Ship α validation now. Train β from α traces later. The unit between them is the Verified Transition Cell.

R15 is a strategic pivot from R7-R14's depth-first architecture rounds to a balanced view: 90-day product wedges that ship to one buyer who hands you money, plus 18-36 month zero-to-one R&D bets, plus moat analysis cross-cutting both. The 6-model panel converged 6/6 on the structural primitive that ties wedges to bets — every Track A product emits trace data that becomes Track D's training corpus.

The primitive is the Verified Transition Cell — a typed unit of valid state change carrying preconditions, postconditions, invariants, confidence, and cryptographic provenance. Frontier models propose; Hypernym validates; committed VTCs become composable units for long-horizon simulation. This is what Hypernym ships once R14's Substrate Lemma propagation engine has substrate density to compound on.

The 6-model panel.

Top-tier from each provider. NDA-bound. First-principles only. Six models reasoning independently arrived at the same structural object with six different names — same convergence pattern as R14's Substrate Lemma.

The convergence is on sequencing, not the wedge.

Track A consensus: 6/6 included a verifier or trace-grading product. 6/6 included context compression / pre-routing. 6/6 included Modulum Router or inference SaaS. 5/6 included persistent memory. 5/6 included domain endpoints. 5/6 included false-positive elimination. 4/6 included Forge OS Solo or IDE plugin. The shared ranking logic was speed to first customer, substrate data generated, and whether the product can be bought by one developer or one team without a network.

Moat consensus: 6/6 honest that Modulum Router and generic inference are revenue wedges, not durable moats unless M5 cost-quality claims remain protected and measurable. 5/6 treated Omnifact verification as medium-to-strong because the generic API is copyable but substrate provenance and domain substrate are harder. 6/6 placed the deepest moat in VTC, Substrate Lemma propagation, Federation Protocol, and domain substrate density.

R16 split: Codex pushed VTC; Claude pushed Substrate Distribution Physics; Gemini and Qwen leaned temporal/counterfactual; Grok pushed world-model continuation; Gemma pushed temporal substrate + counterfactual fork. Synthesis: R16 is "VTC + temporal/counterfactual closure" — subsumes all surfaced candidates into one falsifiable architecture round.

Ten ranked wedges. One-buyer = full value.

Wedges where one customer hands you money in 2-16 weeks. Ranked by panel-converged speed-to-customer × moat strength. Each scored on the data corpus it generates for VTC training (the γ data flywheel).

The decision rule: ship weak-moat wedges only when they feed a strong-moat asset. A router that merely routes tokens is weak; a router that learns cost per verified transition across domains is useful. A memory API that stores summaries is weak; a memory API that builds grounded, invalidatable project substrate is useful. A citation API that attaches links is weak; a citation API that decomposes claims into reusable provenance cells is useful.

Seven ranked moonshots. 18-36 month horizon.

VTC ranks #1 because it's the missing unit that makes "simulation-grade targeted intelligence" operational. If it works, Hypernym becomes a transition-validation layer across frontier models, simulators, and later substrate-native models. Substrate Lemma stays #2 as core but coordination-heavy — its moat improves only after Track A produces substrate density. Substrate-native Modulum (β) sits at #3 — fund it, don't overfund before α products create VTC traces and failure taxonomies.

Four moat archetypes. Be honest about which you have.

Durable structural moats require: proprietary substrate that improves with use, patentable transition/validation mechanics, regulatory or audit embedding, or standard/protocol adoption. Execution moats are not enough in AI infrastructure — well-funded competitors copy API shape quickly.

• Strong structural moats: VTC simulation stack · Substrate Lemma propagation · Federation Protocol after adoption · domain substrate in legal/biomed · Substrate Audit Replay · false-positive elimination in audited workflows. These create switching cost through accumulated substrate, replay obligations, calibration history, or cross-system protocol dependency.
• Medium moats: Omnifact Verify · Trace-Grading · Context Pre-Router · Fine-Grained Citation · Persistent Memory · Domain-Precision Endpoint before deep customer data. These become strong only if they feed a proprietary corpus of claims, traces, refusals, and validated transitions. Without that corpus, they're features.
• Weak fundamentals: Modulum Router · generic inference SaaS · generic IDE plugin · prompt safety gate · broad AI governance suite. Worth shipping if they generate cash, substrate data, distribution, or benchmarks. Kill them if they become support-heavy commodity software.

