Forge is the third architectural plane. Hyperlab + Hypernym is the company that ships it.
R12 produced the total company roadmap (16+ products, three axes, world-model 99.7% completion path, $20–50B/customer hyperscaler economics). R13 deepens it on three dimensions: (1) Forge structurally reframed from "harness" to Substrate Execution Plane (SEP) — the third architectural layer alongside Hypercore and Modulum; (2) Hyperlab + Hypernym as precision intelligence lab + product infrastructure company; (3) the 18-month execution physics that closes the substrate-network-lockup before AWS / Anthropic / Google can react.
4-of-4 panel converged on Forge as a plane/fabric/lattice — not a harness. Each model named its primitive (Grounded Step / Work Claim / Execution Track / Substrate Transaction). 3-of-4 used "Substrate Execution Plane" terminology. Once Forge is named correctly, every R12 product slots into a coherent platform diagram — Hypercore on the substrate plane, Modulum on the inference plane, Forge on the execution plane. The harness vendors (Cursor, Cline, Goose) are competing for the application-plane developer; Hypernym is competing for the plane-between-planes.
Forge is the Substrate Execution Plane (SEP) — a transactional execution engine over typed substrates whose primitive is the Grounded Step. The harness category exists in a substrate-free world; the SEP exists in a substrate-typed world; consumer-facing products like Forge OS are SEP shells, not the SEP itself. Cursor is to Forge what psql is to Postgres: a useful client that touches the system but is not the system.
13 rounds. Each compounds the last.
Each round = parallel cross-model panel + manual synthesis + deployable deck. Convergence detection across model panels = architectural-commit signal. R13 is the largest panel output of any round (282 KB API tier).
5/5 Reality Substrate · Vault · Crafter · Train-B
PDS as unit of product. Hypernym Vault unanimous. Crafter MVP unanimous. Continued Pretrain B at $550K/12wk.
5/5 Attention-Mask Conditioning
M5 unanimous architectural commit. PCHR / MaskGate / Modulum-SparseGate / SAS / Domain-Specific Head Pruning.
6 convergent primitive groups · 33 net-new
5/5 unanimous on Cognitive Gearing as universal hyperscaler primitive.
7 unanimous clusters · climate civilizational pick
3/5 panel (Codex NDA refusal · Gemma local depth ceiling). World-model "half-correct" verdict.
3 meaningful flips · $20–50B/customer · 6-element IP
Soft-ack worked; Codex full participant. Sum-of-All-Parts reverses R10. World-model 5/8 → 99.7% reachable.
16+ products · Three axes · $3T TAM
264 KB panel. Total company roadmap. Forge OS · SAMA · Atom-1.4B · cross-industry correlations · Tesla pitch.
Forge = Substrate Execution Plane · Hyperlab + Hypernym · 18-month execution physics
Largest panel output of family (282 KB). Forge structural reframe (4/4 panel: not a harness, third plane). Org structure (precision intelligence lab + product infrastructure company). 4-product Gantt with falsifier dates. R14 = substrate-propagation engine.
Hypernym competes for the plane between planes.
The harness vendors (Cursor / Cline / Goose / Aider / OpenHands) are competing in a substrate-free world — text plus tool side-effects, no typed reality, no structural confidence, no provenance. Hypernym competes one layer down. Three planes; Hypernym holds two and a half of them.
The harness market sells a UI in the application plane. Forge OS sells a UI but sits on the SEP — the moat is on the plane below the UI. Cursor is to Forge what psql is to Postgres: a useful client that touches the system but is not the system. Five years from now, every harness that survives will sit on top of a substrate execution plane of some kind. Hypernym's bet is that it owns the canonical SEP because it owns the canonical substrate plane (Hypercore) that defines what a substrate is in the first place.
Like the relation, the process, the document, the commit.
A database's primitive is the relation. An OS's primitive is the process. A browser's primitive is the document. Kubernetes' primitive is the pod. Git's primitive is the commit. Forge needs an equally sharp primitive. 4-of-4 panel produced one — different names, same structural object.
The Grounded Step — to Forge what the transaction is to Postgres.
