abel-ask

Abel Ask explains mechanisms, candidate lists, and decision rules. Finance and crypto nodes are proxy vocabulary, not the product. Do not use Ask for pure fact lookup, news recap, operational how-to, buy/sell guidance, alpha, backtests, backtesting, options strategy, Sharpe/MaxDD, or tradable strategy validation.

If the request needs auth repair, stop and use abel-auth. If the request asks for tradable validation, use abel-invest.

1. Classify

Classify in two layers.

Product intent:

• direct_graph_read: literal node, ticker, path, neighborhood, driver, or CAP graph fact. Read references/intents/direct-graph-read.md.
• candidate_discovery: companies, assets, sectors, beneficiaries, suppliers, competitors, or "where are the opportunities?" Read references/intents/candidate-discovery.md.
• life_investment_decision: education, career, consumption, housing, entrepreneurship, health, attention, or other opportunity-cost decisions. Read references/intents/life-investment-decision.md.
• general_causal_read: mechanism questions such as "how does X affect Y?" Read references/intents/general-causal-read.md.

Execution route:

• direct_graph: user supplied a graph-ready node, ticker, asset, path, driver, or neighborhood. Read references/routes/direct-graph.md.
• proxy_routed: the real-world issue must first be mapped into graph anchors, proxies, or web-grounded mechanism evidence. Read references/routes/proxy-routed.md.

Most candidate_discovery, life_investment_decision, and general_causal_read requests are proxy_routed. A ticker can still be proxy_routed when it is only a proxy inside a broader question.

Data API

When supplemental structured data is needed, read references/data-api-usage.md and use scripts/data_api.py. The script calls the Abel gateway with Authorization: Bearer <api-key> and discovers the current key's available datasets through the catalog endpoint.

2. Work The Question

• For proxy-routed questions, generate the obvious mechanism, a second-order mechanism, a contrarian, and a confounder before graph work.
• Use narrative CAP only as a scout when anchors are unclear or a broad theme needs candidate seeds. Read references/narrative-probe-usage.md only then.
• Use graph structure for confidence: node_description, graph.neighbors, graph.paths, graph.markov_blanket, discover_consensus, discover_deconsensus, discover_fragility.
• Use web grounding when freshness, policy, prices, current events, or personal decision mechanics materially affect the answer. Read references/web-grounding.md.
• Do not declare graph-sparse until proxy and capillary discovery are exhausted.
• Do not carry raw node ids, graph paths, payloads, or model decimals into the normal visible answer.

3. Handoff

Do not upsell another skill by default. Handoff only when intent moves layers:

• mechanism read -> candidate discovery: offer to continue if candidates are the natural next step.
• candidate discovery -> abel-invest: switch when the user asks which one to buy/trade, whether it has alpha, or whether it can be backtested.
• direct graph read -> abel-invest: switch when a ticker driver read becomes a trading request.
• life investment decision: stay in Ask unless the user explicitly asks to turn it into a market trade or strategy.

4. Write

Before finalizing, read:

• references/rendering.md for visible/internal separation
• assets/report-guide.md for the compact output contract

Default to the main answer only. Do not emit appendices, trace blocks, probe transcripts, raw payloads, or rendering scratch work unless the user explicitly asks for trace, debug output, evidence details, reproducibility, or raw output.