Six systems. Every count below is a dated, recorded checkpoint — separated into build evidence, user evidence and business evidence, with no claim carried across those lines. Where a number was not independently rerun, the receipt says so.
01 / Recorded checkpoint: 2026-08-02Operating locally
SignalLayer
The strongest local engineering record: a proof-gated AI-search operations system with broad test, route and rendered-QA coverage.
Build evidence
827 tests recorded
The dated checkpoint also records 170/170 routes, 16/16 rendered QA captures and a 258-file export-safety check. These figures have not been independently rerun for this page.
Authorship boundary
Johan-led product direction
Johan defined the product model, evidence gates, approval workflow and review criteria. AI agents performed substantial implementation work.
Outcome status
Internal evidence only
No external customer workflow, production deployment, ranking lift, revenue or commercial-validation claim is made.
Local source record
Checkpointsignallayer/RESUME_HERE.md
Build surface255 core Python modules / 153 test modules found in the archived project
A security gateway prototype with a real TypeScript engine and adversarial tests, paired with an Electron interface that currently uses mock-backed product flows.
Build evidence
Engine and tests exist
The local project contains a proxy implementation, injection-bypass tests, architecture and security reviews, 24 engine-test files and seven application-test files.
Authorship boundary
Direction, not solo coding
Johan owns the threat model, product definition, interface direction and acceptance criteria. Claude agents performed substantial implementation.
Outcome status
No external audit
No production deployment, human penetration test, customer operation, signing, notarization or commercial outcome is claimed.
A broad full-stack build combining a deterministic eligibility engine, serverless payments, monitoring and reusable country modules.
Build evidence
25 country modules
The local record contains 74 generated country and audit pages, a rule engine, Stripe flow, Cloudflare configuration and scheduled monitoring components.
Authorship boundary
Founder and product direction
Johan defined the product, rule structure, customer journey and review workflow. Research, design and implementation were AI-assisted.
Outcome status
Zero revenue recorded
The July external-validation brief records near-zero traffic and zero revenue. No traction, customer or commercial-success claim is made.
The most credible result is a rejection: a strong training result degraded below viability out of sample and was classified as curve-fitted.
Build evidence
4.56 to 0.95 PF
The local master summary records the training and out-of-sample degradation. The project also contains a Python harness, smoke tests, trade exports and diagnostic reports.
Authorship boundary
Validation-system direction
Johan directed experiment design, risk rules and promotion gates. AI assisted implementation. Licensed or published strategy logic must remain attributed upstream.
Outcome status
No performance claim
User-maintained account records are not treated as broker-verified evidence. No expected-return, profitable-bot or deployable-edge claim is made.
Local source record
Research summaryMES_ORB_Project_1/00_MASTER_SUMMARY.txt
A concrete bridge between Johan's photographic practice and software: the tool was executed during the audit and produced the exact output shown in the case study.
Build evidence
Successful local run
The 4,279-line Python tool loaded the source image and generated a worn-Polaroid, matte-warm result with grain and dust. Five processing modes are documented.
Authorship boundary
Creative workflow direction
The tool was built for Johan with AI-assisted implementation. Johan directed the creative behavior, selected the run and reviewed the resulting image.
Outcome status
Owner-tested only
No external user, packaged distribution, production adoption, client-time saving or revenue claim is made.
Hugr, a customizable operating layer its creator architected and runs daily, now available for purchase and customization, has an unusual property: it is built to be wrong safely. Every memory can be corrected or forgotten through audited, append-only mechanisms, and every routing decision is logged with measured latency from live sessions.
Build evidence
1,341 tests recorded
The dated checkpoint records 1,174 green Python tests and 167 green Worker tests, the five shipped V3 components (provider-failure tracking, headless-skill coverage, schedule hardening, outcome credit, agent mailbox), and a correction/forgetting system (supersession-aware retrieval, tombstone registry with retraction, audited cloud forget, conflict dedup). The live routing log holds 646 advisory events with a 127ms median and 253ms p95 latency. Test counts are the repository's own recorded runs and have not been independently rerun for this page.
Authorship boundary
Directed two-agent build
Johan owns the architecture, safety policy and acceptance gates. One AI session implements; a second, independent AI session reviews and re-derives every claim before it is recorded. Disagreements and revised acceptance bars are preserved in the record, not smoothed over.
Operating record
In daily use, owner-operated
Runs inside the owner's daily Claude sessions as Hugr. Subscription AI usage is explicitly recorded as unmeterable; only per-call priced operations are metered.
Swedish-born creative director, filmmaker and photographer — raised across thirteen countries. 20+ years across MTV, VH1, AMC and 180LA — Oscar-shortlisted and an eight-time Promax and Webby Award winner. Known for double exposures, composite worlds and a tactile, cinematic use of light.
