super-research · diff
git:20260902.cb11d43 to git:20260903.7fb3c8c
20 added, 25 removed. Audit A to A.
---
name: super-research
description: Answer one bounded, keyless research question with cited, dated, gap-declared evidence from public platforms.
disable-model-invocation: true
---
Require: one bounded `question` naming its live sources, its `window`
where it has one, a frozen `as_of` at or after the run's own reads, and a
hard per-step cap.
Open the frame, its goal the answered question:
tickets.py frame-open <run> --goal-file <question-goal> --workflow super-research
-
- **Acquisition — one call per named source, launched together.**
+ **Acquisition** — *fan-out*: "One `do` per named item, launched together
+ under the frame; the shape line lists them as one wave."
- tickets.py do <run> --pack orch-research-pack --parent <frame>
- --goal-file <source-goal> --bound "<= 40 tool calls"
+ tickets.py do <run> --pack orch-research-pack --skill research-acquire
+ --parent <frame> --goal-file <source-goal> --bound "<= 40 tool calls"
- Each goal names one source, the window, the `as_of` and the cap, and
- directs the child to enter the `super-research` skill by name with its
- launch prompt forwarded verbatim, at the `SKILL.md` path
- `orchflows list --kind skill` resolves — carry that path, so no child
- searches; a body that opens a frame is this workflow's same-named
- adapter. That skill's Require, Preparation, Never and Return bind the
- child; its Return fills the `artifact: evidence:` line. Time-bounding is
- per operation — the skill's `WINDOW_REACH` table decides — so a windowed
- call whose operation cannot bound time at its origin returns
- `window_not_honored`; name it in the goal, so the child files a gap, not
- a silence. Keep each returned `artifact: evidence:<store-id>` line
+ Each goal names one source, the window, the `as_of` and the cap. Time-bounding
+ is per operation — `research-acquire`'s `WINDOW_REACH` table decides — so a
+ windowed call whose operation cannot bound time at its origin returns
+ `window_not_honored`; name that source in the goal, so the child files a gap
+ rather than a silence. Keep each returned `artifact: evidence:<store-id>` line
verbatim; the next call is handed it.
**Coverage loop, at most two rounds.**
tickets.py judge <run> --pack orch-research-pack --parent <frame>
--artifacts evidence:<id> [--artifacts ...] --goal-file <coverage-goal>
The coverage goal asks one thing: which sub-questions no record answers,
and which typed losses came back. Keep the `findings: <path>` line
- verbatim. Round two exists only for the gaps judge named — one `do` per
- gap, quoting the finding and the source that closes it, then one final
- `judge` over the enlarged set; a gap still open is declared.
+ verbatim. Round two exists only for the gaps that judge named — one `do`
+ per gap, quoting the finding and the source that closes it, then one final
+ `judge` over the enlarged set — and a gap still open is *declare-gaps*: "A
+ gap that remains is written as a gap, `[]` when there is none; silence is a
+ defect."
**Report, one call, the frame's last.**
tickets.py do <run> --pack orch-content-pack --parent <frame>
- --goal-file <report-goal> --bound "<= 40 tool calls"
+ --sheet html-dossier --goal-file <report-goal> --bound "<= 40 tool calls"
- Its goal hands every `artifact:` and `findings:` line verbatim and asks
- for one self-contained rendered dossier — one HTML file, no external
- fetch, legible in light and dark — answering first, then each source's
- evidence dated and cited from its `normalized_locator`, every typed loss,
- contradiction and open sub-question, under the skill's five report rules.
- Keep its `artifact: doc:` line verbatim.
+ Its goal hands every `artifact:` and `findings:` line verbatim and asks for
+ one dossier answering the question first, then each source's evidence dated
+ and cited from its `normalized_locator`, every typed loss, contradiction and
+ open sub-question, and each market's own price string with the markets that
+ already resolved dropped. Keep its `artifact: doc:` line verbatim.
Never: average contradicting sources, read a typed loss as an absence,
quote a community comment without its author and count, or close over the
acquisition children with neither a coverage judge nor an
`unjudged: <reason>` journal line.
Return: `tickets.py frame-close <run> <frame> --done <verifier>`, whose
done is the dossier verifier over that `doc:` identity — every
load-bearing claim cited and dated, every loss stated, every unanswered
sub-question declared.