refactor(skills): deep-dive — 요청한 완료 지점까지 한 조사 담당이 이어 가고, 넘기기 전에 작성한 주장을 원문·계산과 대조한다

harness/skills/deep-dive/DESIGN_NOTES.md

@@ -2,7 +2,9 @@
 
 ## Shape
 
-A short core and three cases. The core holds the default path with its stop points and the few moves that capable models skip unless told. The cases show the rest by example. General evidence-weighing advice is left out on purpose: models already weigh vendor benchmarks, repeated claims and dated sources well without it.
+A core and three cases. The core follows what one research owner does: follow the unsettled question, read what each claim needs and keep its conditions, delegate with the question and evidence intact, and finish at the requested endpoint. The cases show scale and sequence by example. Tool execution stays in the Aside and Outpost skills.
+
+The earlier core left general evidence-weighing advice out, because the recorded blind A/B showed models handled it without help. This draft restores targeted guidance based on the compression failures in `cases/photo-map.md` and the user's report of relying on collectors' summaries for decisive reading. A fixed-material 2×2 and one real investigation (items 7 and 8) showed the aimed moves appearing; scope still widened between correctly recorded conditions and the headings, summaries and recommendations built on them.
 
 ## How it got here
 
@@ -15,12 +17,16 @@ A short core and three cases. The core holds the default path with its stop poin
    - evidence handoffs without a recommendation.
 
    General evidence handling was mostly fine without it. The text was then cut down to those moves plus cases.
+6. **Cross-run review and draft (2026-10-31).** The user's comparison of a motion-design investigation with local runs motivated this revision. The reconstructed record established inspection of relevant original passages and selected video frames; complete reading of every source was not established. A GPT-drafted rewrite proposed one research owner connecting open questions, original evidence, interpretations and the requested endpoint. A local Opus review then adapted the draft to Rubato. The revised skill's behavioral and speed effects remain to be compared.
+7. **Fixed-material 2×2 (2026-10-31).** Current and revised skill, each run by the lead directly and by one delegated owner, on the same request and material. Most preset checks came out the same. Both revised runs dropped the choices that still depend on the user's priorities; the closing sentence was restored. Scope widening and dropped qualifiers still appeared in returns and final answers of both versions.
+8. **Real investigation (2026-10-31).** Photo counts per place for a map video, run by one delegated owner on the revised skill. The owner reframed the question as time allocation, read originals and recorded conditions. Headings, the summary and recommendations still turned bounded findings into rules, and its own time allocation went unchecked. The pre-handoff check was rewritten to compare drafted claims with their passages, calculations and application assumptions (`cases/photo-map.md`, follow-up).
 
-## Two decisions from the user
+## Decisions from the user
 
-- **Stepwise.** The intent is a concrete, stepwise round. The steps stay visible as a default path, and any step can be the stop.
-- **Conclusion on request.** The round ends with decision material. A recommendation comes only when the user asks for a conclusion. Choices that depend on the user's taste or experience go back to them, by comparison or by seeing.
+- **Stepwise (prior decision, replaced).** The earlier design kept a visible default path and allowed any step to be the stop. The requested-endpoint direction below replaces that default.
+- **Conclusion on request (retained).** With no endpoint named, the round ends with decision material. A recommendation comes when the request includes one. Choices that depend on the user's priorities, taste or experience go back to them, by comparison or by seeing. Known user preferences are applied, not asked again. Questions that evidence can narrow stay with the investigation. A remaining choice goes back as a conditional comparison or a difference to see, not as an automatic request for more checks.
+- **Requested endpoint (2026-10-31).** The user's direction: 「단계는 중간 보고로 보이게 두고 요청한 완료 지점까지 진행하는 방향으로 가자」. Run to the endpoint requested: an evidence handoff, a recommendation, or a staged checkpoint. Stages stay visible as progress reports and are not new permission boundaries; stop at one only when the user asked to stop there. The canonical skill changes after the fixed-material comparison and one new real investigation.
 
