Method & Trust

How the Morning Wire is produced — the rules we enforce on ourselves, and the limits we won't hide.

本页说明 Morning Wire 的生产方法：我们对自己执行的硬规则，以及不隐瞒的局限。内容采集与生成由 AI 完成、经编辑流程成刊，不构成投资建议。

1 · How a story is made

Fetch — RSS from Indonesia–Singapore tech & business sources (e27, DailySocial, Katadata, Antara, Business Times, Straits Times), every day at 08:00 WIB via GitHub Actions.
每天 WIB 08:00 自动抓取印尼-新加坡科技商业媒体 RSS。
Triage — keyword relevance gate selects the Top 10 signals of the day; every story keeps its original source URL.
相关性闸门选出当日 Top 10，每条保留原文链接。
Three desks + Chief Editor — Rina Wijaya (market), Marcus Tan (capital), Dr. Aisha Rahman (policy) each analyze the story; the Chief Editor synthesizes. When no LLM key is available, rule templates fill in — and are labelled as such.
三位编辑（市场/资本/政策）分别研判，主编合成；无 LLM 密钥时回退规则模板并明确标注。
Graph context — each story is matched against the Archipelago Capital Graph; matched companies show disclosed investors and link back to the graph.
每条报道与资本图谱匹配，命中公司自动展示已披露投资方并回链图谱。
Publish — static site + RSS + open Funding API, deployed to Cloudflare Pages. No paywall on the wire itself.
静态站 + RSS + 开放 Funding API，部署到 Cloudflare Pages，晨报本身不设付费墙。

2 · Hard rules

Source binding Every factual claim in our analysis must trace back to the story's original source URL. If it can't, it doesn't ship — the story falls back to the rule template.
来源绑定：分析中的每个事实性论断必须能回溯到原始来源，回溯不了就不发。

Confidence labels high — every claim grounded in the source. medium — mostly grounded, flagged inference (mock output always lands here). low — thin or partially off-topic; low-confidence items require human spot-check before any paid-tier use.
置信度三级：high / medium / low，low 级禁止直接进付费场景。

Three desks ≠ independent verification The three perspectives run against the same model family — if the model hallucinates, all three desks can err consistently. The desks exist to force coverage (market, capital, policy angles all considered). Real verification comes from source binding, the confidence gate, and human spot-checks. We do not market this as "cross-validated by three independent agents".
三视角不等于独立验证：三位编辑跑在同一模型家族上，可能一致地错。它们的作用是强制覆盖三个角度；真正的验证来自来源绑定、置信度门与人工抽查。

Budget governance Daily LLM call ceiling (default 50). When exhausted, a fuse trips and remaining stories go Mock — the fuse state is recorded publicly in data/llm_usage/. Stories already published are never re-sent to the LLM.
预算治理：每日调用上限 50，熔断留痕，已发布内容不重跑。

3 · Data & graph

Entity data comes from the Archipelago Capital Graph — an event-sourced Neo4j knowledge graph of Indonesia–Singapore tech capital (335 companies, 72 funds, 405 policy events). Investor information on profiles and cards reflects disclosed graph records with source URLs and confidence levels — the graph deliberately refuses to fabricate uncovered relationships.
实体数据来自 ACG 图谱；投资方信息仅反映图上有来源披露的记录，图谱宁可留空也不编造。

Open surfaces: Funding API · Company directory · Source code (MIT)
开放接口：Funding API、公司目录、开源代码。

4 · Honest limitations

Limitation	What it means for you
Mock mode exists	Zero-key days run on rule templates. Every story is labelled with its `engine` (llm/mock) and `confidence` — check the byline.
Extraction is rule-based	Funding amounts, rounds and investors are extracted by deterministic patterns, not understanding. Treat single-source numbers as leads, not facts.
Graph coverage is evidence-gated	LP chains, ownership percentages and UBO are mostly uncovered by design. "No data" means "no evidence", never "no relationship".
Three correlated desks	See Hard rule #3. Spot-check anything you act on.
One person + automation	This is an automated newsroom, not a staffed desk. Volume is capped at 10 stories/day on purpose.

局限清单：Mock 模式存在且已标注；抽取基于规则不是理解；图谱覆盖是证据门控的——"无数据"意味着"无证据"，绝不意味着"无关系"。

5 · Compliance

本内容由 AI 采集与生成、经编辑流程成刊，仅供信息与研究参考，不构成投资建议。原文版权归原媒体所有。

Content is AI-collected and AI-generated, finished through an editorial process, for information and research purposes only — not investment advice. Original articles remain the property of their publishers.