The honest opportunity ladder · FSV §6
Three rungs. One sequencing rule.
We separate engineering risk (we know how to build it; it needs calibration and productization) from science risk (new instruments and new ground-truth oracles must be designed and proven) from not Mejepa's job. Each rung names what unlocks the next. The sequencing rule is non-negotiable: rung n+1 cannot be sold until rung n is shipped .
RUNG 1 · GREEN · TODAY
Python AI code verification — pre-production, pilot-fielded
The verified surface today. Predictor + 13 frozen instruments + Docker oracle ground-truth + 491-row exemplar registry + ed25519 witness chain. Sold as founder-fielded per-engagement pilots; not a hosted service. Ship-gate target: Panel A correlation 0.95 stable × 4 rolling windows × all Python mutation cells .
| Capability | Status | FSV ref |
|---|---|---|
| Python AI code verification on SWE-bench Lite (300 × 8 = 2,400 instances) | SHIPPED | §1.1, §1.2 |
| 57 mejepa_* MCP tools — claude code, cursor, windsurf, cline, continue, replit integration | SHIPPED | §1.6 |
| ed25519 + SHAKE-256 signed witness chain (offline-verifiable) | SHIPPED | §1.1, §1.5 |
| Mistake-driven update loop (key-local exact-panel state) | SHIPPED 2026-05-20 | §1.5 |
| Panel B cross-panel triangulation (non-overlapping encoders — Goodhart guard) | BLOCKED P0 (#405) | §1.9 |
| Ship-gate green (Panel A ≥ 0.95 stable × 4 windows + Panel B first measurement) | PENDING | §1.2, §5 |
RUNG 2 · AMBER · NEXT (ENGINEERING RISK)
Verified architecture, calibration and productization remaining
The predictor and instruments already work; what remains is operational and engineering — not new science . We do not sell these as products today; we accept named-design-partner engagements with explicit fences.
| Opportunity | What's required | Unlocks when |
|---|---|---|
| Python AI code verification — hosted SaaS | M4 productization + multi-tenant calibration | Ship-gate green + isolation primitives shipped |
| Multi-language verification (Rust / JS / TS / Go / Java / C / C++ / C# / Ruby / PHP) | Per-language parser-backed operators + per-language Docker-oracle equivalent + per-language ship-gate cells | Python ship-gate green first |
| AI coding agent observability (pre-merge gate + Stop hook as a product) | M4 productization + agent-platform partnerships | M4 product surface defined |
| SOC 2 / ISO / FedRAMP evidence packs from Python verifications | Templating layer over the witness chain | Now-buildable; sold via design partner pilots |
RUNG 3 · GREY · RESEARCH (SCIENCE + ENGINEERING RISK)
New instruments and new oracles must be designed and proven
We do not market against these surfaces today and we decline pilots that conflate them with code verification . The frozen instruments today (AST, control flow, data flow, type graph, Docker tests) do not generalize to non-code corpora. Each entry below would require a parallel instrument-design effort and a domain-specific ground-truth oracle before Mejepa could honestly stamp a verdict on it.
| Opportunity | What's missing | Status |
|---|---|---|
| AI document output verification (citations exist, quotes match) | New frozen instruments for document corpora; new ground-truth oracle (citation/quote matching); new corpus; new label hierarchy | RESEARCH |
| AI underwriting output verification (claim summary vs underlying records) | Same instrument-design gap as document verification | RESEARCH |
| AI clinical output verification (clinical summarization, prior-auth, coding) | Same instrument-design gap + HIPAA-native deployment + clinical instrument design + regulatory bar | RESEARCH |
| AI training-data provenance attestation (paired with Leapable.ai OCR Provenance) | Input-side instruments (Leapable) + output-side instruments (Mejepa) + joint witness chain | RESEARCH |
NOT MEJEPA'S JOB
Surfaces we will not build, even with funding
- AI governance policy authoring (Credo / Holistic territory). Mejepa produces verdict + evidence, not policy.
- AI taste judgments — code elegance, writing quality, performance scores. Q4 surfaces formally retired .
- AI agent generation. Mejepa is not an LLM and not a code-writer.
- Continuous-heuristic scoring of "code quality." Binary doctrine refuses.
The sequencing rule
Rung n+1 cannot be sold until rung n is shipped.
This is the discipline that distinguishes a verifier from a brochure. We do not sell legal verification before legal instruments exist. We do not sell hosted SaaS before multi-tenant isolation is built. We do not sell cross-language before the cross-language ship-gate cells are green. If a competitor sells rung n+1 before shipping rung n, the buyer should ask which frozen instruments cover the domain and which Docker-equivalent oracle returns ground truth — those are the falsifiable questions .