[bong-water-water-bong]

Why

docs/PORTING.md step 5 lists .github/workflows/census-{watch,sweep,autopr}.yml; only the sweep landed (census.yml). The watcher's source — Testing/hf_new_models.py and its compiled probe of rcpp_arch_from_string through Testing/census_coverage.py — does not exist in this repository, and the systemd unit that ran it pointed at ~/census-main, a worktree of the pre-split repository that no longer has a main. Every daily run since 2026-09-23 died 203/EXEC ("Unable to locate executable"), so the new-model alert was off for a week with nothing turning red.

What

The registry is generated data now, so the watcher is a lookup, not a compile: no g++ probe and no strip_arch, because registry/architectures.json is keyed by the HF architecture class the census counts.

• tools/census_watch.py — polls the newest text-generation models (createdAt descending), reads each one's class from the config inline in the listing, and compares it against registry/architectures.json (the same read tools/census.py's coverage() makes, so "uncovered" means "not in the number the census reports"). Exit 1 is the alert.
• scripts/census-watch.sh — entry point: log file, exit contract, CENSUS_WATCH_REFRESH=1 to fast-forward a detached checkout (refuses a checkout on a branch or with local changes).
• scripts/1bit-census-watch.{service,timer} — 04:30 UTC daily, from a ~/census-main worktree.
• docs/registry.md — "The watch": the alert contract and how to triage each kind.

State lives outside the checkout ($CENSUS_WATCH_STATE, default ~/.1bit/census-watch-state.json), so a run can never dirty the tree it reads. Gated repos (no config) are reported as unverifiable and never fail the run; derivatives (a quantized GGUF, a LoRA adapter) have no config by design and are skipped, because the release they derive from carries the config and is what gets checked.

No pull_request trigger on the workflow, deliberately. The live watch is red whenever any newest model's class is unmapped — that is the alert — and the census reports a mapped share below 100% by design, so a PR gate would be red on every pull request and mean nothing. The watcher's own contract is checked offline by ctest census_watch (run by ci.yml); the live alert runs on a schedule.

Two bugs the first live runs found, both pinned by the test

• Pagination. HF ignores p= on a sorted listing — every page returned the same first 100 models, so --limit 120 fetched 101 and still reported 120. The walk follows the Link header's cursor instead, the way tools/census.py sweeps the full listing. This one mattered: NcpSmolForCausalLM (below) sits past model 101 and was unreachable.
• A false green. A run that checked 0 models printed "OK — no uncovered class". It now says nothing new since the last run (N models already classified).

Evidence

First clean run against this registry, the 120 newest text-generation models:

census-watch: 120 newest text-generation models checked, 102 with an architecture,
94 mapped, 6 uncovered class(es), 3 unverifiable (gated/no-config),
15 derivative (no config by design)

So six classes arrived that no backend maps — Qwen3_5MVLAAbsorbedForCausalLM (2 models, e.g. TelperionAI/Qwen3.8-27B-GLA-g2), ArgonneModel (2), GPT, Glm5NextForConditionalGeneration, NcpSmolForCausalLM and NimbusMindForCausalLM. Triage is a follow-up, not part of this PR: none is in registry/significant.json, so each needs either a real mapping (only where a pinned backend's code really accepts it) or a recorded reason in significant.json + docs/arch-gaps.md.

Testing

ctest census_watch runs tests/census_watch_test.py: 27 offline checks (no network) driving classify(), report(), remember() and sweep_newest() against synthetic batches, synthetic listing pages and the committed registry. The committed registry always passes, so a regression that made the alert stop firing would leave the daily check permanently green — the same reasoning behind registry_gaps_test.py.

The systemd path was run for real on the development box: the unit fails with exit 1 on the six classes above, the log lands in ~/.1bit/logs/census-watch-*.log, and the 04:30 UTC timer is enabled.

Note for review

The Actions workflow is not in this PR anymore: it needs issues: write, so it goes through RFC #246 first (CONTRIBUTING). The systemd timer is the alert path until then.

[context7]

Docs7 for 1bit-monster/engine

Result	Status	Action
Deployment	➖ Not used	—
Content review	➖ Did not run. This site has no agent runs available this month. Wait for the monthly reset or check your Docs7 plan.	—

Commit 4e5f791