How fast can this deploy against our existing model?

Three to ten working days for the reference implementation. No retraining required — the guard wraps your existing inference endpoint, re-embeds every output, and enforces a per-modality cosine threshold against an enterprise-defined constellation.

Does this require us to retrain?

No. Constellation Guard is a runtime wrap. You keep your existing model, your existing infrastructure, and your existing deployment posture. The guard adds typically 40–120ms per output for embedder forward passes plus cosine comparison.

What audit formats are supported?

W3C PROV-JSON (native), JSONL, and CSV. Every audit row contains the cosine trace, embedder weight SHA-256, centroid SHA-256, threshold set, decision, attempt count, and rejection reason if rejected.

Will this add latency?

Typically 40–120ms per output for embedder forward passes, plus negligible cosine-comparison cost. Distillation to a lighter guard is future work for low-latency contexts.

Can we run air-gapped?

Yes. On-prem container or bare-metal. The reference data, the computed constellation, and any LoRA weights produced never leave the customer environment.

HOME · APPS · APP 9

APP 9 · CONSTELLATION GUARD

Deploys in days against any frontier model. No retraining. Per-output cosine verification with human-readable rejection reasons.

Per-output cosine verification with a human-readable rejection reason on every block — the CISO piece of the AI-risk program that has to pass audit, shipped as a runtime wrap on your existing OpenAI / Anthropic / Mistral / in-house model. Every output re-embedded through a 13-embedder panel, scored against an enterprise-defined centroid, accepted or rejected with a W3C PROV-JSON trail that falls out of the architecture — not a consultancy engagement. This is the substrate that obsoletes the bolt-on AI-governance cohort (Credo AI, Fiddler, Arthur, Lakera, Patronus — ~$5–10B combined equity, 12–24 month absorption window); the CFO number attached to it is $10–50M/yr per large bank in avoided compliance consultancy, because audit is now a property of the model, not a deliverable on a Big-4 SOW.

ARCHITECTURALLY COMPLETE · Prompt-injection resistance demonstrated on prior run; camera-ready retrain pending

Schedule 48-hr POC Book 20-min briefing Strategic inquiry

WHO THIS IS FOR

For the F500 CISO / Deputy CISO / BISO whose pilot has stalled behind EU AI Act, SEC, OCC, or internal risk review.

PROOF STACK

Every claim tagged.

PROMPT-INJECTION RESISTANCE

Arithmetic Layer-2 decoder — cannot be jailbroken by prompting (ARCHITECTURALLY COMPLETE)

COSINE-AUDIT COVERAGE

Every output re-embedded through a 13-embedder panel, rejected off-manifold with a human-readable reason

AUDIT FORMAT

W3C PROV-JSON, JSONL, CSV

DEPLOYMENT MODES

SaaS proxy / on-prem container / air-gapped bare-metal

TECHNICAL DEPTH

How it composes.

Layer 1 — LoRA (learned — vocabulary, rhythm, narrative flow)
Layer 2 — Arithmetic-constrained logit decoder (deterministic token-time filter)
Layer 3 — 13-embedder constellation guard (cosine vs frozen centroid)
Severity tiers (configurable per modality): WARN 0.80 / STRONG 0.90 / HARD 0.95
Runtime deliverables: acceptable_constellation.bin + runtime + W3C PROV-JSON audit log

MARKETS UNLOCKED

What opens downstream.

$150–400B

Regulated-enterprise AI

$52–139B by 2030–2034

Agentic AI in regulated verticals

$10–30B ARR

Post-training-as-a-service

FREQUENTLY ASKED

Questions from every buyer.