Confidential · Hypernym Research Arc

This document contains forward-looking research, primitive specifications, panel deliberations, and strategic positioning for products and architectures that have not yet shipped. It is shared under explicit understanding that recipients hold its contents in confidence and do not reproduce, redistribute, or summarize externally without written consent.

Round R16 · Compound Vision & Execution · 2026-05-10

01R16

Round Sixteen

Compound vision & execution

VTC closures · Modulum product immediacy · local inference packaging · world model continuation · core vision of verifiable computation over reality.

5 streams · 5/6 panels · 2 rounds

02

Headline

The primitive holds.
The spec is partially closed.

03Verdict

Per-stream final verdicts

Where each stream actually lands.

Stream	Verdict	Ship-blocking work
A — VTC algebra	sound primitive · partial adversarial spec	8 attack-vector mitigations + manifold-aware revert sub-spec
B — Product immediacy	sound sequence · partial GA contract	6 abuse-vector mitigations + tiered SLA + work-unit-plus-premium pricing
C — Local inference	sound	MLX M5 throughput gate · refusal-degradation banner
D — World model	partial — R17 carry-forward	LAP + DomainBridge + RedactionBridge + refusal benchmarks
E — Core vision	sound	Maturity-stage communication discipline

04Vision

“

Hypernym builds verifiable computation over reality.

Public framing locked · Stream E

05Convergence

What 5/5 panels agreed on

Eight unanimous commits.

VTC primitive holds. Typed state · transition · invariants · calibrated confidence · provenance hash.
gear_state + GearBridge added to state_before. M5 mask compatibility is a closure requirement.
integration_step for continuous-domain VTCs (PDE-governed substrates).
Three-tier billing — decisive 100%, refused/underdetermined 30%, malformed 0%.
Modulum Edge runtime mode replaces standalone "Hypernym Local". No SKU sprawl.
90-day execution: Verify GA wk 4, Modulum Solo wk 6, Legal Endpoint design-partner beta wk 10.
"Verifiable computation over reality" public framing.
Cost-per-accepted-transition as the unifying executive KPI across all 5 streams.

06

Stream A

Codex named eight attack vectors.
Each gets a mitigation.

07Stream A

Adversarial closures

Eight attacks. Eight typed mitigations.

#	Attack	Mitigation
1	Revert-then-attest race	Read lock on dependency frontier · `trace_epoch` + `invalidation_watermark` in receipt
2	Query-relative load-bearing ambiguity	`load_bearing` as relation, not cell-tag · cached at trace-time · revalidated at query
3	Counterfactual branch-confusion	`causal_model_id` hash + `CausalBridge` · write-set agreement insufficient
4	Observational-to-interventional laundering	`crossover_admissibility` via Pearl back-door / front-door / instrumental criteria
5	Refusal laundering	Default `strict_stop` · refused never demotes to underdetermined · soft mode non-attestable
6	ScaleBridge overclaiming	Per-domain density floors · KL-divergence test · uncertainty propagation
7	M5 mask mismatch	`gear_state.mask_signature` compatibility OR declared `GearBridge`
8	Cross-tenant attestation replay	Attestation scope binds tenant + policy + ontology + jurisdiction + ACL root + validator authority

Per memory rule: Codex iterates more, catches HIGH bugs others miss. R2 confirms the pattern — Codex was the only model to downgrade verdicts after cross-pollination. Synthesis adopts every Codex attack-mitigation pair as required pre-GA work.

08Stream A — algebra

VTC state envelope (R16 commit)

Closed under refusal.

Op	Shape	What it does
apply	VTC → VTC \| refusal	Execute and validate one transition. Invariants checked. Commitment hash sealed.
branch	VTC, Δ → VTC	Counterfactual. Same parent state, different changed/held-fixed sets. Causal-model-ID required.
merge	[VTC] → VTC \| refusal	Defaults to refusal if held-fixed disagreements OR causal-model mismatch.
revert	VTC → trace	Invalidates descendants if load-bearing. Manifold-aware retrieval blocks stale embeddings.
attest	VTC → signed VTC	Signs cell + sources + verdict + tenant/policy/ontology/jurisdiction binding.
query	trace → verdict	Extracts supported / contradicted / underdetermined / refused. Validates dependency proof.

09

Stream B

Verify GA. Modulum Solo. Legal.
Ninety days.

10Stream B

90-day shipping cadence

Three products. One corpus.

i. Wk 4

Omnifact Verify GA

3-tier billing · domain-tier multipliers ($0.02 / $0.05 / $0.10) · work-unit + verdict-premium · audit-token canonical-claim binding · evidence-coverage gate.

ii. Wk 6

Modulum Solo

$100/mo flat → 1M tokens/mo, 100 verified claims/day, 1GB persistent memory. Sales pitch: refusal-correctness, not cost. Single-developer purchase.

iii. Wk 10

Legal Endpoint

Design-partner beta. TrustFoundry $5K floor → 6mo ramp to $20K. 50K-VTC seed corpus. Refusal taxonomy specific to legal procedural claims.

Killed this quarter: generic Modulum Router · Forge OS Solo · Persistent Memory API · IDE Magic. Per R15 Track C: ship weak-moat wedges only when they feed strong-moat assets.

11Stream B — abuse vectors

GA contract abuse mitigations (Codex R2)

Six product-side blockers.

