How I work · the factory · the trace · GUTS · the private work

Every change is a ticket, and the ledger keeps the receipts.

I run my projects through a software factory built on Claude Code. A ticket goes intake, spec, a gate I hold, build, review, verify, a second gate I hold, ship. The models do the keystrokes. The ledger keeps the receipts.

This site is one of those projects. The ticket that rebuilt it is T-1 in the portfolio-hub shop. Its trace is below, exported from the real ledger file as the last step of the build, and it says when it was exported. It is a snapshot; the ledger keeps moving.

The factory ↓The trace ↓GUTS ↓The private work ↓

The factory one ticket, nine stops, two of them mine

The process is versioned, like the code. Each stage writes evidence before the next one starts; a stage that cannot show its work does not pass. The human gates are real gates: nothing merges, deploys, or ships until I have said yes, and the ledger records who said it.

Everything in the list on the right happened to this site. The intake record, the spec with its fifteen testable acceptance criteria, the rendered direction I opened and approved, the worktree the build ran in, the screenshots the builder took of its own pages, the review table, the verify run: all of it is in the repository’s .autodev/ folder, and the trace below is its timeline.

The process does not depend on any one model. Which model runs which stage is a setting in the shop’s config, versioned with everything else.

IntakeThe request becomes a scoped ticket: what this is, whether it already exists, which lane and which gates apply.
RefineThe spec: scope, decisions locked so they are not re-opened, the house style, and acceptance criteria a script or a reviewer can check. For this site it shipped with a rendered direction I could open and drive.
My gate: spec readyI read it and say yes, no, or change. The ledger records “the owner said yes” before anything is built.
Locate and planWhere the change lives in the repo, and the slices it will be committed in.
BuildA worktree and a feature branch. Real artifacts captured, pages written in the voice of the sources, screenshots of its own output looked at and fixed, evidence written per criterion.
ReviewA fresh reviewer, the four questions above, one row per page. Any failing row goes back to build with the reason.
VerifyThe acceptance criteria re-run as commands: the greps, the mobile width, every link, the word counts, the PDF text.
My gate: acceptI open the thing and decide.
ShipMerge and deploy run unattended after the accept, and the trace shows them.

The trace the ticket that built this page, from the real file

portfolio-hub · .autodev/events.jsonlas of 2026-09-05 07:23 UTC · 34 lines

30  process blueprint solo-builder-review imported by the owner
30  AutoDev is set up for shop portfolio-hub (solo-builder-review)
30  a session starts
35  T-1 is new: rebuild-portfolio-hub-in-the-owner-s-voice-ai-native (feature)
35  T-1 passes intake (fable-5-1)
35  T-1 moves on: intake → refine
52  T-1 passes refine (fable-5-1)
52  T-1 waiting on you: gate spec_ready
59  the owner said yes to T-1
59  T-1 gate spec_ready cleared by pm (for the owner)
59  T-1 moves on: refine → queue
59  T-1 passes queue (fable-5-1)
59  T-1 moves on: queue → locate
59  T-1 passes locate (fable-5-1)
59  T-1 moves on: locate → plan
59  T-1 passes plan (fable-5-1)
59  T-1 moves on: plan → build
59  T-1 build: worker started (fable-5-1)
05  T-1 passes build (fable-5-1)
05  T-1 moves on: build → auto-review
05  T-1 build: worker reports back (fable-5-1) — /home/corgea/Desktop/Coding Projects/po…
05  T-1 auto-review: worker started (fable-5-1)
31  T-1 goes back auto-review → build: auto-review FINDING (reviews/T-1.md): how-i-…
31  T-1 build: worker started (fable-5-1)
48  T-1 passes build (fable-5-1)
48  T-1 moves on: build → auto-review
48  T-1 build: worker reports back (fable-5-1) — /home/corgea/Desktop/Coding Projects/po…
48  T-1 auto-review: worker started (fable-5-1)
22  T-1 passes auto-review (fable-5-1)
22  T-1 moves on: auto-review → gate
22  T-1 gate merge cleared by policy:unattended
23  T-1 gate merge cleared by policy:unattended
23  T-1 passes gate (fable-5-1)
23  T-1 moves on: gate → verify

Times are the ledger’s own clock. Each line is the tracker’s plain-English rendering of one event in .autodev/events.jsonl; the export script reads the file through the tracker and writes both this box and js/trace.json.

GUTS the meta-workspace underneath the factory

I created GUTS to organize my multi-agent projects.

Using RAG, an MCP server, and agent-to-agent communication, I used the platform as the shared foundation for roughly a dozen tools and applications, ranging from launching applications to restoring an old inbox. I integrated tmux, OpenClaw, and Tailscale so applications and agent workflows can be used interchangeably across desktop and mobile devices, allowing long-running agentic work to be managed without remaining at a workstation.

