Marjerla — AI systems architect.
Self-taught AI systems architect — sole author of six systems in four months.
Commits as of 17 Aug. Commits are a floor, counted across 6 version-controlled repositories, this website among them. The repository six includes this website; the system six does not.
Vendor dashboards, 17 Aug 2026: DeepSeek 2,141,131,306 (30d) · GLM 370,269,523 (Aug) · Claude 31.9M. Windows differ; lifetime is higher. Gemini orchestrated, never in the number.
Gemini is orchestrated but was never metered — named in the model roster, never inside the number.
Is any of this real?
Six systems and six version-controlled repositories in four months, this website among them.
How the commit floor was counted, and why it is a floor
Two cells counted the commits independently with a recursive census on the same day and returned the same figure; the site publishes a floor rather than that figure, because a point count stops being true the moment the next commit lands. Five honest counts have been taken in this series and every one of them was higher than the last.
The strongest evidence here is not what was built. It is what was deleted.
The four claims that were deleted, and what replaced them
An adversarial audit was commissioned against this site's own copy, and it removed the strongest-sounding claims on it: a security capability the bytes did not support, a medical claim the product's own frozen in-app banner contradicted, an absolute about data storage that a database schema disproved, and an absolute word about audit chains that the product itself refuses in favour of “tamper-evident”. Each was replaced with the mechanism that survived. That record of retraction is the fastest way to establish that a claim on this site has something behind it.
Two smaller proofs. This site states a test count and never says the word “passing” near it, because no committed run report exists to back the word. A shipped release's published hash still recomputes, with a bill of materials listing the same 440 components at two consecutive versions.
Why that figure is the one that held
It is the one live-state number that held unchanged when this campaign re-checked every one of them.
What it does — and does not do.
Each system's page carries a boundary line stating what is built and proven against what is still ahead. That line is a field in the data, not a footnote, and it does not come off to make a system look finished.
Across all of it:
- The medical system is pre-submission and not FDA-cleared. It is in internal validation, it is not a diagnosis, and its own software ships a frozen notice saying it is not a substitute for clinical judgement.
The four other boundaries: compliance, certifications, production traffic, seams
- No compliance conclusions. Where a regulation applies, this site publishes the mechanism and never the verdict — an engine that emits a statistically valid interval is not a compliance programme, and saying otherwise is the vendor grading his own paper.
- No certifications. No SOC 2, no ISO 27001, no penetration test, no bug bounty, no external code review. The control evidence exists and can be mapped to a framework's own vocabulary on request; the certificate does not exist and is never implied.
- Nothing here serves production traffic for a customer. There are no users, no uptime history and no incident record — so any figure that would require them is marked not assessed rather than estimated.
- Where a capability is an honest seam rather than a wired implementation, it is labelled as a seam. A seam rendered as a capability is the exact defect this project already had to correct once, in public.
What happens if I get hit by a bus?
The bus factor is one. Denying it would fail this question, so: it is converted into engineering properties that can be checked before anything is signed.
- It is written down.
What is written down
Each system carries a continuation handoff document written for the next person rather than for the author, machine-readable operating manuals, and a defect ledger that records what broke, what the root cause was at the bytes, and what changed — including the defects found in the fixes.
- The build is reproducible from the repository.
What makes the build reproducible
Releases ship with a content-hashed build manifest and a CycloneDX software bill of materials, and are code-signed with published hashes that recompute. Someone else with the repository can produce the artifact and prove it is the artifact.
- The release pipeline is a script, not a ritual. The gates are mechanical, and they have demonstrably fired against their own author.
- Source-code escrow can be arranged as a condition of an engagement. It addresses continuity without transferring ownership or control of the code.
The honest limit: no second person holds the keys today, and no third party holds an escrow deposit today. The property claimed here is reconstructibility, not redundancy — redundancy is a commercial arrangement, available but not a current fact.
Why me instead of a firm?
There is nothing here about anybody else's work, and there will not be.
- Decision latency. The person who wrote the code answers the question, in the same conversation, without an account layer in between.
- Clean authorship. Sole authorship means undisputed assignment: no contractor chain to reconstruct, no third-party code under a conflicting licence.
- Cost structure. No margin stack, no bench to keep busy, no minimum engagement shaped around a staffing model.
And the same shape is the risk: one person is one person. The continuity properties above are what stand between that fact and a stalled project.
Bounds, and the declared bias
The grades on this page were produced by an audit the subject commissioned against his own work. That is a conflict of interest, and the correct response is to publish the conflict rather than claim it away.
All seven declared biases, in the assessment's own words
- Assessor is an Anthropic model; parts of the assessed work were implemented by Claude-family agents under the operator's direction.
- Strongest available audit trail is self-commissioned (adversarial in content, in-house in origin); no third-party review exists.
- Audit reports were read before the code (anchoring), mitigated by independent byte spot-checks including counter-evidence sites.
- Bands are calibrated to professional teams; the solo-operator-plus-AI-agents modality maps imperfectly onto team-based criteria; scale mismatches are declared, never banked.
- Four-month evidence window; no production traffic, users, or incident history was observable.
- Neither the v2 derivation nor this re-derivation ran a test suite, a build or a server, and neither rendered an application page; both re-verified the runtime-provenance artifacts, fact-ledger lines and shipped strings they cite, and inherited the rest of the campaigns' audited record.