 ## Revising
 
-Add or replace a case when a real run teaches something the current cases do not show. Add a line to the core only for a move that models still skip after seeing the cases.
+Add or replace a case when a real run teaches something the current cases do not show. Change the core when a comparison run shows a move models still skip, or a line that changes nothing.

harness/skills/deep-dive/SKILL.md

@@ -5,35 +5,40 @@ description: "Use when the user asks for a deep-dive (딥다이브, 깊은 조
 
 # Deep-dive
 
-A stepwise research round for a decision that matters. The user decides. The round gives them decision material they can trust, and a conclusion only when they ask for one.
+Investigate to the user's requested endpoint: an evidence handoff, a recommendation, or an explicitly staged checkpoint. Judge relevance and meaning throughout. Give a final recommendation when the request includes one; when the request names no endpoint, hand off the evidence and leave the choice to the user. A research assignment already authorizes the investigation. Keep its stages visible as progress reports; a stage is a stop only when the user asked to stop there.
 
-## Default path
+## Follow the unsettled question
 
-Open the question → collect in parallel strands → recheck what the decision leans on against the originals → curate → independent judgment, when wanted → the user's choice.
+Put the user's words, actual material and unresolved decision together. Preserve fixed facts, explicit limits, and rejected options with their reasons. Current explanations and candidates remain hypotheses; keeping the current approach is a valid outcome. New evidence may challenge an earlier rationale, but does not silently reopen a settled user constraint.
 
-Any step can be the stop: report there, and let the user say whether to go on. Loop back when a step exposes a gap.
+Ask what the decision requires before searching for solutions. Inspect supplied artifacts when they can change that question. Translate the underlying problem into terms used in other fields; treat known authors and projects as leads, not boundaries. Keep strong indirect material whose use is still unclear when it could change the framing.
 
-## What capable models skip unless told
+One research owner—either the lead or the agent entrusted with the whole investigation—keeps questions, original evidence and provisional interpretations together. Let findings change the next search, reading depth or division of work. Parallelize questions that can advance independently, within the accepted budget. Before a favored explanation closes a consequential alternative, look for evidence that could distinguish them; revise the question when the alternative explains more.
 
-- Every brief and collector prompt carries the user's own words, the facts they set, and the options they already rejected, with the reasons.
-- Current candidates and explanations are hypotheses. Keeping the current approach is a valid outcome.
-- A collector in its own session, such as a browser agent, reads only the text it receives. Paste the context and the collection rules into its prompt; a file path is not enough.
-- Seed each strand with the names you already know (authors, papers, projects, threads) as leads it may correct.
-- Curated material interprets the evidence but recommends nothing. End with:
-  - the strongest findings, with the conditions that bound them
-  - the premises that were checked
-  - the evidence that conflicts
-  - the gaps
-  - the choices that depend on the user's taste or experience
+## Read what the claim needs, and preserve its meaning
 
-  Recommend when the user asks for a conclusion.
+Open sources before reporting them as evidence; unopened leads remain leads. Prefer originals. Choose public or internal material according to where the answer can be observed. The owner examines the original material behind decisive or disputed claims, including context-preserving extracts returned by an authorized browser agent. A collector's summary alone is not that check.
 
-## Cases
+Read far enough to test the claim and its limits: for a study, its method, sample, comparison and results; for a product claim, the platform, version and how behavior was established. Skim for discovery, then deepen where interpretation depends on it. Record the location and actual extent of inspection. A few frames do not establish motion at normal speed. Distinguish what the source claims, what was directly observed, and what you infer for this task. Assess provenance and incentives; several retellings of one source are not independent evidence.
+
+Keep each decisive claim with the conditions, counter-results and scope that change its meaning through every summary and handoff. If brevity requires a cut, drop a less important claim rather than its decisive qualifier. Check the outgoing wording against the original: one person's reason does not describe an entire field, and missing evidence does not prove absence.
+
+## Delegate with the question and the evidence intact
+
+Give each research delegate its question, relevant user wording, facts, rejected options and reasons, and whether it owns collection or the whole investigation. Name the record file for the research owner; collectors return material for the owner to record. Ask delegates to read this skill when accessible; otherwise paste its reading and preservation rules. Supply the relevant text of any task material they cannot open. For a judgment consultation such as Outpost, follow its packet rules while preserving the question and decisive evidence conditions.
+
+Return findings with their decisive conditions, source locations, inspection status, conflicts and remaining questions. For restricted sources, return the relevant original passage and surrounding conditions within existing access permissions. The receiving owner follows up on consequential gaps, normally through the same research delegate while the question remains within its remit; it need not repeat reliable checks merely because work was delegated. Aside and Outpost keep their own skills for execution.
 
