01 / RE-DERIVE
Recover reality from live systems
I distinguish current runtime behavior from stale architecture documents, class names, worker self-reports, and optimistic dashboards.
Andriy Baygerych / founder-operator / applied AI systems
I build machine systems that have to prove what happened next.
I am an independent applied AI systems engineer in Norway and the founder-operator of IFP/SAGE. My work sits between live-system archaeology, agent infrastructure, independent verification, authority containment, and forward-deployed problem solving. I use AI workers aggressively, but I do not let a worker certify its own result.
Local workerQwen / target
External workerGrok / no authority
Visual worldEVOKE / benchmark
AuthoritySAGE / unchanged
Current programmeSAGE_MACHINE_GOVERNOR_001
P0-RLPROVISIONAL / REVIEW HELD
Compute integrationISOLATED BENCHMARK ONLY
Open toROLES / CONTRACTS / AUDITS
Current P0-RL provisional findings
35learning and control mechanisms mapped
21with current firing evidence or current output
0verified L2-L4 external-worker improvement loops
1thin deterministic steering component identified
P0-RL primary finding is provisionally accepted. Artifact review and bounded Codex challenge remain pending as of 24 August 2026.
The work, not the title
No conventional IT diploma. A live system is the credential.
I did not enter software through a standard computer-science career path. I built and operate a complex local AI substrate, direct multiple machine workers, and keep negative results in the record when the evidence does not support the original idea.
02 / GOVERN
Keep machine workers inside host-owned boundaries
Task contracts, accepted context, tools, writable paths, verification depth, closure, and authority remain properties of the host system.
03 / VERIFY
Separate activity from demonstrated improvement
A model answer, passing unit test, stored memory, or completed task report is not treated as proof that a downstream result improved.
04 / REDUCE
Find the smallest correct intervention
The Machine Governor audit rejected a new learning database and new general engine. The missing delta reduced to one deterministic steering consumer.
What I can do for a team
Forward-deployed judgment for agentic systems.
I am strongest where a powerful model meets a real operating environment: incomplete context, ambiguous authority, unreliable self-report, legacy behavior, downstream dependencies, and a customer or operator who needs a result that can survive examination.
Live-system re-derivation
Trace writers, readers, consumers, services, state, and runtime identity before accepting architectural claims.
Agent closure and evals
Define whether an invocation merely ran, was consumed, changed a later attempt, improved a held-out result, or transferred across providers.
Authority and containment
Keep model output, evidence, resource use, value, and authority structurally separate.
Provider-neutral worker integration
Design host-owned task, context, capability, verification, rollback, and closure contracts around replaceable models.
Evidence-grade public surfaces
Expose what is proven, what is attributed, what is provisional, and what remains impossible to verify.
Technical programme leadership
Coordinate specialized AI workers, preserve stop conditions, adjudicate artifacts, and prevent unfinished loops from being called done.
Selected engineering record
Evidence that includes the limits.
The strongest work is not a list of features. It is a series of bounded engineering questions, the evidence used to answer them, and the claims that were deliberately not widened.
Provisional
Machine Governor P0-RL
Mapped 35 learning/control mechanisms and found a substantial internal learning ecology, but zero verified external-worker steering or improvement loops.
Read the case Qualified / boundedDecisionBundleV1 P1
Exact-SHA package verification with 315 passing tests, tamper detection, unsupported-profile refusal, and an explicit refusal to claim the broader record.
Read the case Closed on testnetOVL paid attestation rail
A Base Sepolia x402 payment produced an Ed25519-signed receipt that was independently re-derived and rejected four tamper variants.
Read the case Falsified / heldEnergy prediction truth correction
The earlier performance story did not survive governed evaluation. The predictor was held; the public record and verification method remained.
Read the caseOperating doctrine
Instructions steer the current attempt. Verified closure memory should steer the next attempt.
SAGE already learns internally. The active programme asks a narrower and harder question: can independently verified consequences change how the next unfamiliar machine is tasked, constrained, checked, and trusted?
Agreed compute architecture
Four roles. One authority boundary.
The external model may be more capable. The visual model may generate worlds. The local model may work offline. None of them inherits state authority from capability.
Qwen 27B target
Primary local and offline language worker. Exact package identity and runtime envelope must be verified before installation or routing.
Identity check requiredGrok
High-compute reasoning, research, architecture and coding. Its outputs enter through provenance, admission, verification and closure.
Authority: noneEVOKE
Candidate local visual-world and simulation worker. Installation is isolated; residency, speed, thermals and licensing require measurement.
Benchmark pendingSAGE
Sole owner of accepted state, decisions, admission, independent verification, closure, retained steering and revocation.
Authority unchangedIFP / SAGE
The system is the proof environment.
IFP is the wider platform direction. SAGE is the live local substrate: state, learning, admission, evidence, regulation, operator truth, and current Machine Governor work. External models are replaceable workers; SAGE is intended to remain the owner of state, closure, and authority.
External workerGrok / Qwen / Codex
WorkerDoortask + context + capability
Resultcandidate / action / artifact
Verificationindependent of worker
Track-2 + G1 + KASfacts + shadow + class
steering_updatethin deterministic consumer
Next invocationmaterially changed envelope
Live operator surface
The substrate remains visible.
The embedded surface is served from infinity-folder.org. It is a separate live system surface, not a static claim generated by this portfolio site.
Embedded cross-domain content can be blocked by the live site's frame policy or by a browser extension. The independent link above is the authoritative fallback.
Work with me
Roles, bounded engineering work, and serious technical conversations.
I am open to forward-deployed AI, agent systems, AI evaluations and reliability, applied AI architecture, technical product work, and focused consulting where the output must be independently checkable.
A bounded first engagement
Agentic AI closure and steering audit: inventory claimed memory and feedback paths, classify L0-L4 closure, locate where history fails to change later behavior, identify authority leakage, and propose the smallest deterministic correction with held-out verification.