Topic Atlas

Third of the atlas family. outreach-atlas maps people, discourse-atlas maps niche writing, this one maps what is known about a topic: the claims, the numbers, the mechanisms, the cases, the tools, where sources disagree, and what nobody has settled.

Goal: answer the brief's questions from pages someone actually read, and leave a field map that the next session can extend instead of redoing. Success means: every atlas item carries a verbatim quote, its URL, its date and its evidence kind; every lane ends with a coverage section naming what it searched and what it could not reach; every claim the synthesis rests on was spot-checked on its page by the primary; disagreement between sources appears as a listed disagreement, never a silent pick; what the research could not establish is listed under Nulls with what was tried. Stop when: SYNTHESIS.md answers the questions or names the holes, the counts line is current, and OptMem points at the directory.

Scale to the question

A "X vs Y" question takes three lanes whatever the size: A alone and B alone, neither told it is a comparison, then C seeded with the axes A and B surfaced. Comparative writing is a different population of sources than either solo pass.

Check what is already here before any spend: $MEMO_CMD recall '<topic regex>', then search $RESEARCH_ROOT and the local decision corpus. Configure paths in config.example. A hit turns "what are the options" into "what changed since we decided Y".

The one structural rule

Fleets write raw/; judgment writes the atlas and the synthesis. No script, fleet job or regex edits a curated file. The primary reads lane output critically, spot-checks on the page, and hand-picks what earns a line. Curation is the merge.

Layout

$RESEARCH_ROOT/YYYY-MM-DD-<slug>/
├── BRIEF.md       the questions, the generic context, lanes with their seed queries,
│                  runtime, ruling if one was asked for
├── raw/           briefs/<lane>.md, out/<lane>.md (+ .err), FLEET_PROGRESS, FLEET_DONE;
│                  per round: raw/r2/...; append-only, quote-don't-edit
├── atlas.md       (atlas size) counts line, then curated items by section
├── SYNTHESIS.md   verdict first, then the questions answered, cited to lanes and pages
└── CHECKS.md      what the primary verified on primary pages, date, outcome

Reusable patterns from prior runs: a synthesis with a "facts checked today" section, a second round aimed at gaps, and a "what we missed" lane closing each round.

Item format

- **One shared connection may receive any event** — one hub holds the connection and routes
  events. "<verbatim excerpt from a fetched vendor page>" (source URL, source date)
  · evidence: vendor doc · status: checked <date> · bears on Q2

• One claim per item. A number carries its unit, its date and what it measured; "2,000 users" without a date is not a number.
• evidence: is one of measurement, vendor doc, paper, case, practitioner report, opinion. A model's training is none of these; a finding without a fetched page stays out of the atlas.
• status: is checked <date> (the primary opened the page and the quote is there), unchecked (lane-reported only) or UNVERIFIED (the lane could not open a primary page). contested → D3 points at the disagreement entry.
• Counts line, first line of atlas.md: items 84 · checked 31 · UNVERIFIED 9 · disagreements 4 · lanes 8/8 reported. Only checked items are quoted outward.

Sections of atlas.md, in this order, each omitted when empty: Claims · Numbers · Mechanisms · Cases · Tools and projects (stars, licence, last commit, as read on the page and dated) · Disagreements (both sides, each cited) · Open questions · Pointers (people, venues, canon reading; a person here is a pointer to their writing, never a contact: no addresses, no hunt) · Nulls.

Flow

Brief. Write the questions as questions, numbered Q1…Qn; a lane answers questions, not topics. One prior fleet was re-run mid-flight because topic-shaped lanes were too constrained; the ruling was to use the reader's actual questions. State seed queries per lane and license the lane to derive more; a lane whose queries miss a phrasing misses the lane. The context paragraph is generic and public: nothing about the operator, no vault text, no CRM, no private project names, because the default runtime logs prompts. A brief that cannot be made public stays on the subscription tier (below). Ask the operator for a ruling only when lanes or spend would differ materially by the answer; a curiosity question is its own ruling.

Map. Round 1 is wide: one lane per angle, 6 to 12 lanes, each a report rather than a list. Lanes are aimed at pages that describe something that happened or was measured; a lane aimed at rosters and overviews returns overviews. Every brief carries the partial-output rule: write the report incrementally, a partial report with a coverage note beats an exhaustive search that times out with nothing.

