stackfund-crowd-scenario

name	stackfund-crowd-scenario
description	Rehearse (not predict) how Taiwan retail ETF investor archetypes might react to an already-computed market event, producing a second-order reaction-chain narrative plus a categorical, NON-AUTHORITATIVE crowd_consensus (bearish / neutral / bullish — no numeric scalar). Scenario stress-test, not sentiment forecast. Use for the Pro crowd-scenario report. It never decides any number, never influences any action, and never writes back to the decision layer.
license	MIT
version	1.0.0
metadata	{"hermes":{"tags":["scenario","crowd","consensus","ETF","Taiwan","narrative","non-authoritative"],"firewall":"non-authoritative"}}

StackFund — Crowd Scenario Engine (FACE / non-authoritative)

Tier A disclaimer (load banner): 情境推演·合成人格·非真實民意·不是預測·未經回測;此層只產敘事與一個非權威標註,不決定任何數字、不回寫決策層、不下任何證券單。

This is the headline FACE, deliberately firewalled from the decision path. It reads a frozen ScenarioSeed (bucketed ordinal context only — never raw numbers) and emits a CrowdNarrative (narrative + persona reactions + a categorical crowd_consensus ∈ {bearish, neutral, bullish}, non_authoritative = true). There is no numeric / decision-shaped scalar; the model only writes persona reaction text.

When to use

• Generate the Pro report's second-order reaction-chain narrative for one event.

How to run

# dry-run = zero LLM, zero spend (CI + demo fallback safe)
python ${HERMES_SKILL_DIR}/scripts/run_scenario.py --symbol 0056 --scenario 0056_cut

--scenario is a free-form context tag (bucketed to ordinal context — never raw numbers). Useful labels: 升息 / 降息 / 0056_cut(配息調整)/ 電子權值回檔 / 高股息追捧. Same seed + label → deterministic consensus.

Hard rules (see references/firewall.md)

• G1 output whitelist: narrative / persona samples / stance counts / chain / a categorical crowd_consensus label (no numeric scalar). Any price/NAV/NTD pattern is a schema violation → abort.
• G2 non_authoritative is hard-wired true.
• G3 L4/L5 never import this skill's artifact (import-graph CI).
• G4 the model computes no number; numeric tokens in persona text are stripped.
• G5 determinism: references/seed.lock.json is the single source of truth.

G1/G2 are schema-encoded + tested. The committed crowd_narrative schema pins them: additionalProperties:false is the firewall's teeth — any numeric / decision-shaped scalar is rejected — and non_authoritative is const: true. tests/test_schema_validation.py asserts a conforming narrative passes and a tampered one fails.

Worked example

User: 推演 0056 在降息情境下散戶可能的反應

python ${HERMES_SKILL_DIR}/scripts/run_scenario.py --symbol 0056 --scenario 0056_cut → emits (the crowd_narrative schema pins the shape — a sneaked scalar fails it):

{"artifact_type": "CrowdNarrative", "crowd_consensus": "bullish", "non_authoritative": true,
 "synthetic_population": true, "n_personas": 30, "seed_id": "seed_0056_…",
 "narrative_md": "## 群眾情境推演 …"}

Present it as scenario rehearsal — never a number, never a call:

〔情境推演 · 合成人格 · 非權威 · 已與決策層隔離〕在這個降息假想情境下,合成散戶人格 整體偏多(crowd_consensus = bullish)。這是反應鏈推演,不是預測,也不回寫任何權重或委託。 真正的決策請看 stackfund-etf-analysis 的確定性引擎;兩者分歧時,以引擎為準。

References

• references/personas.md — closed-set archetype taxonomy + behavioural priors.
• references/seed.lock.json — rng_seed / model_id / temperature=0.
• references/firewall.md — the executable firewall contract.
• references/output-schema.md — the committed crowd_narrative schema in schemas/, checked by tests/test_schema_validation.py.
• references/clean-room.md — AGPL boundary (honest provenance).