The Grounded Step is the smallest atomic unit carrying: substrate handle (the typed slice of reality the step operates against), claim (what the step proposes to add, modify, or assert), confidence vector (mechanical × structural × compositional), provenance edge (which prior step / claim / model run / substrate revision produced this step's input), invariant signature (the substrate-type-checked predicate the output must satisfy, or be rejected). Forge's invariants are stated over Grounded Steps. Atomicity per step. Consistency via invariant signatures. Isolation via per-substrate locks and federation junctions. Durability via the Claim Lifecycle Ledger.
What every other agent stack lacks. What Forge solves.
Codex's framing — the cleanest articulation of why Forge is a new category, not "harness with extra features." Five missing invariants in today's agentic-development stack. Forge is the answer to all five.
11 of 16 R12 products run inside the SEP.
Once Forge is named correctly as the SEP, every R12 product slots in cleanly. The harness is one client of the SEP among many — the harness market is competing for the application plane while the SEP is a layer below.
H5 Hypercore ComposerH6 Hypercore PerimeterH7 Hypercore Relation GuardH9 Hypercore Audit ReplayM7 Modulum RouterM8 Atom-1.4BM9 Magic EverywhereP7 Substrate Diff EngineP8 Claim Lifecycle LedgerC7 Forge OST4 SAMAC9 Verified Agent RunsThe other 5 R12 products (Modulum Calibration Cloud · Modulum Edge Runtime · Substrate Pack SDK · Semantic Cost Observatory · RelationTruth Benchmark) sit alongside the SEP rather than inside it — they're SEP-adjacent infrastructure. Eleven products' runtime IS the SEP. That's why Forge is the connective tissue.
Precision intelligence lab + product infrastructure company.
The founder-named structure. Two arms, one company, two audiences, one architecture. DeepMind-style two-brand split — not OpenAI-style one-brand-blurry, not Cursor-style pure-product, not FAIR-style pure-research.
Scope
- Substrate-authoring craft — methodology, ontology design, composition-kernel craft
- Composition Type Theory — Gap 3 work; CMU PL / MIT CSAIL / UPenn academic partnership
- Federation cryptography — substrate-lineage attestation (Gap 2)
- World-model substrate experiments — Crafter / SWE-Bench / climate / drug discovery falsifiers
- 75%-noise universality follow-on — extending Modulum patent posture
- Academic partnerships — substrate co-authorship
- Public benchmarks — LongMemEval-Grounded, RelationTruth, etc.
Headcount profile
- PI-class research leads
- Graduate student researchers
- Substrate-craft engineers
- Benchmark engineers
- Paper-writing leads
- Academic-partnership program managers
Funding · IP
- Funded by Hypernym product revenue (capital flows up product → research)
- Patent-track work flows down to Hypernym; methodology/foundational papers go academic-public
- Trade-secret algebra (calibration, head-detection, composition kernels) stays proprietary
Scope
- Tier 1 — 4 Hypercore products (Engine · Magic · Omnifact · HyperRemember)
- Tier 1 NEW — 5 Hypercore (Composer · Perimeter · Relation Guard · Substrate Exchange · Audit Replay)
- Tier 2 NEW — 5 Modulum (Router · Atom-1.4B · Magic Everywhere · Calibration Cloud · Edge Runtime)
- Tier 4 NEW — 4 Public (Forge OS · Forge Control Plane · Verified Agent Runs · Substrate Pack SDK)
- Tier 5 NEW — 3 Trust (SAMA · SPAS · RelationTruth Benchmark)
Headcount profile
- Engineers — substrate · runtime · frontend · devrel
- Sales (vertical-domain enterprise)
- Customer success (Sample → Pilot → VPC motion)
- Product management
- Marketing
Revenue model
- Per-customer enterprise — Hypercore Engine, $1–10M ARR per vertical
- Per-seat developer — $40–80/mo Forge OS, Magic plug-ins
- Per-token — Modulum Router OpenRouter
- Per-call — Hypercore APIs
- Per-substrate-hour — SAMA
- Licensing — Modulum to hyperscalers ($20–50B/year/customer)
Two-way mobility between research and product.