Check. Before anything is curated or quoted: open the page behind every item the synthesis will rest on, confirm the quote and the date, record the outcome in CHECKS.md. In a prior fleet, all sampled quotes verified, while another synthesis caught a lane's wrong claim about a vendor rate limit by reading the primary page. Grade each lane in one line: reported, thin, failed, off-brief.

Aim. Round 2 goes where round 1 left holes: the Nulls, the disagreements, the claims that rest on one source, plus one "what we missed" lane that reads round 1's coverage sections and hunts the adjacent angles. Rounds live in raw/r2/ and so on; a rerun never overwrites an output file.

Curate (atlas size). Grep the atlas before adding; dedup is a curator's grep. An item without a quote stays in raw/. A section past ~40 items means the admission test is loose: tighten to what bears on a question.

Synthesize. SYNTHESIS.md opens with the verdict or the answer, then takes the questions in order, each answered in a few sentences with lane and page citations, then Disagreements, then Nulls with their detection limit (which queries, which sources, which were unreachable) and the strongest remaining route. A null closes nothing; retiring a question is the operator's ruling. Judgment work: runs in the primary, never in a fleet.

Point. OptMem P pointer to the directory and an E event for each fleet completion with the counts; the entity hub if the topic has one. Publish SYNTHESIS.md or the atlas as an artifact when the operator wants to browse or share. Commit the research repository.

Runtime

Default is the free route, because topic briefs are public inputs (the operator ruling 2026-09-17: continual mapping and web-research loops over public inputs run on free models, not the Codex subscription). omp-search runs Inkling inside free-silo: the model sees an empty jail, search is Codex-hosted through the bridge, and --oneshot leaves it read-only, so the report is its stdout:

D="$RESEARCH_ROOT/$SLUG"; POOL=3
for b in "$D"/raw/briefs/*.md; do
  n=$(basename "$b" .md); o="$D/raw/out/$n.md"
  [ -s "$o" ] && { echo "SKIP $n" >> "$D/raw/FLEET_PROGRESS"; continue; }
  while [ "$(jobs -rp | wc -l)" -ge "$POOL" ]; do sleep 4; done
  ( timeout 1000 omp-search --oneshot --max-time 15m "$(cat "$b")" > "$o" 2> "$D/raw/out/$n.err" \
      && grep -q '## Coverage' "$o" && echo "OK   $n" || echo "FAIL $n" ) >> "$D/raw/FLEET_PROGRESS" 2>&1 &
done; wait; touch "$D/raw/FLEET_DONE"

Launch it detached (nohup … &), poll FLEET_PROGRESS; the Bash tool's 120 s cap kills a foreground fleet. Pool 3 to 4: every :free run on the key shares 20 requests a minute and 1,000 a day, and a 429 storm means the pool is too wide. --model thinkingmachines/inkling:free swaps in the larger model when a lane needs depth. --search tavily is the fallback when Codex search quota or rate limits bite, with worse sources.

codex-fast --web-search on --oneshot --file "$b" < /dev/null (Luna, pool 8 to 12, as phone-control-hub/run-lunas.sh) when the brief must carry something private, when the question needs the subscription model's judgment, or when a free lane came back thin. Claude-side fan-out (Agent with quick-research or general-purpose on sonnet) is the small-fleet path for 2 to 4 lanes when no directory is wanted; it draws on the session's shared WebSearch budget, which 15 workers exhausted in minutes on 2026-09-06. No Fable subagents. Nuanced judgment (which of two contradicting sources is right, what the field is converging on) does not fleet on the bulk tier; it runs in the primary or on sol.

Runner mechanics, pool semaphores, the xargs -P trap, killing by fleet path and the Codex browser-connector rule are in outreach-atlas/references/fleet-recipes.md; read it before the first fleet of a session. The lane brief template is references/lane-brief.md.

Rules

Read-only everywhere except the lane's own output. No logins, no scraping behind walls, no browser automation, and no brief invites an "alternative way to load" a page. Quotes verbatim or not at all; an unknown field says unknown; nothing about the reader in any lane output. Versions, prices, model rosters and APIs turn over within months: the lane records each source's date, and the synthesis says which findings rest on a source old enough to have turned.