Ticker Universe Builder

Build one market at a time. Every market is an independent universe; there is no cross-market merge and a name listed in two markets is two members.

Registered markets, each with reviewed rules, its own theme table and a report in its own language: us, cn (zh-Hans), jp (ja), in, hk (zh-Hant), kr (ko), uk, tw (zh-Hant), de (de), fr (fr), ca, au, br (pt-BR), crypto. Anything else builds too — see step 6.

Read examples/README.md first. The ClawHub package ships all worked inputs; run python examples/build_examples.py once to recreate the dated reports before inspecting them. Examples ship for US Light/Medium/Heavy/Max and CN/Crypto/HK/JP/KR/UK Medium; US shares one snapshot. For other markets and depths open the closest worked input and the matching overlay. Max still requires a qualified Heavy seed. Read the example spec and report summary first; inspect relevant candidate rows programmatically rather than loading multi-megabyte snapshots into model context. Archived replay uses assets/legacy-policy.json explicitly and does not certify the current coverage contract.

Route the request

1. Read references/methodology.md and references/tier-profiles.md, then references/coverage-plan.md.
2. Read the market overlay at references/markets/<market>.md — exactly one. For an equity market read references/markets/equity-common.md first: the universe boundary, the fund-versus-basket redundancy test, the cash-management exclusion and the rule about regressing against the theme rather than the index are shared by all of them, and the per-market file covers only what is true there and nowhere else. crypto has no shared part; read references/markets/crypto.md alone.
3. For an existing universe, also read references/maintenance.md.
4. Read references/data-contracts.md before writing any JSON.
5. Follow references/source-policy.md for evidence and provider use.
6. Read references/measurement.md before filling in any metric.
7. For output paths, artifact formats and delivery, read references/output-artifacts.md.
8. To check a universe after the fact, read references/evaluation.md.

If the market or the depth is missing, ask only for the missing choice. Default the depth to medium when the user asks for a generally useful universe without naming one.

Start from a watchlist the user already has

If the user brings an existing TradingView export, do not retype it:

python scripts/universe.py import --watchlist theirs.txt --market us \
  --output snapshot.draft.json

The draft carries their tickers and their sections as a starting taxonomy, and nothing else — a txt file does not say what is still listed, what anything is for, or how liquid it is. Every candidate comes back ineligible and complete is false, so the draft cannot build until the research below has been done against it. Tickers that do not belong to the named market are reported rather than dropped.

Build a new universe

1. Write build-spec.json from the user's request. Use the reviewed plan budgets unless the user asks for a smaller ceiling. Write coverage_plan first: stable economic sectors/branches, reviewed leader roster, necessary peers, scope, four entity ceilings and reference instruments. Counts are ceilings, never a reason to pad. When migrating a Core, run audit-core and resolve every original code before publishing Heavy; a retained code is not automatically a leader.

2. Start the taxonomy from the published one rather than inventing themes per run — two universes of one market built on ad-hoc taxonomies cannot be compared:

   python scripts/universe.py taxonomy --market us --profile light
   

   Edit display themes for readability. Their number and weight do not determine economic budgets. Set stable parent-sector caps/weights in coverage_plan before looking at optional candidates. Merge sparse themes with economically adjacent themes through coverage_plan.display_groups; preserve underlying duties. Light/Medium may use broader groups; Heavy/Max share the same map. Max follows Heavy group entity proportions, with only integer rounding surplus, and stops at the minimum qualified expansion. Carry forward the user's existing economic-driver map when available, but recheck ambiguous assignments against current business disclosures. Preserve the method, not inherited mistakes. Separate the primary observation purpose from secondary businesses; fibre/cable is not an optical module, EDA is not generic enterprise software, and oilfield service is not shipbuilding. Apply the same legacy principles to every market: primary business, earnings/value capture, persistent catalyst, then supply-chain position. Provider sectors and product tags discover candidates; they do not replace that map. Declare each theme's purpose and representative_roles (contracts). Preserve market/sector gauges separately from companies. Do not generate weights from how many names the provider happened to return.

3. Check the table before researching a single candidate. This is the step that is cheapest to redo now and most expensive to redo later:

   python scripts/universe.py taxonomy --check taxonomy.json --market cn --profile light
   

   This checks display structure only, not economic feasibility or per-theme seat floors. The formal build checks economic branches, necessary representatives, sector caps and export limits.