How Hyperlab and Hypernym Compose
Hyperlab → Hypernym (research flows down)
- Patent-track research → product features (Composer · Relation Guard · Federation Protocol)
- Substrate-craft methodology → customer deployment quality
- Benchmark publication → marketing / sales moat
- Academic partnerships → standards-body authority
- Trained substrate engineers → Hypernym customer-facing roles
Hypernym → Hyperlab (revenue flows up)
- Customer-substrate accretion data → research input
- Production telemetry → research priority signals
- Product revenue → research budget (no VC dependency)
- Customer engineers identifying gaps → rotate into Hyperlab
- Real-world failure modes → falsifier experiment design
The case against four alternative shapes.
Hypernym is the product-infrastructure company that owns substrate as a category. Hyperlab is the precision-intelligence research arm that develops the load-bearing IP — Composition Type Theory, federation cryptography, substrate-craft methodology — that no product-only competitor can replicate without the research foundation. The capital flows up: product revenue from Hypernym funds Hyperlab's 24-month research programs; research output flows back as patents, methodology, and benchmark publications that compound the Hypernym product moat. Two brands, one company, two audiences, one architecture.
The four highest-leverage products. Week-by-week.
Hypercore Composer (closes Gap 1) · Atom-1.4B (cheapest-train proof) · Forge OS (consumer flagship) · SAMA (B2B agent layer). Each ships on a falsifier-gated trajectory.
16 → 75 headcount over 18 months. Substrate engineers don't exist yet.
| Quarter | Hyperlab | Hypernym | Total | Critical hires |
|---|---|---|---|---|
| Q1 | 4 (founders + 1 PI + 2 substrate-craft eng) | 12 (founders + 6 product eng + 2 sales + 2 CS + 1 design + 1 ops) | 16 | CTO of Hyperlab · VP of Customer Success |
| Q2 | 6 (+1 PI, +1 federation cryptographer) | 18 (+2 eng, +2 sales, +1 marketing, +1 DR) | 24 | Substrate Craft Lead · Compiler PL Hire (for Gap 3) |
| Q3 | 9 (+2 grad students, +1 paper-writing lead) | 25 (+5 eng, +2 sales) | 34 | Chip Co-Design Partnership Lead |
| Q4 | 12 (+2 substrate-craft, +1 PI) | 32 (+5 eng, +2 sales) | 44 | VP of Engineering · Head of Marketing |
| Q5 | 16 | 42 | 58 | Head of Federation |
| Q6 | 20 | 55 | 75 | (Stable scaling) |
The substrate-engineer hiring funnel — the unique challenge. This talent pool doesn't exist today; Hypernym must create it. Three-channel recruiting strategy: (a) academic from CMU / MIT / Stanford PL groups (Hyperlab apprenticeship); (b) re-skilling backend engineers via 12-week internal substrate bootcamp; (c) customer-facing deployment engineers from Osmium / TrustFoundry rotated into Hypernym substrate engineers. Hyperlab apprenticeship is the load-bearing pipeline.
Modulum deployment path #2. The 18-month-window-relevant milestone is the MOU, not the silicon.
Silicon programs take 18-36 months. Full silicon ships post-window — but the partnership MOU at month 5-6 establishes the lockup before competitors can react.
| Partner candidate | Architecture fit | IP risk | Timeline |
|---|---|---|---|
| Cerebras (primary) | Wafer-scale matches Modulum routing primitives well | Low — already commodity-routing-friendly licensing | MOU M5 · tape-out M9-12 · production M24-30 |
| Tenstorrent (alt) | Less mature than Cerebras but more architectural-co-design willing | Medium — newer company, IP terms negotiable | MOU M6 · tape-out M12-15 |
| Apple Silicon | Best for consumer products via MLX | High — Apple's IP terms punitive for licensors | Partnership M9+ if Apple opens program |
| Groq | Speed-advantage for inference | Medium-high — architectural lock-in to their hardware | Standalone licensing track, no co-design |
| Custom fabless (Mead-style) | Maximum control | Maximum cost + risk | Out — too long for 18-month window |
IP partition. Hypernym retains Modulum routing primitives + calibration algorithm + head-affinity tables. Partner retains physical-layer / process-node IP. Joint IP for the silicon-routing integration block. Partition survives partnership dissolution because either party can re-implement against alternative partner.
