks (relative-representation analysis) | No usable alignment between small models of two families |
| FB-07 | ArcTide — The Reversible Mind Knowledge round-trips between graph and weights under a ledger. Graph as cross-family medium. Ledger as title. Partly written down in pieces; unified loop not found in literature. Patent search required. | ArcTide E-1 round trip (Phase 1) | E-1 and E-2 both fail their pass signals twice |
| FB-08 | Pocket as team member's double Each engineer's ArcNode trains a personal adapter from their approved work, owned by them, usable across the Grid with receipts. Consent/contract terms required before production. | One volunteer's personal adapter on ArcNode with consent terms | Volunteer finds it unhelpful or terms unworkable |
Open decisions
For Charles — status as of 2026-10-18
- D-GR-01 (BLOCKING Phase 0) — Foundation strategy. Recommendation: Option A (one home base, others as teachers). Option C allowed if bake-off shows a 10+ point QA-006 gap for Development.
- D-GR-13 (BLOCKING Phase 1) — Where trace-derived training data lives. Resolve jointly with ArcHub D-AH-05 and ArcRouter port 4000 capture thread.
- D-GR-06 (proposed) — llama-server now; vLLM at next hardware tier; Ollama stays for fixed pairings.
- D-GR-08 (proposed) — Noema Check is never a component of the model it judges.
- D-GR-09 (proposed) — Use mergekit now; replace with Grid-owned code in Phase 3.
- D-GR-11 (proposed) — Ratify GRID-DOCTRINE-TIDE-001 (ArcTide) as a foundational doctrine.
- D-GR-12 (proposed) — Package name ArcGraft; doctrine name ArcTide. Collision check pending before external use.
- D-GR-02 (open) — First department for Phase 1. Proposed: Marketing; confirm with P0-04 worksheet data.
- D-GR-03 (open) — Client adapter ownership terms. Requires legal review (contract clauses per A-11).
- D-GR-04 (open) — First hardware upgrade: one 16-24 GB GPU, aligned with ArcNode M0 pilot.
- D-GR-05 (open) — Burst cloud GPU policy for fusion and full fine-tunes. DATACLASS ceiling and per-run spend cap required.
- D-GR-07 (open) — Long-run shape: single Noema MoE vs routed department cores. Decide on Phase 3-4 evidence.
- D-GR-10 (open) — Teacher roster owner: one accountable reviewer for license clearance. Required before Phase 3.
- D-GR-14 (open) — Finance as a sixth department. Carry from PKG-NOEMA-001; no change from this package.