#	Vector	Mitigation
1	Pricing manipulation via refusal retries	Work-unit floor · retry-link billing groups for near-duplicate claims in time window
2	Refusal-bounty abuse	Per-company auth aggregation · organic-traffic gate · dedup-yield gate · bounded per account
3	Audit-token forgery / laundering	Receipt binds canonical claim hash + AST + evidence bundle + policy + verdict + replay window
4	Citation-shadowing attack	Verify reports `evidence_coverage` · adversarial retrieval (search contradictors)
5	Schema drift product↔corpus	Promotion pipeline ObservedRecord → CandidateVTC → VerifiedVTC · versioned · rejection reasons
6	Cross-customer replay leakage	Selective disclosure modes: full · redacted · commitment proof · regulator escrow

12Stream B — SLA

Tiered SLA (Codex R2)

Don't promise audit-grade at router latency.

fast_check

p95 ≤ 500ms

Non-attestable. For draft-quality validation in conversational UX.

verify

p95 2–5s

Decisive verdict. Attestable. Includes evidence retrieval + claim decomposition + adversarial check.

attest

p95 5–30s

Federation-grade receipt. Cross-bridge composition. Regulatory replay.

13

Stream C

Modulum Edge.
A runtime mode, not a SKU.

14Stream C

Local inference packaging

Edge mode. Not a product.

Hypernym is the company. Modulum is the product. Local is a deployment mode — cloud, on-prem, edge. SKU sprawl avoided. Substrate sync stays in the cloud. Differentiation is substrate-grounded inference, not "model running offline."

MLX M5 throughput gate. ≥1.5× Ollama baseline, OR Edge ships CUDA-only first.
Refusal-correctness gate. Within 5pp of cloud Modulum on the same workload.
Refusal-degradation banner. Customers running biomed/legal on Edge see "calibration −X% vs cloud."
Local attestation paradox resolved. Local receipts = "private receipts." Federation-grade requires sync. Visible in API and UI.

15

Stream D

Federated VTC graph.
R17 must close LAP.

16Stream D

World model continuation

Year-1 dense single-domain. Federation deferred to R17.

i.

Federated VTC graph ≤10% capital

Year-1 = single-tenant dense biomedical/legal graph. Designed as the first shard of the federated graph. Density precondition: per-domain calibrated floor.

ii.

PoVT staking + Lemma royalties

Anti-poisoning: stake bond to commit Lemma; reverted/contradicted slashes. Anti-hoarding: micro-royalty per third-party compose.

iii.

Substrate Manifold Index

Continuous-vector projection for O(1) LLM-critical-path retrieval. Invalidation-aware via attestation-tag cross-check.

R17 carry-forward (cannot federate without): Lemma Arbitration Protocol · DomainBridge cross-domain attestation · RedactionBridge for regulated industries · published refusal-correctness ground-truth benchmarks · continuous-domain integration_step calibration tables · per-domain density floors · validation-vs-inference cost margin model.

17Vision

Stream E — three maturity stages

What we sell now. What we don't yet.

i.

Decision support

Verify · Modulum Solo · Legal Endpoint. Customer holds the decision; Hypernym attests claim validity. Production-ready post-GA.

ii.

Attestable recommendation

System proposes; customer ratifies; receipt commits. Production-ready post-GA. The substrate flywheel.

iii.

Autonomous execution

NOT in core pitch until refusal · do-calculus · M5 gear · attestation closures pass benchmarks. Long-term arc, not Q3 2026 marketing.

18Vision — non-negotiables

Eight non-negotiables

The principles that cannot move.

1

Substrate ownership

Customer corpus belongs to customer. Lemmas commit to federation only by stake.

2

Refusal as first-class output

Refusal is a feature, not a defect. Sales motion frames "we refuse, they hallucinate."

3

Calibrated confidence

ECE published, not scalar. Calibration class on every cell.

4

Provenance hashing

Deterministic audit replay. Hash chains over inputs / outputs / validation trace.

5

Audit replay

90-day deterministic minimum. Required by FDA / courts / regulators.

6

Cross-domain bridges

Typed, attested. Gear · Scale · Causal · Domain · Refusal · Redaction.

7

Invariant preservation

Hard constraints cannot be violated. Compose() refuses on invariant violation.

8

Economic closure

Anti-poisoning + anti-hoarding economics. Otherwise federation degrades to centralized arbitrage.

19KPI

Unifying KPI (Codex R1+R2)

Cost per accepted transition under audit.

Every algebraic refinement, product wedge, runtime mode, world-model bet judged by the same scalar. CPAT collapses Streams A/B/C/D into one optimization target. Refusal-correctness is the orthogonal second-axis KPI — refusals don't count toward "accepted" but count toward calibration moat.

CPAT = (
  substrate_retrieval_cost +
  decomposition_cost +
  validation_cost +
  bridge_check_cost +
  federation_hop_cost +
  attestation_cost
) / accepted_transitions_count

// per dispatch · per product surface
// reported to Cost-Tracker (paperclip)

Standing principle

The corpus is the moat.

Every product Hypernym ships either produces VTCs, consumes VTCs, or composes them. Anything that doesn't is cash extraction or a distribution experiment — not core strategy.

R16 closed · R17 carry-forward queued