Avoid confusing patentability with monopoly. M5-conditioned routing, substrate validation, and transition cells may be filable, but patents do not replace proof. The commercial moat appears only when customers believe the measured output: lower false positives without higher false negatives, fewer unsupported claims at the same task completion rate, shorter context without recall loss, longer simulations without calibration drift. Every wedge ships with a benchmark harness as part of the product, not as a research afterthought.

The strongest compound loop: verify claims → log failures → classify failure modes → convert repeated failures into substrate lemmas → use lemmas to improve future verification → distill the resulting VTC corpus into substrate-native Modulum. Products that don't contribute to this loop are cash extraction or distribution experiments, not core strategy.

Verified Transition Cell. Like the relation, the process, the commit.

The VTC schema — merged from all 6 panel proposals.

Three load-bearing fields: failure_mode prevents the α-to-β corpus from becoming noisy (a bad proposal, stale evidence, missing substrate, and invariant bug should not train the same correction). scale_level prevents molecular claims from composing directly into organism-level claims without a declared bridge. commitment makes simulation replay auditable; without a hash chain, VTC becomes another unverifiable trace format.

Validity invariants: state_before and state_after must be typed under the same domain schema or a declared ScaleBridge. Hard invariants cannot be violated. Preconditions must be satisfied before application. Postconditions must be checkable or explicitly marked unobserved. Every accepted VTC must have a hash commitment over inputs, outputs, invariants, and validation trace. Composition is legal only when the first cell's state_after satisfies the second cell's preconditions and no invariant contradictions cross the boundary.

Six operations. All return a VTC, a VTC trace, or a typed refusal.

Closure matters: if any operation returns untyped narrative, long-horizon simulation collapses back into hallucination. merge and revert need strict treatment in R16 — incompatible branches can look semantically compatible while hiding contradicted assumptions; substrate updates will invalidate earlier cells and a simulation platform that can't revoke descendants accumulates stale certainty.

The 6/6 panel consensus is γ.

The recommended β target is not a general frontier model. It's a substrate-native proposer/validator optimized for domains where Hypernym owns substrate density. Atom-scale handles extraction, routing, compression, local validation. 8B-class powers domain endpoints and memory. 30-70B-class becomes relevant only after VTC traces prove enough signal for long-horizon reasoning. Trillion-scale training is not required if the company remains targeted rather than general.

Ten products that become possible only with VTC at the center.

The common product pattern is not "replace the domain simulator" — it's "wrap proposers with transition validity." That distinction keeps scope sane. Hypernym should not build a climate model, AV simulator, chemistry engine, and macroeconomic model from scratch. It should validate, compose, calibrate, and audit the transitions those systems produce, then train substrate-native proposers only where repeated validation failures show frontier or incumbent tools cannot generate the right candidate transitions.

What the panel said we hadn't addressed.

• First-class counterfactual primitive — what changes, what stays fixed, which branches are admissible
• Temporal-causal substrate — decay, lag, persistence, reversible vs irreversible transitions, uncertainty over long horizons
• Sensor-to-substrate ingestion — telemetry, lab measurements, LIDAR, climate stations, clinical time series, video
• Multi-scale composition — molecule → cell → organ → organism → institution → planet (scale bridges)
• Human targeting operator — converts expert intent into typed query + invariants + evidence preferences + refusal threshold
• Simulation-scale compute economics — 10⁶ validation steps require caching, invariant precompilation, hierarchical validation, approximate checks with escalation, early stopping
• Substrate governance — source decay, contradiction propagation, revocation, confidence aging, audit trails. If a paper is retracted, a sensor recalibrated, or a legal rule changes, descendants must be re-scored. Not compliance — simulation correctness.

The single benchmark VTC must pass.

35 / 25 / 40.

The allocation is intentionally not 70% R&D even though VTC is the defining bet. Without customer traces, VTC becomes an elegant internal format. Without VTC R&D, the wedges become a loose API business. The portfolio forces the two sides to share substrate artifacts: every Track A product must emit reusable cells or cell-adjacent traces, and every Hyperlab milestone must improve at least one Track A metric.

Per-tier packaging across the AI stack.

Eight wedges and bets to avoid.

VTC + temporal/counterfactual closure.