Six architectural assumptions. Six experiments that kill them.
| Assumption | Falsifier experiment | By | |
|---|---|---|---|
| 1 | 75%-noise universality holds across new architectures | Run head-importance analysis on Llama 4 / DeepSeek V4 / GLM-5 / GPT-OSS-120B; verify ≥70% noise across all 4 | M3 |
| 2 | Compositional Planner generalizes across substrate types | Test on legal (TrustFoundry) + climate substrates beyond biomedical (Osmium) | M6 |
| 3 | Substrate Federation Protocol junction confidence calibrates in production | Deploy across 3 federated substrates; measure junction-confidence vs ground-truth | M9 |
| 4 | Atom-1.4B competes with frontier 8B-class models | Run vs Llama-3.1-8B on 6 substrate-typed benchmarks | M16 |
| 5 | Hyperscaler $20–50B/customer/year survives integration overhead | Day-90 hyperscaler pilot at AWS / GCP candidate | M5 |
| 6 | Forge OS hits 5K paying users in 90 days post-launch | Launch W12 + 90-day measurement | M15 |
Kill-switch protocol. Any falsifier that fires triggers a 14-day re-plan cycle, not an immediate halt. The roadmap survives 1–2 falsifier failures by re-sequencing; 3+ failures invalidate the 18-month execution architecture and trigger a strategic re-evaluation.
The benchmark + paper schedule that converts the hyperscaler sales cycle.
| Month | Publication | Venue |
|---|---|---|
| M3 | RelationTruth Benchmark (composition-layer hallucination) public release | arXiv + benchmark website |
| M6 | LongMemEval-Grounded Cert methodology paper | NeurIPS workshop submission |
| M9 | Atom-1.4B technical report — the press anchor | arXiv + tier-1 press |
| M12 | Hypercore Composer architecture paper | ICLR submission |
| M15 | Substrate Federation Protocol cryptographic spec | RFC + arXiv |
| M18 | Composition Type Theory v1 (Hyperlab + academic partner) | POPL submission |
The press anchor is Atom-1.4B at M9. The article that lands: "8B-class model trained from scratch on Hypernym's substrate-typed architecture beats Llama-3.1-8B on 6 benchmarks at 1/5 the parameter count." This is what converts the hyperscaler sales cycle from 12 months to 4 months — pitch becomes "here is the model, beating yours, on your benchmarks."
The fourth axis we're still missing.
R12 + R13 specify the technical primitives, the company structure, the execution physics. None specifies the propagation engine — the structural mechanism by which substrate accretion compounds across customers. Without it, the substrate-network-lockup is fragile.
The Substrate-Propagation Engine
What is the substrate-propagation engine — the structural mechanism by which the Nth customer's substrate quality is materially better than the (N−1)th's because of the prior N−1 customers — and how does Hypernym engineer it as a load-bearing primitive (not as a marketing artifact) such that the substrate-network-lockup is irreversible rather than merely first?
R14 must produce: (a) first-principles definition of the substrate-propagation engine; (b) architectural primitive at its center (analogous to Grounded Step at Forge's center); (c) network-effect math (under what conditions does customer-N+1's substrate quality strictly dominate customer-N's, and what's the half-life); (d) specific Hypernym product line that is the propagation engine; (e) falsifier — what measurement at month 18 proves or disproves that the propagation engine is operating as designed.
The R12 → R13 progression went: roadmap → company structure + execution. The R13 → R14 progression should go: execution → moat compounding. That sequence is the irrevocable-path architecture taken to completion.
R13 in three observations.
If R7 said "PDS is the unit of product", R8 said "M5 is the mechanism", R9 said "Modulum makes knowledge operational at a layer below code", R10 said "softmax-level operation is audit-grade truth", R11 said "the full system makes world-model precision a substrate-engineering problem", R12 said "Hypernym is the substrate company in an industry that mistook itself for a model company" — then R13 says: Forge is the Substrate Execution Plane, the third architectural layer beneath every Hypernym product. Hyperlab + Hypernym ships substrate as both research IP and product infrastructure simultaneously. The 18-month execution architecture closes the substrate-network-lockup before AWS / Anthropic / Google can react.