- Direction-of-travel: v2 was a post-audit re-grade anchored to the campaign's own criteria file — six scores moved down, one held, none rose; this run widened the evidence set to all six version-controlled repositories and moved nothing, in either direction. Every movement, and every non-movement, is derived from a published band table rather than from the audit's mood.
Eighteen further dimensions are unassessable from a codebase — collaboration, mentorship, hiring, production reliability, incident response, customer impact, user research, independent security validation and peer-relative standing among them. They are named in the assessment record and they are not graded here, because a grade without evidence is the thing this whole apparatus exists to prevent.
The assessment
CLAUDE CODE / LEGION ASSESSMENT · OPUS 5 · 2026-08-17
4 months · solo + AI agents · no external review
- Software EngineeringB+ · 7/10good
Parity-tested numeric cores, fail-closed pipelines, a harness that tests the test system; only one suite ever witnessed green, dead aspirational modules remain in-tree, and this website itself ships no test and no CI.
Dropbox IC4 · Monzo L5
- Systems ArchitectureB · 6/10good
Six systems share one enforced grammar — refusal paths, tamper-evident chains, operator gates; unwired seams (a described boundary that gates nothing) and zero production scale bound it.
iSAQB CPSA-A · AWS SAP-C02
- AI Systems EngineeringC+ · 5/10average
Refusal and grounding engineering is real and has held — zero ungated answers across every training pair to date; on the MLOps maturity ladder itself, no eval harness and no experiment tracking: average.
Microsoft MLOps · Google MLOps
- InnovationB · 6/10good
Pre-registered hash-pinned validation and a ship gate red on its own product — honesty process innovations in daily use; most products are not yet introduced, and the protocol exists because published figures once had nothing behind them.
Eurostat CIS · Oslo Manual
- Workflow & OrchestrationB · 6/10good
A locked multi-agent build law with enforcement that demonstrably fired — 207 hash-pinned exploit replays on every push, greps that caught their own author; DORA's delivery half is not assessed: nothing runs in production.
DORA capabilities · SLSA Build L1
- Security EngineeringB · 6/10good
2-of-N approval, dead-man quarantine, tamper-evident chains, hybrid PQC on Meridian North, ten response headers and no third-party surface here; the worst overclaims in the portfolio were security claims, and no external test exists.
OWASP SAMM v2 · NIST SSDF 800-218
- DesignC+ · 5/10average
A measured accessibility floor on the audited pages and a token design system with propagation law; graded from bytes and recorded audits — no page was rendered by the grader, and no full WCAG conformance audit exists.
WCAG 2.2 · NN/g UX maturity
Seven separate measures — never averaged into one number.
Re-derived (v2, 17 Aug 2026) against the campaign's own criteria file: whole-point scores from its published band tables — Dropbox/Monzo ladders, iSAQB CPSA-A, AWS SAP-C02, Google/Microsoft MLOps maturity, Eurostat CIS, DORA capabilities, SLSA, OWASP SAMM, NIST SSDF, WCAG 2.2, NN/g UX maturity. Re-run the same day across all six version-controlled repositories — this website among them — with no score moved: the site adds shipping discipline and security defaults on one side and no test, no CI and no code review on the other, and under the band rule neither clears a content that was not already cleared. Evidence: a 10-cell Legion audit, an 8-cell frontier recon, and independent byte re-verifications. No composite — unlike sticks are never averaged. Claims discipline is priced in the counter-evidence, not graded as a dimension.
Scores derived on Fable 5 (v2, 17 Aug) and carried unchanged; the re-derivation across all six version-controlled repositories ran on Opus 5 this run — no score moved. Both are runtime-pinned, not self-declared.
Fund the next system
This work is self-funded. Contributions pay for what arrives with an invoice attached — compute, code-signing certificates, hardware, and the hours spent auditing what has already shipped. The next system then gets built without a client deciding what it is.
- Systems
- 06
- shipped, each stating its own boundary
- Commits
- 880+
- counted across 6 version-controlled repositories, this website among them
- Tokens
- 2.54B+
- three vendor dashboards read 17 Aug 2026; each vendor's own window
Where it goes
- Compute — the model spend, and the largest line by a distance.
- Code-signing certificates, so a desktop build installs without a warning that teaches users to ignore warnings.
- Hardware — local inference is the architecture, and that is a machine someone buys.
- Audit time — the hours spent re-checking claims that already shipped.
What it does not buy
- No influence over what gets built, or in what order.
- No priority support, and no support commitment of any kind.
- No access to anything that is access-restricted.
- No ownership of anything, and no financial interest in anything.
- No place in a queue, and no promise of a future thing.
Terms, in plain words
- This is a contribution — a gift, acknowledged. It is not a purchase and not a pre-order, and it buys no deliverable.
- Contributions are not tax-deductible.
- Contributions are received by an individual, not by a company or a registered organisation.
- Refunds on request within 30 days, no reason needed — write to the address in the footer.
- Payment is handled by Stripe, on Stripe's own page. This site never sees, touches or stores card details, and no account is created for you.
There is no membership ladder, no recurring option and no tiers beyond the amounts above. Rungs that would deliver something — a mailed record, a named backers list — are not built, and a benefit that does not exist is a promise waiting to be broken. If they are built, they will appear here with what they deliver stated before anyone pays.
Written down plainly: this page exists before the audience does. The social controls on this site are not wired to a profile yet, and a funding page with nothing upstream of it converts approximately nobody. It ships because it has to exist first, not because it is expected to earn this quarter.