GUTS is a private factory that runs on one always-on machine. Its repository is the factory’s spine: it owns the canon, the seam (contracts/), and the assembly record (manifest.yaml). It vendors nothing. Each organ is its own git repo, pinned there; the meta-repo only references the machines, it never contains them. Pinning is a manifest, not submodules.

The seam is the single point where organs and satellites touch. The dependency direction is one-way and enforced: contracts/ depends on nothing; the organs depend on it; nobody reaches past the seam. Naming that rule is what stops coordination from ever swallowing the machines again.

The public GUTS architecture site ↗

Layer	Organ	Role
L0 · Runtime	GUTS Bridge	Browser control surface over tmux + Tailscale Serve
L1 · Memory	GIMS	Append-only institutional ledger; every action a sentence
L2 · Coordination	GONS	The foreman and command room. References machines; does not contain or execute them
L3 · Interpretation	GEDS	Relevance engine, five modes over events, tensions, and anomalies
L4 · Intent	GOMS	Intent compiler: goal → project → sprint → task → action. Owns decompose → execute → PR review

Constitution · five rules

Everything is a sentence in GIMS: subject · verb · object · at · by · under. Corrections are new verbs.
Nobody talks directly. Every message routes Agent → GONS → GIMS record → recipient.
Memory before autonomy: the ledger precedes any coordinator or executor.
The human holds the gate.
One hallway at a time.

Photo: Brett Sayles, Pexels · decoration: the seam is a patch panel, one-way and enforced

The private work, as it stands described, not linked; these repositories are private

GOMS Generalized Orchestration Management System · L4, intent

An autonomous machine that takes a goal for a software repository, plans the work, executes it through skilled agents, reviews the resulting PRs, and merges or re-queues, without waiting for human approval in normal operation. In the spine it is the intent compiler: goal → project → sprint → task → action. It owns the decompose → execute → PR-review lifecycle.

Where it stands, as of July 2026: the last item in the build order, “generalized GOMS”, and the far end of the autonomy work: closing the live GONS → GOMS loop so the factory runs unattended, veto preserved. Phases 0 and 1 landed; phase 2, safety and surfacing, is next.

GONS Generalized Operations Nervous System · L2, coordination

The foreman and the command room. It references machines; it does not contain or execute them. All messages route through it, and as of July 2026 its coordination ledger is GIMS-native grammar: a work order, a coordinate verb, a ledger record, one runlog per work order. The One Body unification landed one ledger, one loopback front door, one green per organ, and the GONS scribe and pulse.

GEDS Generalized Evaluation & Decision System · L3, interpretation

GEDS is an experimental relevance and anomaly engine; its core technique is structural salience. It watches an ordered stream of events and tries to identify when the underlying structure of the situation changes: a hidden authority becomes visible, a ledger stops being trustworthy, a routine resource turns out to be controlled by someone else, or a contradiction surfaces inside an otherwise ordinary scene. It is not a trained classifier: the current implementation uses frozen sentence embeddings, a dynamic threshold, scaffolded tension descriptions, local contradiction checks, and labeled corpora for evaluation.

Where it stands, as of June 2026: evaluated on labeled narrative corpora, and the reference evaluation is strong on both twist recall and false-positive precision: 40 of 40 planted twists caught and 20 of 20 false-positive probes clean across the twenty reference corpora, with the out-of-sample sci-fi and magical-realism sets clean as well. It is meant to sit beside an LLM-driven narrative system, campaign simulator, or long-running story engine as background pressure monitoring: noticing that a story has crossed into new structural territory before an LLM smooths it away. In the spine, that is the interpretation layer.

GLP Strong a collaboration · pre-MVP

Helping people on GLP-1 medications stay strong while losing weight, by making protein-focused nutrition, strength-supportive habits, and progress tracking simple. It is a consumer wellness product, not a clinical one: it does not diagnose, does not recommend or change medication doses, and questions outside that boundary are refused and escalated to the person’s care team, never answered by a model. The demo being built will show onboarding with versioned health-data consent, food search over an 8,204-food catalog, a coach chat behind a gated AI pipeline and a hard refusal boundary, and curated program and recipe libraries the AI selects from and explains but never invents. The canonical product repository belongs to a partner; my checkout is the working mirror, and work ships upstream as pull requests.

Where the spine stands, as of July 2026

Assembled and green: per-organ CI 5 of 5, the manifest verified 5 of 5. Build order: remote runtime (done), the Bridge (done), the GIMS event spine, GONS core, GONS console, GEDS mode 1, GEDS modes 2–5, generalized GOMS. Everything is a sentence, nobody talks directly, memory before autonomy, the human holds the gate, one hallway at a time.