HUMANOIDLEDGER
方法論——引用任何數字之前請先睇。方法論本身就係信任資產：下面寫嘅係規則，唔係摘要。

Methodology

How the Humanoid Ledger is built, what each number measures, and what it cannot tell you. Last updated 2026-11-08.

1. 本站係咩

一本人形機械人嘅公開數據賬簿：核實過嘅規格、一條發布年表、官方影片，再加全網已分析嘅報道——以及由呢啲報道建起嘅三本賬（宣稱 vs 實測、評分、問題）。賬簿獨立，同所列任何廠商都冇隸屬關係。記賬嘅規矩好簡單：有出處先記，冇核實嘅寧願寫「未公布」，都唔估一個數字上去。

2. Seed data — verified before entry

The 12 models, 31 release events and 8 official videos were each checked against a vendor page or a trade-press report before entry (seeded 2026-11-06). Every row carries its source link. Prices are vendor-published only: target prices, estimates and unconfirmed reports live in the notes field, never in the price column. Where a figure is unpublished or unverified, the cell says Not stated rather than a guess.

3. Live coverage — sources and the relevance gate

Hacker News (Algolia API), Google News RSS (with Bing News RSS as fallback when Google rate-limits the Worker egress), tech-blog RSS (The Verge, Engadget, Wired, TechCrunch, IEEE Spectrum, The Robot Report, PCMag, Tom’s Guide, 9to5Google, Road to VR, UploadVR), YouTube search pages, and Reddit via Google News. Ingest runs hourly. An item counts for a model only when its headline names the product — vendor + model where the name alone is ambiguous (e.g. “Apollo” only counts with “Apptronik”, “Digit” only with “Agility”, “Atlas” only with “Boston Dynamics”). Items that fail the gate are counted as skipped, not silently dropped.

4. Per-article analysis

Each article is analyzed once by Workers AI (Llama 3.1 8B, max 2048 tokens, JSON only): a 1–3 paragraph summary plus structured extraction — sentiment index, comparison, usage time, recommendation index, per-aspect scores, verifiable claims, issues, and the author’s verdict. The model is instructed never to invent numbers, quotes, durations or facts not present in the evidence; undeterminable fields are null, never invented.

5. Evidence checking (what survives, what is dropped)

Two hard guards run on every analysis before it is stored. First, a compared product is kept only if its name appears in the evidence text (title + snippet + body excerpt); anything else is dropped as model invention, and the compares-with flag is recomputed from the surviving list. Second, a verdict quote is kept only if the exact quote appears in the evidence text; otherwise it is stored as null. Aspect scores are restricted to eight aspects — locomotion, dexterity, battery, payload, autonomy, safety, software, price — each scored −100 to +100 only when the article actually discusses that aspect.

6. The scoring decision — what the three retained fields are, and are not

Decision (2026-11-07): the sentiment index (0–100), the recommendation index (0–100) and the author verdict (buy / wait / skip / mixed, with the author’s own words) are retained in full, in the article cards, in the metrics pages and in the three ledgers. They are kept because they measure the source content and its author — the stance of an article, and the conclusion its author reached — not this site’s judgement of any product. This site issues no rating, no ranking, no recommendation and no prediction of its own: there is no site score anywhere on the ledger, no “best model” list, and no verdict written by us. That is the whole of the exception, and it is the only exception: the general rule for this ledger is the same as the no-scores rule — record what happened, cite the source, add no judgement of our own. The three retained fields are presented with their provenance (which article, analyzed from snippet or full text, with what confidence), with nulls counted and never filled in, so a reader can always see how thin the underlying signal is — for example, the recommendation index is not determined in 308 of 329 analyzed articles, because most coverage is news, not a buying recommendation.

7. 三本賬

宣稱 vs 實測攞每篇分析入面嘅 key_claims；只有模型確實講咗實測值，宣稱先會帶實測值列出（截至 2026-11-08，289 條宣稱入面有 55 條）。其餘係廠商或者來源嘅宣稱，按原樣記錄——淨係列出，唔作驗證。評分彙總 sentiment_index 同 recommendation_index（分布、計數、淨係對非空值取均值）以及作者結論計數。問題按 issues[] 嘅逐字相同聚類，再按型號分組。某型號喺賬入面冇行，就寫「無記錄」，絕對唔寫成 0。即時端點：/api/ledger/claims-vs-measured、/api/ledger/scores、/api/ledger/issues（全部接受 ?model=<key>）。

8. Limitations — read before quoting a number

9. Corrections

Errors are corrected in the open and logged here, not silently fixed.

10. 歸檔——2026 年 11 月 7 號新增

Five vendor documents are archived as a baseline (Figure 03 announcement, Unitree G1 product page, Boston Dynamics Atlas product page, Agility Digit solutions page, 1X NEO order page), each with a SHA-256 fingerprint of the raw fetch and of the cleaned text, a 6-character seal and a 64-bit SimHash. The baseline is the first archived version; every later check compares the live document against it and records any change in doc_changes (currently 0 rows — 0 changes means nothing has changed since the baseline, not that nothing is watched). Live endpoint: /api/archive; page: /archive/. The Tesla Optimus official page could not be baselined: tesla.com returns 403 to this fetcher. It will be retried with a rendered fetch; until then it is listed on /archive/ as not baselined, not silently omitted.

11. 開放數據——2026 年 11 月 7 號新增

Every dataset behind this site is open JSON with no key. /data/ lists the catalogue with denominators and coverage; /api/open-data returns the same catalogue live, including a row count for every D1 table. Live endpoint: /api/open-data.

12. 二期表——持久表，2026 年 11 月 7 號回填

Four persistent tables were backfilled from the 330 analyzed articles (snapshot 7 Nov 2026, 20:10 EST) and are served live with a ?model= filter. experience_records (55 rows): every key_claim that carries a measured value, classified by metric (locomotion, dexterity, battery, payload, autonomy, safety, software, price, other) — the value is the number as stated in the article, not an independent lab measurement. review_scores (224 rows): one row per analyzed article that carries a sentiment index, a recommendation index or a verdict — metadata only (reviewer = article source, score / 100, date, URL, aspects tested), no quoted opinion text. failure_modes (93 rows): issues clustered per model by identical wording, with mention counts, first/last seen dates and up to 3 source URLs. resale_snapshots (1 row): the G1, with used_price_median NULL — no used-sold market exists for humanoid robots in any reachable public source (see §13). Live endpoints: /api/experience, /api/review-scores, /api/failure-modes, /api/resale.

13. 歷史——過去、而家同未來三段完整，仍然係 0 嘅項同原因

賬簿按歷史、而家、未來三段完整記錄，絕對唔會以而家嘅 0 做起點、淨係等未來變化先記錄。歷史係基線，冇基線，未來變化就冇意義，所以每個維度都要按公開來源同已採文章所及嘅範圍回填。2026 年 11 月 7 號回填結果（330 篇已分析文章，2026-11-07 20:10 EST 快照）同已核實嘅公開記錄：

14. 來源