/architect · Technical
Engineering record.
What is built, what is next, what is backlog. How the model is kept honest. The decisions, with the one that was withdrawn the same day.
Built, next, backlog
- builtDependencies: the reach of a change counted over the estate's closure, modules, actions at module level, other models from import names, the graph that grows a click at a time.
- builtThe 42-test catalogue and the action plan, deterministic, with the owner's state on every card.
- builtSize: list, time and version sizes solved from the cell counts, each axis priced, over-dimensioned line items with the owner's decision on each.
- builtChat with tools over the parsed estate; every object name a chip; the verdict line under every answer.
- builtNotes typed and scoped, siblings that contradict, supersede that keeps the record; flagged when a later export moves the object.
- builtConversations kept with the estate, held on the snapshot they started on; continue or branch when the estate moves.
- builtSnapshots with history, backup and restore as one zip, deletion that clears retrieval.
- builtTracing of every conversation as one tree, when keys are set.
- nextProject and design documents, meeting notes and change logs ingested next to the structure, so the answer can cite the document that decided it. Built on Grounded, being joined.
- backlogSaved views, UX pages and what feeds each card; which action uses which view.
- backlogModel history: what changed, by whom, over years.
- backlogA model book regenerated from the export, one page per module.
- backlogExports pulled through the Anaplan API on a schedule.
85 tests in the Architect's own suite; the engine's tests run again at every pin. The repo is private; the engine and the grammar are public.
How the AI is constrained
- The toolbox is bound to one estate and one snapshot. No estate identifier ever arrives from model output.
- Tool results carry citation ids. The model may only cite ids it was handed; anything else is dropped or flagged. The question is no longer "is this name real" but "is this citation real".
- Notes and uploaded text enter the prompt inside a delimited evidence block framed as data, never as instructions.
- Retrieval per turn is the pinned estate facts, the matching notes for the scope, the conversation tail and whatever the tools fetch. No embeddings, no vector store.
- Anything not established by an export, a note or a tool result is answered as unknown. Absence of a recorded reason is a finding, not a gap to fill.
- Relationships, counts and measurements are computed in code. The model explains them, connects evidence and proposes investigations.
Limits the app prints on its own pages
- Cross-model reach is inferred from import names at model level. The exports carry no line-item mapping across models and none will be fabricated.
- Where an action is linked only at module level, it is shown at module level.
- Change reach describes dependency reach. It does not prove a value will change or that an edit is safe. Coverage gaps sit next to every count.
- Marking an action done records a decision. It does not verify anything in Anaplan.
- Local storage is not local AI processing; Settings says which provider sees what.
- Missing information disables or qualifies the affected analysis. It never becomes a zero or a confident conclusion.
Decisions
- Identity
- Real accounts, many to one workspace. Nobody creates an account from the web; the owner is created on the console and hands out invite codes. The earlier "first user becomes owner" was built and withdrawn the same day.
- Persistence
- SQLite from the standard library plus a content-addressed blob store. Nothing to install beside Python.
- Frontend
- React and Vite, the built files committed, so a pip install alone still runs it.
- Graph
- Cytoscape, loaded on demand, grown one click at a time instead of drawn whole.
- The engine
- The Model Health Check and the grammar installed at pinned commits, never as editable links.
What is mine and what is not
The grammar, the test catalogue, the size solver, the evidence layer and the app are mine. The fictional estate on every page here is mine too, four models with faults planted on purpose and listed in the repo. The one real estate the Architect was built against is a client's and stays private: nothing from it appears on this site.
Public record
What has been built.
- Page/architect site family, nine use cases answered from the exportsOne real question per page on a fictional estate, re-set as HTML with the evidence drawer; the free check, the grammar, the CLI and the engineering record under Technical
- Page/architect project page, what it does today and what it cannot do yetThe landing page for pilot estates, with the four missing layers named
- PostAnaplan has the number. Finance still has the work.
Run Anaplan in-house?
Start with the Model Health Check on your own exports.
The 42 tests and the action plan run free, hosted or on your own machine, and nothing is stored. Then give me 30 minutes and bring the three questions you would ask about your estate.