Semantica

What it is

Semantica is the coordination layer for the agents that build the projects on this site, and for Luke himself. Its product is two desks that face each other: Luke’s desk, where the things only he can decide wait for him, and the agents’ desk, where refereed work packets wait to be run. A card is a view of an object plus the actions that unblock the other side. An empty agents’ desk after a run is success; a cron job is not a feature.

Work arrives as a packet: the paths it may touch, a referee command that decides whether it is done, and a goal. The kernel that runs packets is small on purpose. Its size is a convention the agents enforce with their own test, about 1,500 lines of Python and 800 lines of desk markup, so that one person can still read all of it; it is not a rule Luke wrote. The project manager that stamps, reserves and schedules work lives inside the Umwelt engine as its Semantica consumer, so every write to the desk goes through Umwelt and the desk is itself an Umwelt consumer.

It replaced a 95,000-line predecessor, Semantica 2, that worked in places and could not be carried. That tree is frozen, not deleted, and the lesson is written into every charter since: hundreds of kernel lines, not tens of thousands. Semantica 1 was a complex-valued semantic engine, where meaning had a magnitude for content and a phase for context; that idea stays as lineage and a planned paper, not as shipped code.

Luke’s personal front desk is a role inside Semantica: email triage through AttentionShield, tasks, reminders and bills behind one keyed door, with a dial that earns autonomy from watch to run. Its demo world ran a synthetic week in Umwelt’s gate and passed its five proofs.

This site is one of its stations. The things only Luke can do are filed here as asks in Semantica’s own dialect and served from this site’s desk feed so his desk can pull them; the lifecycle rules Semantica enforces are published in this site’s repository and run as a referee in its CI; and this site’s repository carries the reporter that reads Semantica’s desk on its host and posts the host report and Semantica’s face to this site. Until that reporter is installed on the host, the machine signal on the home page reads NO CONTACT, and the page says so rather than guessing. The site you are reading is, in that sense, Semantica’s public face.

How work moves

• Observation, then recommendation, then commitment: CI and ingest observe, issues and pull requests recommend, a merge, a label or a lease commits.
• Statuses: proposed, ready, leased, working, then blocked or review, then completed or abandoned. Illegal transitions are rejected, and adding evidence never changes a status.
• A claim is worth 0.25, a wall 0.40, a referee 0.92. Adapters never promote: nothing moves up a tier without the test that tier names.
• A missing measurement reads dark, never zero. A skipped CI job is not a pass.

What runs

• A desk server on a local port with a handful of read routes (the current desk, the manager’s history, the knowledge index, a read-only observatory of Umwelt’s view of the work) and two write routes, both answered by Umwelt’s Semantica consumer.
• A command line: list the enrolled work, run one packet, wake and run until the queue is empty, scout the host for what is alive, or serve the desk.
• The stamped manager inside Umwelt: every request is stamped and content-addressed, one writer is reserved at a time, repeated request ids return the saved receipt, and an interrupted request answers 409 rather than repeating its effect.

What exists today

A private repository, 441 files, main branch last changed 16 October 2026 and pushed 21 October 2026; a kernel of 1,498 lines in 17 Python files, under its own 1,500-line convention; 129 tests in 77 files plus 69 one-test example packets; no continuous integration, the suite runs by hand. The desk runs on a local port and over Luke’s private network; not a public service.

The stamped manager’s pilot (notes dated 12 October 2026): one task selected and run on the real engine daemon, its outcome verified and the policy advanced from version 1 to 2; reconstructing the runtime ran zero duplicate tasks; stopping and restarting the worker recovered the identical field hash with 12 recorded sensor rows. Four synthetic comparisons of the classical and the coherent selection policy returned the same rankings, which the notes call a tie in choice, not evidence of an advantage.

The front desk: a standard-library Python daemon with eight tabs, all live for Luke, and five of five proofs on a synthetic week in Umwelt’s gate (28 July 2026). It owns no task database and no Google client of its own; it composes upstreams.

The lifecycle referee runs in this site’s CI, this site’s desk station serves its open asks in Semantica’s dialect, and the host reporter that posts Semantica’s face lives in this site’s repository with its own tests.

Not done

Not a public service, and nobody but Luke sits at the desk.

No continuous integration: the test suite runs by hand. The Umwelt dependency is real and unpinned.

Nothing in Semantica posts to this site yet. The reporter that will do so is written and tested here, and waits for Luke to install it on the host.

One packet has not yet been run end to end with an independent check, with a forecast recorded before the attempt and resolved after.

The Gmail round trip through AttentionShield is unverified, and the front desk’s cruft-capture gate, catching 70 to 80 percent of logistical cruft, is unmeasured.

Next

Install the reporter on the host so the home page’s machine signal turns live and the Semantica card carries a face measured from the desk.

Let Semantica’s desk read this site’s open asks, so Luke’s two desks face each other.

Run one packet end to end with an independent check: record a forecast before the attempt and resolve it against the result afterward.