5
Working papers, v1 — one research program on shared infrastructure
Released
CTC AI Operations · Evaluation & safety research · data-sovereign
Safety, logic, and alignment of frontier AI models — independently evaluated.
A boutique evaluation lab: senior depth instead of crowd, sensitive data processed on-premises under NDA — no cloud-leakage risk. Paired with CTC Research Lab — the applied-research unit whose five workstreams on trustworthy evaluation and multi-agent safety run entirely on sovereign hardware.
~200 tok/s
Local batch judging throughput, single stream, on the sovereign stack
Measured
32 GB
One GPU. Judge, sandbox, and pipeline co-resident — no spill
Measured
≈10×
Cost-reduction target per 1,000 judgments vs. all-cloud baseline
Target · predicted
CODE→CAPITAL
From the engine room of model evaluation into a dedicated wealth- and reinvestment architecture.
§01
What we deliver — CTC AI Operations
Root-level evaluation of how frontier models reason, where they fail, and whether their behaviour holds under pressure — judged by a senior practitioner, not distributed annotation.
01
Root-level model evaluationLogic, correctness, robustness — assessed at the level of reasoning, not surface output.
02
RLHF & preference-data qualityEvaluation of the training signal itself, not only the model's responses.
03
Dataset QA & annotation auditsChecks for consistency, bias, and coverage gaps across labelled data.
04
Agentic-AI safety & red-team auditsBehaviour under adversarial conditions and open-ended tool use.
What an engagement returns
→Scored datasets & calibrated verdictsRubric-based scoring with documented failure modes and edge cases.
→Frozen rubric suitesVersioned and content-hashed — a verdict you can re-derive months later.
→Red-team findings you can act onAdversarial probes scoped to your system, with prioritised, reproducible results.
→Reproducible audit artifactsPinned configurations and hashes — defensible in an audit, not taken on faith.
Engagement flow
Scope→
NDA→
On-prem evaluation→
Findings + artifacts
§02
The dual engine — Operations × Research Lab
Every mandate feeds the method. The method sharpens every mandate. Operations asks whether a given model is safe and correct; the Research Lab asks how that can be measured and proven at all — two separate levels, deliberately built to reinforce each other.
Can frontier and multi-agent AI systems be evaluated trustworthily, reproducibly, and cheaply on sovereign commodity hardware — a single 32 GB GPU rather than a datacentre?
W1Hybrid Evaluation PipelineExecution-grounded local FP4 judging, anchored by a sampled cloud arbiterInfra measured
W2Contamination-Resistant Code EvaluationBenchmarks regenerated from live repositories to resist memorisationPilot measured
W3Three-Tier Agent WorkstationSingle-residency time-multiplexing of three agent tiers in 32 GBInfra measured
W4Multi-Agent Safety EvaluationMeasuring the emergence gap between fleet risk and single-agent safetyAgenda
W5Sovereign Personal AssistantLocal-plus-remote assistant with zero content egressDesign
§03
Infrastructure — the CTC AI Operations stack
Confidentiality is structural — so the stack itself is the trust signal. Sensitive material is processed on-premises, inside a hardened sandbox, on hardware we own: demonstrable data custody, not raw compute.
§04
Why CTC
→
Industry and research under one roofOperational evaluation access that pure academics lack, paired with methodological depth that pure vendors don't offer.
→
Data sovereignty by designOn-premises processing under strict NDA. The value is demonstrable data custody, not raw compute.
→
Senior depth, not crowdEvaluation carried out by a principal practitioner — depth and judgement over distributed volume.
→
Claims you can checkEvery research claim is stated as a falsifiable criterion and labelled measured or predicted — including on this page.
→
NDA disciplineConfidentiality is structural. Client identities stay private; references remain in the abstract.
→
Interdisciplinary founder DNADeep AI craft paired with a finance- and capital-markets mindset — the basis of the "Code to Capital" thesis.
§05
Model & growth thesis
PHASE ICurrent focus
active core practice · entity in formation
AI Operations & Research Lab
The evaluation practice and CTC Research Lab’s five-workstream program. The senior-evaluation core is delivered today by the founder under mandate for leading AI labs — the company is being formed around that practice.
PHASE II
year 2 · in preparation
CTC Advisory
Secure-by-design IT infrastructure, RAG systems, and professional client presences for the DACH mid-market — a capability the company's own brand and web infrastructure already demonstrates.
PHASE III
target state · separate entity
CTC Wealth
Reinvestment and wealth architecture — license-compliant in its own regulated structure.
Code to Capital, Inc. is the umbrella entity in formation. CTC AI Operations is the commercial evaluation practice — carried today by the founder’s senior mandates and being formalised into the entity; CTC Research Lab is the applied-research unit built on the same infrastructure. Later phases are optionality, marked as such.
§06
Leadership
Founder & CEO · Principal Investigator
Marian E. Arenskrieger
Owns the evaluation practice, the Research Lab, the transfer of live mandates into the entity, the structural and tax setup, and the long-term capital architecture.
→arenskrieger.dev · Principal-Investigator profileCo-Founder · VP of Commercial
Commercial & client-facing seat
Active in adjacent AI contracting (voice auditing, AI ad review); builds the go-to-market layer of the Advisory arm and grows into higher-value evaluation work.