Hypernym Compound Research — plain-language status.
This document explains, in plain language, what we've done across 14 rounds of compound research, what each artifact actually is, what it does, what it enables, and where it integrates across Hypernym Labs. It is the bridge document you take to the team before R15 to make sure everyone is aligned. Companion to the eight per-round decks (R7 → R14).
What this whole thing IS · One paragraph
Compound research = a sequence of 14 ideation rounds, each consisting of: (1) a confidentiality-preambled prompt, (2) sent in parallel to 4–6 reasoning models (Codex / Claude / Gemini / Grok / local models), (3) synthesized by hand into a decision document, (4) packaged into a deployable HTML deck on Cloudflare Pages. Output is decision artifacts, not code. Nothing has been built. The decks are for cofounder review and team alignment.
Each round produces one architectural commit.
Convergence (3-of-5, 4-of-5, 5-of-5) is the alignment signal — when models reasoning independently arrive at the same primitive, that is the load-bearing decision.
R7 family · Foundation
2026-04-30 → 05-01
Locked
What it is
Foundation round. Asked: what is Hypernym's atomic unit of product?
What it produced
PDS (Persistent Domain Schema) as the unit of product (5/5 unanimous). Hypernym Vault product (5/5). Crafter MVP (5/5). Continued Pretrain B at $550K/12 weeks (5/5).
What it enables
Every downstream round assumes PDS as the structural primitive sold to customers.
Integrates with
Hypercore (4 products: Engine, Magic, Omnifact, HyperRemember).
Status
Locked. Don't relitigate. Deck: hypernym-vision.pages.dev
R8 · Mechanism
2026-05-02
Locked · Patent-track
What it is
Asked: what's the inference-time mechanism that makes substrate-aware models work?
What it produced
M5 Attention-Mask Conditioning (5/5 unanimous). Five sub-mechanisms: PCHR, MaskGate, Modulum-SparseGate, SAS, Domain-Specific Head Pruning.
What it enables
Modulum's 3.04× decode speedup and 75% noise universality — the load-bearing model-runtime patent.
Integrates with
Modulum (inference + memory layer). Specifically the runtime that sits over OpenAI/Anthropic/Bedrock/Together/Groq.
Status
Locked. Patent-track filed/filing. Deck: hypernym-vision-r8.pages.dev
R9 · Unlocks
2026-05-03
Locked · Patent-track
What it is
Asked: given M5, what unlocks at the hyperscaler tier?
What it produced
Cognitive Gearing as universal hyperscaler primitive (5/5 unanimous). Six convergent primitive groups, 33 net-new ideas.
What it enables
The "$20–50B per year per hyperscaler customer" pitch (R11) lands because Cognitive Gearing is what hyperscalers buy.
Integrates with
Modulum runtime + chip-partner sequence (Cerebras, Tenstorrent, Apple Silicon).
Status
Locked. Patent-track. Deck: hypernym-vision-r9.pages.dev
R10 · Softmax
2026-05-05
Half-correct · reframed in R11
What it is
Asked: at the softmax layer, what's the audit-grade truth claim?
What it produced
7 unanimous clusters; climate emerged as the civilizational pick. World-model verdict 5/8 reachable (later REVERSED to 99.7% in R11).
What it enables
The civic / public-good positioning (climate substrate as a non-commercial Hyperlab project).
Notes
Codex refused NDA on this round (fixed in R11 via Form B soft-ack). Local Gemma Q4 hit depth ceiling on 23K-character prompts (fixed via MLX-native Q8 + MTP drafter, complete today).
Status
World-model claim REVERSED in R11. Climate cluster still standing. Deck: hypernym-vision-r10.pages.dev
R11 · Meta Reframe
2026-05-06
Locked
What it is
Asked: take the sum of all parts — what does the full system actually mean?
What it produced
Three flips. (1) World-model 5/8 → 99.7% reachable as substrate engineering, not model training. (2) Hyperscaler economics $1–5B/yr → $20–50B/yr per major customer. (3) 6-element IP-protection strategy (binary-blob + black-box ABI + calibration-as-service + VPC-bounded + no co-located papers + joint chip co-design).
What it enables
The pitch math. The patent strategy. The world-model 99.7% completion path.
Integrates with
Investor decks · CTO/COO briefings · academic-partner conversations · hyperscaler design-partner pitches.
Status
Locked. This is the pitch math. Deck: hypernym-vision-meta.pages.dev
R12 · Roadmap
2026-05-08
Roadmap reference
What it is
Asked: convert the architectural commits into a total company roadmap.
What it produced
16+ products across 4 tiers (Hypercore Tier 1, Modulum Tier 2, Public Tier 4, Trust Tier 5). Three axes (precision, distribution, defense). $3T TAM by 2028. Tesla pitch (vertically integrated where it matters, partnered where it doesn't).