-Read the one closest to your run.
+## Finish where the requested work is complete
+
+Continue, within the accepted budget, while a credible, accessible next check could materially change the requested result. Do not stop merely because one explanation sounds coherent or a collection phase ended. When a necessary experiment, observation or user judgment falls outside the requested stage or cannot be obtained here, state what would settle it. Mark proposed checks as unperformed; distinguish an access, time or budget limit from sufficient evidence.
+
+Before handing off, compare the result with the user's limits and rejected choices. Check the decisive claims in drafted headings, summaries and recommendations against their supporting passages or calculations, including the assumptions needed to apply them here. Carry corrections through the current record and outgoing text. State what the evidence supports, what conflicts, what remains open, which material choices still depend on the user's priorities, taste or experience, and why the investigation stops here. Keep the research record in a file the next owner can open, and point to it from the working note. Rewrite changed judgments in place; log consequential search pivots, exclusions and superseded judgments with their reasons, and keep source records beside it. Mark missing history rather than reconstructing it as fact.
+
+## Cases
 
-- `cases/photo-map.md`: the origin run. A UI question was widened, six strands ran, a claim was corrected on recheck, and a split judgment went to the user as mock-ups to decide by seeing.
-- `cases/skill-catalog.md`: a harness-policy question run as an evidence handoff that stopped at the digest.
-- `cases/brief-contrast.md`: one request briefed with and without the moves above.
+Read the closest case when calibration would help; its scale and sequence are examples.
 
-Aside and Outpost keep their own skills for how to run them.
+- `cases/photo-map.md`: widening a visual question, indirect evidence, original-source correction, conditions lost during compression, and a follow-up where conditions recorded correctly still hardened into rules.
+- `cases/skill-catalog.md`: bounded findings, public-evidence gaps, and stopping at an evidence handoff.
+- `cases/brief-contrast.md`: carrying user context into collection without preselecting the answer.

harness/skills/deep-dive/cases/photo-map.md

@@ -53,3 +53,13 @@ Two slips got through, and they are the lessons:
 - An outside judgment received the digest, the user's words and the failed attempts. It reframed the question from "how to mark a cluster" to "what one mark stands for, and where a tap leads".
 - An independent session that knew the history argued with parts of it. It also caught the lead re-proposing a structure the user had rejected weeks before.
 - The two judgments disagreed on one point. That point went to the user as something to decide by seeing: real-scale mock-ups of both options, built from the user's own photo data. The research did not pre-empt that choice.
+
+## Follow-up: photo counts per place (2026-10)
+
+A later investigation for the same family of products asked how a map video should show one place with 3, 5, 10, 15 or more photos. One delegated owner ran it on the revised skill. It reframed the question from "pick one of three compositions" to "how a place's seconds are shared between big viewing, fast flow and a bundle", read the papers and product pages directly, and wrote their conditions into the record. Review still found:
+
+- Conditions written correctly in the evidence notes did not stop headings, the summary and the recommendations from stating stronger claims.
+- A tracking capacity measured under hard conditions ("as low as 1-2 objects", identical dots) became a general limit for photo cards. A flash count from an accessibility rule became a minimum display time per photo.
+- The owner had reframed the problem as time allocation, then did not put its own recommended compositions into the place's seconds. They did not fit.
+- Rereading the record did not catch these. Checking each drafted claim against its passage or calculation, with the assumptions needed to apply it, did.
+