Texas AI Docket

texasaidocket.com

A public, fact-checked record of AI decisions in Texas: who is deciding, by when, whether the public still has a way in, and a fetched primary source behind every fact.

Three things run off this repo:

• The docket — every AI infrastructure decision in Texas, re-verified daily, published as a page, as open data under CC BY 4.0, and as a corpus written to be read by LLMs.
• The daily carousel — one verified Texas AI story a day, planned in forensic detail and rendered as bespoke code, delivered as a post-ready email draft.
• The Texas Grid Watch — a daily numeric record of the ERCOT grid's position, built around the load factor, which is the one number here that is genuinely about AI: a data center draws at four in the morning close to what it draws at five in the afternoon, so large constant load lifts the overnight floor faster than the afternoon ceiling.
• The Texas Water Watch — daily reservoir storage for 119 Texas reservoirs, rolled up by metro. The driest metro in Texas is in the Permian, which is where much of the new load is landing. The page publishes both numbers and draws no conclusion from them.
• The ask box — a question about the record, answered in the reader's own browser from an index that shipped with the page. Nothing is sent anywhere, and tests/ask_engine.mjs cuts the network and asks every catalogued question anyway.

Two sibling repos carry the rest: TexasAIDispatch (the narrated video engine) and TexasAIScanner (the Bottleneck Scanner backend).

How it fits together

ERCOT  ──► gridwatch collector  ──► ledger/gridwatch/readings.jsonl ──┐
TWDB   ──► waterwatch collector ──► ledger/gridwatch/water.jsonl    ──┤
                                    (both cron, never a routine)      ├──► site_build ──► docs/ ──► Pages
research ──► claims ──► docket routine ──► ledger/docket.json ────────┘        │
                              │                                                ├──► ask index + catalogue
                              │                                                └──► llms.txt, feeds, JSON-LD
                              └──► carousel routine ──► runs/ ──► Gmail draft

docs/ is generated. It is a pure deterministic function of the ledgers, and site_fresh_check.py proves it by rebuilding into a temp directory and requiring byte equality. Nothing hand-edits the published site.

The laws

Numbers are computed, never generated. Every numeral published traces to a quoted source or to code that computed it from data, and numeral_lint fails the build otherwise. It is stated publicly on the site, because it is the reason a reader should believe a number here.

Every path has exactly one owning actor. Several unattended routines share this history and each ends in a phase whose job is editing its own machine. Prose is not a boundary against that; ownership.yaml is.

docs/ is generated and never hand-edited, proven by rebuilding into a temp directory and requiring byte equality.

A gate that cannot go red proves nothing. Every checker here self-tests by reintroducing a known-bad input and requiring itself to reject it, before its real run is allowed to mean anything.

See HANDOFF.md for what has to happen outside this repo to turn it all on.

Running the guards

Both structural gates self-test before they run.

# No automation may write outside its lane.
python3 scripts/shared/ownership_check.py --self-test
python3 scripts/shared/ownership_check.py --actor daily --diff origin/main...HEAD

# Is the port done, and is what we moved actually wired up?
python3 scripts/shared/port_audit.py --self-test
python3 scripts/shared/port_audit.py
python3 scripts/shared/port_audit.py --summary      # just the progress table

Why there is an ownership map

Several unattended routines share this one git history, and each of them ends in a phase whose entire job is editing its own machine. ownership.yaml gives every path exactly one owning actor, and the check runs at commit time and again in CI. An automation that wants to change something outside its lane records a proposal in its run record and stops.

Enable the hooks

Cloning does not install hooks. Once per clone:

git config core.hooksPath .githooks

pre-commit enforces the ownership map. post-commit mirrors every commit to origin immediately, because this code runs in ephemeral containers where committing is not durable and only pushing is.

Repo layout

Path	What
prompts/	the routine prompts; each is the source of truth for its routine
knowledge/	shared/ Texas research and design doctrine, carousel/ deck craft
config/	brand.yaml shared voice and tokens, then per-surface config
ledger/	committed state: the public docket, plus per-actor memory
scripts/	namespaced by owning actor: site/, carousel/, gridwatch/, shared/
assets/	fonts, art libraries, Texas geodata, places gazetteer
docs/	the published site. GENERATED, never hand-edited
runs/	shipped artifacts, merged to main each run
out/	per-run scratch, gitignored

CLAUDE.md is the law: attribution, merge policy, the ownership model, the Grid Watch rules, and the house style rules that do not bend. A run's progress across contexts lives in out/<date>/run_state.json and in its committed run record. No routine writes a worklog. knowledge/shared/GATE_LESSONS.md is permanent, and is the record of faults that shipped with every check passing.

Status

Under construction. port_audit.py --summary reports how much of the machine has landed.