What it enables
Sequencing decisions. What ships when. Hiring-funnel sizing.
Integrates with
Everything.
Status
Locked as roadmap reference. Individual product greenlight decisions made per-product by R&D team. Deck: hypernym-vision-roadmap.pages.dev
R13 · SEP + Company Structure
2026-05-09 morning
Architectural commit signal
What it is
Asked: reframe Forge structurally. Define company structure. Specify 18-month execution physics.
What it produced
- Forge = Substrate Execution Plane (SEP) — third architectural plane alongside Hypercore (substrate) and Modulum (inference). Forge is not a harness like Cursor/Cline; it's a layer below them. (4/4 panel.)
- Grounded Step = SEP's primitive (substrate handle + claim + confidence vector + provenance edge + invariant signature). Like the transaction in Postgres.
- Hyperlab + Hypernym = DeepMind-style two-brand split. Hyperlab = research lab (long, deep, academic). Hypernym = product company (fast, customer-facing). Capital flows up from product to research.
- 18-month execution physics — 4 highest-leverage products (Hypercore Composer, Atom-1.4B, Forge OS, SAMA), 16→75 headcount, Cerebras MOU at M5–6, Atom-1.4B press anchor at M9, falsifier table.
What it enables
Org structure clarity. Hiring plan. Chip-partner sequencing. Press cycle.
Integrates with
Forge code (this repo, build OK) · Hypercore (your team, alignment first) · Hyperlab (your hires) · VC conversations.
Status
Locked structurally. Individual hiring/chip-partner decisions are R&D + cofounder decisions. Deck: hypernym-vision-r13.pages.dev
R14 · Substrate Lemma + Propagation Engine
2026-05-09 evening · today
Architectural commit signal
What it is
Asked: R12+R13 specified the products and execution. R14 closes the missing axis — the propagation engine that converts first-mover advantage into compound lockup.
What it produced
- Substrate Lemma = the typed unit of substrate refinement that propagates between customers (5/5 panel-convergent on the structural object — tightest convergence of any round).
- Three-product propagation engine:
— Hypercore Distill (customer-facing): watches Grounded Steps, distills lemma candidates with privacy preservation (TEE-attested K-anonymous ε-DP).
— Substrate Refinery (Hyperlab back-end): multi-stage quality gate, cryptographic attestation, registry maintenance.
— Substrate Exchange (public marketplace): federation marketplace for opt-in cross-customer lemma trade. - Network-effect math — typed lemmas give competitors β≪1 in the compounding equation (copying type system = provable IP infringement). Untyped propagation gives β≈1 (no advantage).
- M18 falsifier — Compound Onboarding Lift: customer-7 week-4 quality must be ≥1.5× customer-1 week-4 quality.
- Spanner-shaped patent claim — single combination patent covering 6 elements (none novel alone, the combination is patentable).
What it enables
Patent strategy. Substrate-network-lockup math. M18 OKR design.
Integrates with
Hypercore Distill into Engine pilots (Osmium, TrustFoundry). Substrate Refinery into Hyperlab. Substrate Exchange as Tier-5 trust product. M5 attention-mask routing rules become lemma
body field. Federation Protocol becomes the cross-substrate validation step in the Refinery's gate.Status
Architectural commit signal. This is a research output for cofounder review — not a build greenlight. Build decision belongs to you + R&D team. Deck: hypernym-vision-r14.pages.dev
The layer map.
Each round's primitive sits at a specific architectural layer. Hyperlab sits horizontally across all of these — research feeds patents and methodology back into every layer.
Application planeCustomer's product
The customer's actual product — the medical AI, the legal-research tool, the codegen IDE, the financial-analyst dashboard. Unchanged. Hypernym does not compete here.
Execution planeWhere work happens
Forge / Substrate Execution Plane (SEP) R13 · Primitive: Grounded Step R13 · Multi-agent: SAMA R12. Cursor / Cline / Goose are substrate-free degenerate execution-plane competitors. Forge is the substrate-aware version one layer below them.
Propagation engineWhere compounding happens
Hypercore Distill → Substrate Refinery → Substrate Exchange R14 · Primitive: Substrate Lemma R14. The mechanism by which customer-N+1's substrate quality dominates customer-N's because of the prior N customers.
Substrate planeTyped reality
Hypercore (Engine, Magic, Omnifact, HyperRemember + new: Composer, Perimeter, Relation Guard, Substrate Exchange, Audit Replay) R7 · Unit of product: PDS R7. Hypernym defines this category.
Inference planeModel serving
Modulum (Router, Atom-1.4B, Magic Everywhere, Calibration Cloud, Edge Runtime) R7-R12 · Mechanism: M5 Attention-Mask Conditioning R8 · Primitive: Cognitive Gearing R9. Sits over OpenAI / Anthropic / Bedrock / Together / Groq / Cloudflare Workers AI.
