Pro question

Authority: docs/project/OPERATING_CONSTITUTION.md section 5. The DM obtains one adequate independent
scientific review at its four decision points and adds Pro when it offers distinct expertise,
framing or unresolved-disagreement value. This skill authors that Pro question; it does not add
a second consultation requirement. No owner reminder or Root forwarding is needed within
authorized direction work. Pro advises; scientific direction correction and material dissent
follow constitution section 2, and other in-scope choices remain with DM. Consultation is not approval.
Select the reasoning needed: evidence synthesis, competing failure explanations,
simple-model/literature bridge, targeted revision, candidate generation or confirmation criticism.
Do not turn every result into a consultation or require a fixed number of new ideas.
Pro reasons from the repository at the supplied sha and explicitly supplied primary sources;
it does not inherit the author's chat memory. Use an existing adviser conversation when useful.

Before authoring

Name the impending scientific decision and the unresolved question that can change it. Check
whether a complete prior Pro answer already covers that decision and its evidence and premises
remain materially applicable. Reuse that reasoning for routine implementation, planned verification,
execution and collection; no new question or reuse receipt is needed. For confirmation the advice
must cover the actual claim, comparison and fixed plan. If the question, premise or evidence has
materially changed at a section 5 decision point, obtain focused scientific review and add a Pro
follow-up only when it offers distinct value; show what changed. Engineering review is separate
from scientific review. Preserve consequential independent-review disagreement in the existing question
when it bears on the decision; do not present only DM's preferred interpretation. Keep any frozen
review exception within its original object. Portfolio follows an owner request or explicit
project-management delegation under constitution section 2.

Steps

1. Write the section. Append to docs/research/candidates/<direction>/NOTES.md:

   ## Pro question <YYYY-MM-DD> <slug>
   Conversation: <URL from the NOTES.md header, or "new">
   Question: <one focused question; which judgment or choice the answer can change>
   Standing: <current explanation, the motivating observation/gap, supporting and contrary
     evidence, and links to the prior notebook interpretation, runs and claim note>
   Context: <concrete governance/method/evidence files and sections selected using
     references/pro-reading-context.md; source_sha unless an explicit frozen/evidence sha is given>
   Prospective cost: <planned fits and dominant non-fit work for the proposed next action>
   Constraints: seeds and matched baseline per constitution section 8; no training, no edits
     outside the empty "### Answer" subsection; write only there on branch <branch>.
     Read the question at the pinned source, but fetch the latest target file before editing
     and use its actual blob SHA. Preserve all other bytes; stop on overlapping edits.
     On successful write, report the actual commit. On write failure return the complete
     answer in chat, not just a SHA, status message or link.
   Return: <what the evidence strengthens/weakens/leaves unresolved; source-grounded
     explanation or prototype with assumptions and omitted MARL coupling; next useful
     observation and strongest alternative. A targeted revision predicts an intermediate
     change and native consequence, with fit and non-fit cost. Candidate count is task-specific;
     no new idea is required. For criticism also return MATERIAL_DISSENT yes/no.>
   ### Answer

2. Commit and push that file by pathspec on shared main, unless the owner requested another
   authoring branch. Record the full sha.

3. Compose the message with repository, branch, source_sha, target_path,
   question_heading and answer_heading (### Answer), the source-pinned URL and the exact
   answer-only write instruction. The source is immutable; the write target is the latest
   version of this file on the named branch. Use a unique question heading, not a reused slot.
   Include the reading/source-precedence instruction from the context reference in the actual
   message; do not send only a bare question URL or assume existing chat memory is current.

4. Send directly. Use the host's browser procedure in the authoring session. On WSL use
   hmasd-jev-pro-transport; use hmasd-chatgpt-pro-transport for the Agentify Desktop route.
   On the Jev route the reviewer is ChatGPT 6 Pro, with 5.6 Pro as the fallback when the 6 Pro
   quota is short (owner, 2026-09-27): both are on the same account and read the repository through
   the same GitHub connector, so one committed question serves either model and is not rewritten
   for the fallback; the transport records which model answered. When the question is an
   independent scientific review, include the review specification itself (reading order, the
   numbered questions, the return format with MATERIAL_DISSENT yes/no) in the question text.
   These names identify procedural skills, not subagent roles. After Send is accepted, a POSIX
   Codex session hands passive observation to tools/hmasd_wait.py using its verified task UUID.
   Other runtimes use a verified deterministic observer with native/manual continuation;
   they do not inherit Codex queue support. Do not delegate waiting to a model. Authorized
   direction consultation needs no Root relay or new per-question owner approval, and the
   App-only cross-task messaging restriction does not apply to Jev Pro.

5. On return, fetch and inspect before integrating. Apply the browser procedure's complete-answer
   checks to the specified target and immutable answer commit; a short receipt is not an
   answer. While Pro owns the subsection, the DM and its leaves do not edit it. Reconcile an
   uncertain write before taking back that subsection. Paste a complete recovered fallback
   only after verifying no answer already landed and the question is unchanged; commit and
   note "saved from chat". For a new Agentify-account conversation, put its URL in the notebook
   header. Jev-account URLs remain only in the local operation file; shared records use the key.
   Use the current branch/file version; never overwrite concurrent changes with the pinned copy.

6. Respond in writing. In the next notebook entry record what you adopt, modify or reject,
   which prior judgment changes (or remains unresolved), and why the selected next action follows.
   Preserve contrary evidence and hypotheses inherited from prior work. Advice or consensus is
   not an independent experiment. Apply section 2's scientific-review responsibility and
   material-disagreement resolution; do not create a second reflection report.

Method context

Pro does not inherit local skills or role instructions. In the existing question, inline the
brief methods that matter or explicitly ask it to read named sections of the scientific-tools
and, when relevant, research-engineering skills at a full published sha. Keep current methods
distinct from frozen experiment inputs; supply only relevant foundation/primary passages.
Check referenced paths exist at the published revision. Include the strongest alternative and
discriminator; fit counts alone need not describe a nested search or sweep's dominant cost.
This is question context, not a packet, recursive reading assignment or additional Pro round.
Use pro-reading-context.md to choose the concrete reading
set for synthesis, diagnosis, prototype reasoning, hypotheses or claim review, and an owner-triggered Portfolio
question. It also describes explicitly requested control reviews without adding a routine trigger.
On return, assess the answer's source use and unread material gaps before adopting recommendations.

Never

No packet, TASK or HANDOFF files, request ids, registry, binding records, finality labels or
mandatory result-review loop. Uncertain send state means reconciliation through the same browser operation on the same
message, never a second send. A method or wording change does not resend a question.