4. Optionally fetch a dated research bench with fetch before researching admissions:

   python scripts/universe.py fetch --market cn --prices-until 2026-09-28 --limit 1250 --output temp/cn
   

   See references/providers.md. Fetch is optional and never chooses membership. Inspect the manifest, rejected histories and raw receipts; provider classification is a starting point, not issuer due diligence. Research the eligible universe against the taxonomy. Record facts in snapshot.json; never pass a claim to the scripts hidden inside prose. Research structural representatives before extensions. Add sourced admission records (leader/peer/satellite), not just roles: size/turnover leadership alone does not establish business leadership, and low R² alone does not establish useful independent information. Use candidate reason and evidence to explain the observed variable; several complementary leaders may share a theme. Unmapped candidates stay deferred, never in a permanent OTHERS or provider-industry catch-all. Do not promote a candidate's role merely to satisfy a coverage constraint.

5. Declare in measurement how each metric was produced. A window-dependent statistic — liquidity, factor_r2, beta_strength, beta_stability — must be computed, not estimated, and the builder refuses to accept it as judgement. If you have a table of daily bars, compute them instead of arguing with the gate:

   python scripts/universe.py measure --prices prices.csv --benchmark BINANCE:BTCUSDT.P \
     --source https://data.binance.vision/ --into snapshot.json --output snapshot.measured.json
   

   See references/measurement.md. If you have no price table, the candidates that need those metrics do not belong in the universe yet.

6. Supply a dated active listing check, admission reason, strong evidence and per-ticker measurement_record. Cite current sources for listing status, venue, liquidity and every non-obvious admission. If an essential fact cannot be verified, exclude the candidate or mark the snapshot incomplete. For a market outside the fourteen, also research its rules and declare them in the snapshot's market_spec — venues, symbol shape, identity rule and legacy-compatible breadth/guidance, with tier 1 or tier 2 evidence. Current sizes still come from coverage_plan.budgets. See references/data-contracts.md; do not guess a venue code or a symbol format, and say in your answer that the rules were declared, not reviewed.

7. Run:

   python scripts/universe.py build \
     --spec build-spec.json \
     --snapshot snapshot.json \
     --output output
   

8. A blocked or underfilled build starts a repair loop, not an immediate final error. Read references/recovery.md: preserve the checkpoint, research the diagnostics, remeasure and resume. Normally allow up to three materially different repair rounds within the user's scope. Stop sooner when the accessible source universe is exhausted or a required external input is unavailable; explain that boundary and retain a continuation. A small mapped snapshot does not prove source exhaustion. Repair necessary coverage before researching optional depth. Max must add at least 30% of Heavy's entity count, entirely as qualified Beta. The shortfall handler defaults to auto: a count-only gap within 5% of the required entity count may be delivered as partial, with qualified members, unchanged Heavy and planned group quotas, explicit counts/growth, and -partial filenames. Every other check must pass. Use --shortfall-action retry to require full growth and analyze/repair instead; deliver still cannot waive the 5% bound or any other check. Larger gaps remain needs_research. Diagnose business/gauges, missing groups and source/capacity limits, repair and resume; never call partial delivery complete. Unspent capacity above the minimum is allowed. Reference instruments do not count toward growth. Never weaken gates or hide the difference between a ceiling and actual membership.

9. The command prints the path of every artifact it wrote; they are named {market}-{profile}-{as_of}. Run validate on the universe path even though the builder validates before writing. Never present an output that fails its declared delivery contract. A validated partial has passed: true, qualified: false; disclose its shortfall prominently. Review the content too: each economic duty still has a qualified representative, gauges remain observable, and role shortages are understood. A filled count with poor duty coverage is not a completed research result. Core supply far below its policy target calls for role/business research, not padding.