Chip partner layerJoint silicon co-design
Cerebras / Tenstorrent / Apple Silicon / Groq R13. Joint chip co-design is the 6th element of the IP-protection strategy. The 18-month-window-relevant milestone is the MOU at month 5–6, not the silicon itself.
What's a decision vs what's still research.
| Item | Type | Who decides |
|---|---|---|
| PDS as unit of product | Locked decision | Already ratified across rounds |
| M5 attention-mask conditioning | Locked decision | Patent-track filed/filing |
| Cognitive Gearing | Locked decision | Patent-track |
| 6-element IP-protection | Locked decision | Cofounder + legal already onboard |
| 16+ product roadmap | Roadmap reference | Per-product greenlight = R&D team |
| Forge = SEP framing | Architectural commit | Affects how Forge code structured (you can decide) |
| Grounded Step primitive | Architectural commit | Forge spec — you can decide |
| Hyperlab + Hypernym split | Org structure proposal | Cofounder consensus needed |
| 18-month execution Gantt | Plan proposal | Cofounder + R&D consensus |
| Substrate Lemma primitive | Architectural commit signal | R&D team consensus before building |
| Hypercore Distill / Refinery / Exchange | Product proposal | R&D + product team greenlight |
| M18 Compound Onboarding Lift | OKR proposal | Cofounder + R&D consensus |
| Spanner-shaped combination patent | Patent strategy | Legal + cofounder consensus |
Build/Don't-build map.
You CAN build (in this Forge repo)
- Forge SEP changes — architectural primitive plumbing
- Grounded Step type system — Forge spec implementation
- Forge OS shell — consumer-facing harness UI
- FIX_DESIGN_RIGHT — current sprint, audit fail-open hardening
- Eval suites, sprint FSM, dispatch tooling — Forge infrastructure
You should NOT build alone
- Hypercore Distill — touches Hypercore product directly
- Substrate Refinery — Hyperlab product
- Substrate Exchange — marketplace product
- Atom-1.4B training — needs $550K + cofounder sign-off
- Modulum runtime changes — patent-protected layer
- Anything that touches Hypercore, Modulum, or substrate-product code
To avoid scope creep before R15.
- Building Substrate Lemma code without team consensus (it is a spec, not implementation)
- Atom-1.4B training without budget approval ($550K/12wk per R7 — needs cofounder sign-off)
- Chip co-design before MOU signed with Cerebras (M5–6 milestone)
- Public RFC for Substrate Lemma before patent filing (would compromise IP-protection)
- Hyperlab academic partner outreach before legal frames the IP-share terms
- Customer pilot expansion beyond Osmium / TrustFoundry / Amble before propagation engine is at least specced
- More compound-research rounds beyond R15 without your explicit "fire R16"
What R15 will ask.
R15 — Substrate Distribution Physics
Where compounded substrate physically lives, and how it reaches every customer at hyperscale latency.
What is the substrate-distribution physics — the structural mechanism by which compounded substrate (lemmas, refinements, attestations) is physically distributed to every customer at hyperscale latency? Where does the substrate live? Who serves it? What is the edge-cache vs hyperscaler-bundle vs direct-to-customer trade-off?
Why this is next
Three of five R14 reviewers explicitly named distribution as the next missing axis. The propagation engine answers what compounds; it doesn't answer where the compounded substrate physically lives or how it reaches customer 7's deployment in single-digit milliseconds.
What R15 will produce
- Distribution architecture (edge / regional / hyperscaler / customer-VPC)
- Substrate cache primitive (probably a separate primitive from Lemma — panel will name)
- Hyperscaler-bundle vs direct-to-customer economics
- Latency math (what is 99p latency for
query(substrate, ctx)?) - Falsifier — what M18 measurement proves distribution is operating
What R15 will NOT do
- Propose new primitives that contradict Substrate Lemma
- Relitigate Hyperlab/Hypernym structure
- Change the 18-month execution Gantt (sequencing-only adjustments allowed)
- Specify chip-level instructions (separate R16+ topic)
What you decide before R15 fires
- Fire R15 as specified above, OR
- Pause Hypernym track, refocus on Forge build work (FIX_DESIGN_RIGHT, audit hardening), OR
- Modify R15 framing — different question, different scope
Default if you say nothing: I will not autonomously fire R15. The Hypernym track is paused on your word.
Four sentences.
1
R7→R14 = 14 rounds of strategic R&D ideation. Output: decision artifacts (decks). 8 deployable HTML decks at distinct CF Pages URLs, no overwrites.
2
Nothing has been built that touches Hypercore / Modulum / Hypernym / Hyperlab products. The Substrate Lemma is a spec on paper, not a TypeScript interface or schema.
3
Forge-only changes can proceed in this repo (FIX_DESIGN_RIGHT etc.). Anything touching Hypercore/Modulum/Hypernym needs team R&D alignment first.
4
R15 is teed up but not fired. You decide.