Our venture runs start from portfolio/flagships.yaml, a short list of deep projects with a buildable: line saying what a cloud session can do with no outside input. Several of those items need a public data host: raw camera files from raw.pixls.us, chant scans from archive.org, Mexico's SAT error matrix from omawww.sat.gob.mx.
The rows recorded whether each host worked, in prose, and the prose kept going wrong. platform/ci/friction_rank.py --live --ungated --days 2 ranked this file first: 13 run reports in four clusters between 2026-11-13 and 2026-11-15. Within one day, raw.pixls.us was written up as "answers 403", then rewritten as "serves files with curl -L" by a run it served, then reported refused at CONNECT by the next. archive.org flipped the same way on 2026-11-13. Each stale line cost a run its first stretch: one skipped a buildable item that was reachable, another planned on one that was not.
The egress proxy's answer depends on the session, so no sentence written down stays true. The fix is to stop writing the sentence and ask the host instead.
- A flagship row can list
hosts: [<url>, ...].portfolio/router/validate.pychecks it is a list of http(s) URLs. allocate.py --venture <slug>probes those hosts, plus any host name in the row's prose, with the same HEAD test the research programmes'HOSTSfiles already use (research/_shared/pick_track.py): any HTTP answer counts as reachable, and a failed connect or proxy CONNECT counts as refused.--no-probeskips it.- The first live reading, from this session: github.com answers; raw.pixls.us refused. That is the opposite of what the row said an hour earlier.
The photo-editing row no longer claims raw.pixls.us "serves files", and the file's header asks rows not to record reachability in prose. Four new tests cover the parser, the extractor (it must not mistake README.md or calc.mjs for a host) and the report line.
This follows a rule of thumb we keep coming back to: when runs keep doing the same check by hand, a script should do it. A host probe takes a few seconds, and a person or an agent re-reading a stale note takes much longer.