Project summary

SEVERANT is a 7-layer heterogeneous AI architecture where safety is an architectural property, not a training outcome. Its core differentiator is L6, an ethical constraint layer that does not train, is formally verified in Lean 4, physically encoded into write-locked Phase Change Memory, and active throughout the training pipeline of every other layer.

A working prototype (SEVERANT-0) is operational. The formal verification suite is complete. The L2 causal knowledge base stands at 3.9 million entries. This grant funds the next critical phase.

Full documentation: github.com/EvangaleKTV/SEVERANT

What are this project's goals? How will you achieve them?

12-month milestones:
Months 1-3: L2 causal knowledge base completed to 10 million entries
Months 4-6: L2 training initiated with L6 active throughout every training step
Months 7-9: L2 training validated and confirmed stable under L6 constraint
Months 10-12: L3 hierarchical memory architecture prototype operational
Each milestone is a concrete proof point. L2 training is the first demonstration that formally verified ethical constraints can shape trained weights at meaningful scale, not just in a prototype.

How will this funding be used?

Full budget breakdown at $320,000:
Researcher stipend, 24 months (incl. tax) - $100,000
GCP compute, 12 months - $8,000
GPU fine-tuning, L2 construction - $8,000
NullForm entity registration - $8,000
Research equipment - $25,000
Conference and community engagement - $15,000
Accounting and operational overhead - $11,000
PCT international patent filing - $55,000
Trademark registration, key jurisdictions - $32,000
Legal counsel, IP and entity structure - $28,000
Contingency - $30,000

Compute costs are materially reduced by AI-assisted pipeline tooling. Claude is used as a validation and structuring layer in L2 knowledge base construction, replacing annotation overhead that would otherwise require paid labour or significantly more compute.

Minimum funding ($30,000): Covers 6-month researcher stipend and compute. L2 knowledge base construction is completed but L2 training is deferred pending additional funding. Progress is preserved and the project continues at reduced pace.

Full funding ($320,000): Complete research infrastructure funded without compromise. Stipend extends to 24 months, IP protection secured across key jurisdictions, hardware prototyping toward L3 begins within the grant period. The 3-year roadmap executes without interruption.

Who is on your team? What's your track record on similar projects?

Sole researcher: Evangale Jooste, Cape Town, South Africa.
Built independently and without external funding:
1. SEVERANT-0 operational, hosted on GCP, accessible via OpenRouter
2. L2 causal knowledge base at 3.9 million causally structured entries
3. L6 formal verification suite complete: 21 predicates verified in Lean 4, adversarial suite 19/19 pass, build status clean
4. Full architecture, roadmap, and research documentation publicly available
No institutional affiliation. No lab. No team. The above was built by one person.

What are the most likely causes and outcomes if this project fails?

The most likely failure mode is funding running out before the L2 training milestone is reached, forcing the project into a reduced-pace holding pattern on personal resources. The architecture and verification work already completed would not be lost. SEVERANT-0 would remain operational. Progress would slow significantly but not stop.

The second failure mode is a technical problem in the L2 causal construction pipeline at scale. This is mitigated by the fact that the pipeline is already operational at 3.9 million entries and the primary remaining work is scaling what already works.

How much money have you raised in the last 12 months, and from where?

$0 external funding raised to date. The project has been entirely self-funded. Active applications currently pending:

Foresight Institute AI for Safety Node: $100,000 requested, submitted April 2026, decision expected June 2026

Long-Term Future Fund: $80,000 requested, submitted April 2026, decision expected within 21 days