Economist and data analyst · Neiva, Huila, Colombia
jhonstevenalvarezruiz.vercel.app · LinkedIn · humanizar.tech · HegelFlow · Contribution audit · Achievement audit · Bloquitos · Type system
Economist and data analyst. Together with Steven Vallejo Ortiz I build the system that turns scattered AI agents into a company that actually operates.
I work with three materials that academic convention keeps apart for no good reason: the critique of political economy, the architecture of multi-agent systems, and the logic that lets you say when a set of parts constitutes a whole and when it remains a heap. These are not three fields. They are one problem approached with three instruments.
humanizar.tech — development companies, built from agents
An agent is not a company. The claim is neither rhetorical nor quantitative: it is categorial. Adding agents no more produces a company than adding coins produces capital, or adding individuals produces a body politic. What constitutes a company is a determinate form of organization — division of labour, attribution of responsibility, governance of capacity, accounting for spend — and that form is contained in none of its parts.
The opposite error has a name. It is the hypostatizing of what exists only as a relation: treating a structure as though it were a thing. Gustavo Bueno dismantled it in Ensayo sobre las categorías de la economía política; Luis Carlos Martín Jiménez carries it through in El mito del capitalismo, where he shows that "capitalism" names a myth precisely because it performs that hypostasis. Today's AI market repeats the operation on its own ground: it calls a process an "agent" and expects an organization to emerge from multiplying it.
CAUCE V3 begins by denying that premise. It does not orchestrate agents that already exist — it produces the teams and imposes the form that makes them accountable. From two to N specialized teams, governed from a single control plane.
| Durable routing | Expiring reservations and automatic redistribution of work when a process dies. No task is orphaned because its executor went down. |
| Dead letters | Failed operations retain the complete message body. A failure that leaves no trace is not a failure — it is a silent loss. |
| Multi-harness | Four distinct engines under a single protocol contract. The engine is interchangeable; the contract is not. |
| Typed protocols | Separate fields for response, delegation, notification and artifact. A confused type is a decision nobody made. |
| Identity and isolation | Per-client certificates and separation at the storage layer. Channels do not talk across tenants — not by oversight, by design. |
| Human gates | Deployment, secret access, deletion and spend each require a person. The system does not decide for you. |
The method is the one Juan Íñigo Carrera demands of any critique that means to be serious: not to represent the process from outside, but to reproduce its movement. A control plane does not describe the organization — it exercises it, which is why it can be audited.
And the architecture is Hegelian in a precise, non-decorative sense. The company is not presupposed as given: it starts from the minimal determination — a request in a channel — and is left to develop through its mediations into the concrete result. This is the requirement Stephen Houlgate defends in The Opening of Hegel's Logic: a presuppositionless beginning, whose categories are not imported from outside but generated in the development itself.
| Project | What can be verified now |
|---|---|
| HegelFlow · production | Multi-user Scrum and Kanban workspace with accounts, audit trails and PostgreSQL isolation. Its public build shows the real dashboard and reports over explicitly fictional data; CI passes 44 unit tests and 11 PostgreSQL integration cases on Next.js 16.3.3. |
| gh-before-you-contribute · v1.2.0 | Read-only GitHub CLI extension and Action. Schema v1 keeps bounded, typed and source-attributed evidence; v1.2.0 adds validated native verdict, policy_verdict and issue_verdict Action outputs from one audit while preserving requested text/JSON and strict exit semantics. Its 62 deterministic regressions, hosted composite-Action check, CI and both CodeQL analyses succeeded; incomplete or malformed evidence still fails closed. |
| gh-achievement-audit · v0.1.0 | Read-only GitHub CLI extension that separates achievements visibly rendered by GitHub from supporting public evidence without guessing unpublished criteria. Its 54 deterministic tests on Python 3.10 and 3.14, public API smoke test and CodeQL analyses for Actions and Python succeeded. The producer excludes private events before report assembly; schema v1 and an independent semantic validator reject incomplete or malformed reports. |
| Bloquitos · v1.1.0 · play | Its browser/game core has no runtime package dependency; desktop packaging uses Electron. The release publishes six Windows, Linux and macOS assets with SHA-256 checksums; 67 tests cover game and release invariants. |
Every link above points to running code, a reproducible check, or an immutable release. None is a roadmap claim.
| Repository at measured revision | Source | Test | Test / source | Reachable commits | Workflow files |
|---|---|---|---|---|---|
| automatizacion-evidencias-adso @ a66a12dd2655 | 8,405 | 341 | 0.04 | 8 | 1 |
| bloquitos @ d2dfb32b63d2 | 3,011 | 740 | 0.25 | 10 | 2 |
| colmat-x-automation @ f5a34b73a6ec | 2,112 | 1,294 | 0.61 | 1 | 1 |
| gh-achievement-audit @ 8c9da850a2ff | 1,292 | 1,520 | 1.18 | 2 | 1 |
| orquesta-ia @ be4d50a0316b | 9,263 | 5,874 | 0.63 | 27 | 1 |
| sincategorematico-bot @ 3fe8b6342874 | 1,133 | 134 | 0.12 | 4 | 0 |
| Aggregate | 25,216 | 9,903 | 0.39 | 52 | 6 |
This block and its SVG are generated from the same measured report; CI rejects any altered, missing or duplicated published value.
