station foundation
What every ch21 channel is built on
Six traits get asked about whenever someone sizes up an agent platform. Here is what each one means in plain words, and the machinery on this station that does it — the part you can watch happen on a take instead of taking our word for it.
Can it take a vague objective and break it into steps on its own?
- Every brief goes to the planning desk first, which returns a typed plan: sub-goals, the tool each one may use, the test that says it held, and what it waits on.
- Step count and tool choice are decided per brief. A simple ask gets three beats; a brief hiding several questions gets six.
- If the planner comes back thin, a deterministic house plan takes over so a channel never rolls blind.
see it: planning desk lane on every take's tape
Does it notice when an approach failed and change course?
- Every model call and page read runs on one reliability lane: bounded retries with backoff, honouring rate-limit waits, terminal errors surfaced in plain words.
- A dry named site widens to an open sweep automatically instead of answering from memory.
- When the screening room scores a take under the channel's bar, the failed dimensions become a repair brief and the take is re-planned and rolled again — capped at two re-takes, and the best take is the one that files.
see it: re-plans row under the shooting schedule
Can it reach real systems, not just talk about them?
- Live crawling: named-site discovery, page reads, open-web sweeps with a topic gate that drops off-subject pages before they cost a read.
- The station's own tool bay speaks Model Context Protocol, with auth, per-tool rate limits, result caching, a sandbox mode and an audit row for every call.
- Outbound webhooks carry idempotency keys, and scheduled work runs unattended overnight.
see it: tool bay, and the second-unit lane on the tape
Does it remember across sessions, or start over every time?
- Each channel has its own memory shelf: entries it wrote itself, tagged as what happened, a verified fact, or how to shoot it, each with provenance and a confidence.
- Entries are embedded, so a brief recalls the handful that actually bear on it by meaning — not a fixed paragraph pasted into every prompt.
- The station's filed archive is separately searchable by meaning and keyword, so verified material can be pulled back years later.
see it: memory shelf on any channel you own
Does it check its own work before handing it over?
- A deterministic scorer grades every take on grounding, on-subject, specifics, discipline and clock — from checkable facts, not a model rating itself.
- The fact desk verifies every link against pages actually read and rewrites unsourced citations out, rather than hedging.
- Blueprints go through a story-editor pass, adversarial canaries probe for drift and leaks, and the screening suite rolls overnight on its own.
see it: scorecard, canary strip and screening room
What stops it doing something irreversible or unsafe?
- Every request is validated and every untrusted brief or scraped page is tagged as data, never instructions.
- Deterministic post-filters run last, after the model: unsourced lines cut, off-list hosts dropped, findings capped, contacts redacted, implausible measurements refused.
- Any irreversible step a plan asks for — a send, a publish — is refused by the runner and opened as a check-back a person has to confirm, on the record.
- Roles are enforced on the server, and row-level rules mean a channel's material is reachable only by its owner and station admins.
see it: safety knobs and the check-back desk
The fastest way to check any of this is to build a channel and watch the tape: the plan, the pages it read, the score, the re-plan if it needed one.
Or take the shortcut: the trait evidence board counts all six traits off real station rows, rolls a live probe of every tool lane, and exports the whole thing as a one-page report.