ROUND R17 · FEDERATION ACTIVATION + STRATEGIC SURFACE
2026-05-11 · 11 streams · 5/5 R1 · 5/5 R2 · 2 rounds at max budget
R16 closed the algebra. R17 closes the positioning. Modulum proof doc lands mid-round (BABILong-validated, 131k effective context, frontier-lab regression confirmed) and the panel converges on Hypernym as the licensed Truth Maintenance System for institutions — not an AI feature, not a verification API, not a memory layer. The corpus is the moat; the certificate is the product.
R17 produced design decisions. No code shipped. The 90-day product plan from R16 (Verify GA · Modulum Solo · Legal Endpoint) is still the build path; R17 layered the strategic position on top.
Codex R2 framing of the proof doc. Frontier labs advertise 1M-token context that performs at 200-400k. Hypernym advertises 131k that performs at 131k. Hypernym sells effective retained context as the audited unit.
Claude R2 framing of the proof doc + panel convergence. Frontier labs can't fix attention math via post-training (Opus 4.6 → 4.7 regressed 32-46pp). They'll license Hypernym wholesale for the verification layer they can't build internally.
Every panelist independently landed on a position that the others' surfaces complement. The synthesis adopts all four as facets of one product.
Customer frame: courts, regulators, hospitals, auditors — not AI labs.
Protocol frame: Verisign for AI. Hypernym signs VTC validity cryptographically.
Distribution frame: Hypernym beneath every frontier lab as licensed wholesale layer.
Ecosystem frame: VTCs attest against any LLM weights — frontier and OSS.
Post-R2 cross-pollination + proof-doc evidence integration.
| Stream | R17 Verdict | Key resolution |
|---|---|---|
| 1 — LAP | sound (stratified) | Factual oracle-priority + interpretive reliance-class arbiter; tiered slashing |
| 2 — DomainBridge | sound | First-class VTC + loss semantics + regulator-published mappings + bridge royalties |
| 3 — RedactionBridge | sound | Build substrate layer + audit ZK partner; policy compiler UX |
| 4 — Refusal benchmarks | sound + dynamic | Static (procurement) + dynamic adversarial bounty (frontier-lab pressure); CC-BY-SA-4.0 |
| 5 — integration_step | partial | Bounded industrial twins Y1; climate Y2+ Hyperlab R&D only |
| 6 — Density floors | sound | Floors attach to reliance_class, not raw domain |
| 7 — Cost margin model | sound | Reuse-ratio = master variable; insurance-backed CPAT for high-stakes classes |
| 8 — Substrate strategy meta | sequenced Y1/Y2/Y3 | Y1 Hyperlab-owned + Modulum-cloud-API · Y2 Exchange Model · Y3 multi-modal federation |
| 9 — Skipped surface audit | pivot — regulatory Y1 critical | FDA + FedRAMP + SOC2 + HIPAA escalated to Year-1; hardware Y1 partnership scope |
| 10 — Competition / Positioning | sound (unified) | TMS-of-Institutions + Verisign + Infrastructure-tax + Model-as-Substrate |
| 11 — Anti-thesis | pivot — 3 actively breaking | Validation cost unbounded (insurance offsets); federation Y1-unwinnable (sequenced); algebra incomplete (Reliance-Class + Hollow Nodes + Info-Gain Monotonicity) |
High-convergence decisions that anchor R17 closeout and R18 scope.
state_before. Collapses LAP arbitration tier + DomainBridge cost + RedactionBridge disclosure + density floors + refusal correctness + pricing + liability into one field. (Codex origin.)Hypernym - Modulum Attention - First, Not Lost.pdf (2026-05-08) landed during R1 dispatch. Modulum shipped as live OpenAI-compatible API.
Measured BABILong qa1 (n=100, same Gemma 4 31B, same hardware, only Modulum varied):
| Context | Vanilla | +Modulum | Δ |
|---|---|---|---|
| 32k | 87.0% | 89.0% | +2.0pp |
| 64k | 74.0% | 80.0% | +6.0pp |
| 128k | 60.0% | 69.0% | +9.0pp |
Improvement widens as context grows. Signature of a kernel-level architectural fix, not benchmark-tuning. Reproducible via public BABILong runner.
Frontier-lab regression confirmed:
| Model | Length | Result |
|---|---|---|
| Opus 4.6→4.7 | 256k MRCR | -32.7pp |
| Opus 4.6→4.7 | 1M MRCR | -46.1pp |
| GPT-5.5 | 200-400k | 74% multi-needle |
| Opus 4.6 self-report | 48% fill | "recommends restart" |
| Modulum (Gemma 4 31B) | 131k full | 131k effective |
Frontier moved 32-46pp the wrong direction in 18 months. Hypernym moved 9pp the right direction. Cost differential: ~5 orders of magnitude. Modulum was developed by a single lab, on a single 31B-parameter model, without retraining from scratch.
Modulum API endpoint smoke-tested during R17 dispatch returned 503 ("Tundra backend on :8090 not running"). Front-end alive, backend down. Pre-GA reliability is a strategic gap. Codex R2: "Shipped but unreliable damages trust faster than unshipped." Year-1 reliability ops are now critical-path.
R18 = "Institutional Reliance-Class VTCs as the licensed protocol layer." Unicorn-trajectory if all 4 close.
One court system (Delaware Chancery for commercial law), one regulator (FDA / NIH via Hyperlab), one Big-Four auditor (PwC or EY).
Most likely Anthropic (safety-positioned brand). Wholesale per-call licensing for "Hypernym-verified" badge on Anthropic reliance VTCs.
Munich Re or Lloyd's. Underwrites specific reliance classes (court-filing, regulatory-submission, clinical-support).
Published taxonomy of 7 reliance classes as open standard. Competitor labs implement against the spec. Hypernym writes; everyone follows.
Hypernym is unicorn-trajectory by Year-2, $5-10B by Year-3. The institutional licensing layer + insurance economics + frontier-lab wholesale + open standard creates a moat no frontier lab can buy their way out of via training spend.
Each round produces 3-5 outliers worth tracking; these are R17's signal.
MSA per frontier lab. Wholesale per-call pricing. Lab markets "Hypernym-verified" badge at retail markup. Distributes Hypernym beneath the entire industry.
Nutrition-label-style receipt for every model call: effective context length · retrieval confidence by depth band · dropped-evidence risk · contradiction scan status · refusal obligation. Standardizes how customers compare AI systems.
Munich Re / Lloyd's partnership. Client pays insurance premium. Hypernym takes commission. Removes procurement convince problem entirely — liability offloads to a familiar contract.
R18's seed scope. Critical path items first.
R16 closed the algebra. R17 closed the positioning. The four R1 framings — Truth Maintenance System for Institutions · Certificate Authority of Machine Reasoning · Infrastructure-tax / Akamai-for-verification · Model-as-a-Substrate — are facets of one product: the licensed authority that signs all VTCs in the AI ecosystem.
Modulum-shipped reality means Year-1 doesn't need substrate accumulation; it can sell effective retained context today. Frontier labs are visibly regressing while Hypernym moves forward at 5 orders of magnitude lower cost. Regulatory pathway is Year-1 critical-path. Insurance-backed CPAT removes procurement convince problem. Frontier labs become customers, not competitors.
R18 ships the licensed protocol layer. If 4-of-4 anchors close, Hypernym is unicorn-trajectory by Year-2, $5-10B by Year-3.