Every raw count above is computed from six selected public repositories at the
exact commits recorded in
metricas/repositorios.json. “Source” and “test” mean
physical lines in regular .py, .js, .mjs and .sh blobs only; HTML, CSS,
templates, documentation and assets are outside this measure. Files under test/,
tests/, __tests__/ or pruebas/, and filenames using conventional test or
spec tokens, count as test. “Workflow files” counts direct regular .yml and
.yaml blobs under .github/workflows/. The generated Bloquitos browser bundle is
explicitly excluded as a duplicate of its maintained modules, and that exclusion is
versioned in the manifest. “Reachable commits” traverses the complete parent graph
from each pinned revision, regardless of author. Schema v1 fixes full SHA-1 commit
IDs. Only regular-file blobs contribute to file and line counts; symlinks and
gitlinks/submodules are excluded, and the six pinned trees contain zero gitlinks.
The verifier reads Git blobs rather than mutable worktrees, rejects shallow history,
disables lazy fetching, and fails closed when an object required for the count is
missing. CI recreates the local clones without checking out their worktrees,
runs the regression suite, and accepts only canonical github.com/SirHegel/...
sources whose repository name matches the manifest. It requires both the
measurements and every published README/SVG value to reproduce exactly. Base counts
are exact. Test/source ratios are rounded half up to the nearest hundredth; language
shares use deterministic integer largest-remainder allocation with an alphabetical
tie-break.
Security is not a layer added once the system already works. It is the constraint that decides what the system may do, and therefore comes first.
No credential enters a repository. Every project ships in two modes: the public version, without secrets, and the real configuration, which never leaves the machine. Where it matters, a scanner runs before every commit and aborts on a credential pattern or a local path. Continuous integration rejects the push if anything slips through.
Deleting a secret from a file does not remove it. It stays in history. The only valid repair is to rewrite history and rotate the credential. The second is not optional; the first without the second is theatre.
What I claim is tested. If a system says it works offline, a test cuts the network and verifies it. If it says it will not corrupt its state, a simulation of thousands of cycles checks the invariant. A claim with no test behind it is a promise, and promises do not deploy.
I work economic theory from the tradition that refuses to take equilibrium as its starting point, because equilibrium is not a fact: it is an assumption that decides in advance what the model will be able to see.
From the School of Oviedo I take the categorial apparatus: political economy is not ordered by magnitudes but by categories, and confusing the plane of magnitudes with the plane of categories produces operative myths — "capitalism", "the market", "technology" — that function as the subjects of sentences in which there is no subject at all.
From Juan Íñigo Carrera I take the demand on method: critique does not consist in setting one model against another, but in reproducing in thought the real movement of the form under critique. Whoever only represents, describes. Whoever reproduces, explains.
Selected for distinct engineering constraints; the order is editorial, not a technical ranking.
| Repository | Why it is here | Engineering evidence |
|---|---|---|
| automatizacion-evidencias-adso | An engine that regenerates, validates and audits deliverables. It verifies office documents by unzipping them and reading their XML, exports and checks PDFs, and runs a privacy audit over every extracted part of every file. | Blocking CI and privacy audit |
| orquesta-ia | Local multi-account orchestrator: routing by task type, accounting per quota window, limit detection and cross-auditing between models. Public antecedent of CAUCE. | SECURITY.md, secret scanner and regression suite |
| colmat-x-automation | Transactional state, OAuth 1.0a and a double safety against accidental publication. | Transaction and publication-safety tests |
| sincategorematico-bot | Three surfaces over one core: Telegram bot, local HTTP dashboard and desktop application. Ownership-claim protocol with SHA-256 and expiry; the token lives at 0600 outside the repository. |
Ownership protocol and local secret isolation |
| bloquitos · play | Browser, installable app, desktop executable and local database. The browser/game core has no runtime package dependency; desktop packaging uses Electron. | Release checks and game regression suite |
«Die Eule der Minerva beginnt erst mit der einbrechenden Dämmerung ihren Flug.»
Hegel, preface to the Elements of the Philosophy of Right
Jhon Steven Alvarez Ruiz · Neiva, Huila, Colombia · jhonstevenalvarezruiz.vercel.app