10. Deliver the HTML report, Markdown, TradingView TXT, authoritative JSON and validation paths following the output contract. If no destination was requested, use ticker-universes/<market>/<profile>/<as_of>/ in the user's project, passing an absolute --output path. Keep checkpoints, use a new empty directory for revisions, and disclose source cutoffs and partial status. Keep rendered paths in place; deliver links or copies. English is always generated; non-English markets also get their market-language report, in both HTML and Markdown. HTML is a complete, offline browser report; keep it beside its companion artifacts for relative links. Read paths from artifacts.html_reports / artifacts.reports, keyed by language. Fixed headings come from locales; supply authored report_translations.en for human content (names, brief reasons, themes and descriptions). Check the English report's content as well as its headings; the CLI does not translate prose. Supply a one-sentence reason_summary for each candidate: explain the business observation, preferably at most 60 Chinese characters or 160 characters in other languages. Keep detailed reasoning and evidence in reason/admission/JSON; HTML and Markdown are readable briefs, not logs. State the business directly, without repetitive prefixes such as “观察” or “Observe”.

If you have a price table covering the window after a universe was built, run evaluate --universe U --prices P before proposing the next set of changes. It reports whether the instrument saw the largest moves, which exclusion code cost the most, and which metric ordered anything — see references/evaluation.md. It is not a backtest and never reports what the universe "returned".

To compare two universes — two sessions, two months, two people — run diff before.json after.json. It leads with market_spec, because for a declared market two sessions that researched the venue list differently did not build two versions of one universe.

Use --seed heavy.json for Max. It must be a qualified Heavy from the same plan and source date; only sourced satellites with measured liquidity may be added. For a narrower depth, rebuild from the same reviewed roster. Plan for at least ceil(0.30 * H) new entities, where H is the actual Heavy entity count. Check sector, satellite and export capacity before research; the new growth minimum does not waive those gates. Update Heavy first when facts or necessary representatives change.

The model researches core business leadership and quality, economic branches and information gain. For Beta, verify broad business/token identity, complementarity to named core members and sourced market cap; do not spend time on detailed profitability or tokenomics analysis. Eligible Beta are ranked by equity market cap (stocks) or circulating USD market cap (Crypto) within the planned distribution. Listing, sourced liquidity and measurement provenance remain mandatory. Price beta, R² and stability are auxiliary descriptors for coverage satellites, not admission thresholds or ranking inputs. Missing factor metrics are disclosed; legacy high-beta replay retains its numeric gates. Python owns evidence contracts, protected coverage, sector/satellite ceilings, identity, hashing and rendering. Never hand-write the final txt. See the 0.7 design.

Maintain an existing universe

1. Load universe.json, not just the txt. The JSON carries the version and the audit history.

2. Refresh liveness, venue, liquidity, theme leadership, factor redundancy and event evidence as the market overlay specifies.

3. Produce a small changes.json against the exact base_version_hash. Prefer NO_CHANGE when fresh evidence does not justify churn.

4. Run:

   python scripts/universe.py maintain \
     --universe universe.json \
     --changes changes.json \
     --output output
   

5. A hard finding means no new universe. Repair the proposal; never bypass the validator. Report warnings, additions, removals, turnover and deferred candidates in the Markdown result.

A verdict must come with an operation. Judging a theme obsolete or missing and writing it down as a note for a later round is how a universe rots: use ADD_THEME, UPDATE_THEME, REMOVE_THEME, REFRESH, MOVE and REPLACE in the same round, or record the candidate in deferred so the next round inherits it.

Non-negotiable boundaries

• Never invent a ticker, venue, listing state, liquidity number, theme relationship or source.
• Never select a name solely because it is popular or recently rose.
• Preserve benchmarks and anchors before adding satellites.
• Light is leader-only; Medium covers most reviewed leaders; Heavy completes the necessary leader/peer skeleton plus at most 20% satellites; Max adds at least 30% to qualified Heavy, entirely as Beta for complete delivery; a small count-only gap may be delivered explicitly as partial under the recovery contract. Max still has at most 35% satellites overall. The latter is a quality ceiling, distinct from the required expansion minimum.
• Every optional satellite must explain its incremental value relative to named core members. No BREADTH_PROXY/tactical fallback, no padding to satisfy growth, no hot-theme budget inflation.
• Preserve direct reference instruments and explicitly explain substitutions; references do not consume leader seats. A passed contract does not independently establish leadership truth.
• No return guarantees, allocations, order instructions or trade execution. This skill produces an observation instrument, not investment advice.
• Keep each TradingView file at or below 1,000 tokens including ### section headers.