Changelog
What changed, and when
Last updated: September 27, 2026 · Updated as verdicts change
Everything that changed on the evidence layer, dated and real. Every entry corresponds to a shipped change; score history is machine-readable. If you would rather not check back: subscribers get one email when something that matters actually changes — a verdict, the score, a new tool — and silence otherwise.
Changes
| 2026-09-27 | Four independent graders beside our verdicts, one disagreement; check the records in your browser New page /grader-consensus puts our eight verdicts next to every dated public grading of Situational Awareness we could find, each quoted verbatim with a link, under a counting rule printed on the page. Four graders so far (Delisle on LessWrong, Reeves on AGI Friday, Harris on the EA Forum, Dubach on his own site): 16 comparisons, 9 exact agreements, 6 in the same direction, 1 outright disagreement — Reeves calls knowledge work unresolved where we say On track. Two predictions (intelligence explosion, superintelligence) have no outside grade. /calibration gains a button that hashes the anchored records in your browser, checks them against the OpenTimestamps manifest and recomputes the Thesis Tracker from /data.json; nothing is fetched until you click. It also describes the forecast bot's sealed, Bitcoin-timestamped ledger, which has not submitted a forecast yet. |
| 2026-09-26 | AGI consensus board published; records anchored in Bitcoin; calibration page reads live data /agi-prediction-markets had been a 404 since 9 September: the weekly fetch could not find the small AGI contracts and Kalshi had renamed its price fields, and the step was set to pass anyway. Rebuilt as a consensus board — Polymarket, Kalshi, Manifold and Metaculus on 'AGI before 2027/2028/2030/2035/2040', cross-venue median and spread, recompute formula, machine-readable at /agi-consensus.json and MCP get_agi_consensus. data.json now carries each prediction's flip condition, resolution date or stated blocker (they were only on the homepage before). /calibration now regenerates from the ledger snapshot and states the Brier-eligible n (0). data.json, index-history.json, the board and the odds history get an OpenTimestamps proof per version from today (/ots/; pending until the calendar's transaction is in a Bitcoin block), so a version stamped from now on cannot be quietly backdated. The installable skill at /skill.md was being overwritten by the page mirror; fixed. |
| 2026-09-14 | Graded Amodei's call to slow down Dario Amodei published "We Must Pace the Frontier" on September 12 and Sam Altman endorsed it the same day. We checked it against all eight predictions: none moved. An announced intention is not a measured change in the rate of capability gain. The new page records what would have to show up in the data before a verdict does move. |
| 2026-08-27 | Refreshed the intelligence-explosion verdict page The page had not been touched since 30 June while the dataset was revised on 17 August. It now carries a dated status check quoting the ledger verbatim — still Pending, precondition undemonstrated — and names what is not the blocker: compute scaling reads On track. Cross-linked to the AI-2027 scenario explainer and the OOM compute page. |
| 2026-08-16 | The dataset now lives in git too Every verdict and every Thesis Tracker reading is mirrored to a public GitHub repository, rebuilt nightly. It only changes when a grade changes — so its commit history is an independent, timestamped record of when we changed our mind. |
| 2026-08-15 | Listed on the official MCP Registry com.agiscorecard/agi-scorecard is now on registry.modelcontextprotocol.io — AI agents can discover and call the Thesis Tracker, verdicts, market ledger and site search natively, no auth. |
| 2026-08-15 | First on-site subscriber The inline form converted its first reader end-to-end (homepage footer, Aug 14). The address is stored first-party; it syncs to the newsletter the moment the beehiiv key is configured. |
| 2026-08-14 | A dated status line on the most-read page "When will AGI arrive?" now opens with the two numbers that actually move: 504 days until the nearest forecast is settled, and 45 days with the Thesis Tracker stuck at 62.5 — the disagreement is moving, the evidence is not. |
| 2026-08-13 | English stock coverage completed Every US-listed name SunWatch tracks now has an English page — 18 of 18 — plus a new /en/stocks index. The names that are still Chinese-only are listed and labelled as such rather than quietly left out. |
| 2026-08-13 | The exposure tool now points at the one paid product /ai-stock-exposure is the most-used tool on the site and had no path to SunWatch, the only subscription on this domain with a real price. It has one now — with the public hit-and-miss ledger linked, not a claim. |
| 2026-08-13 | New tool: Your AGI Timeline One slider places the year you expect AGI against nine real published forecasts — bull, base case or bear — with a shareable result. Free, embeddable. |
| 2026-08-13 | Holdings-update promise made accurate The invest pages no longer promise a same-day rewrite of every 13F. Holdings change only after a filing is read line by line — and subscribers hear the day the page changes. |
| 2026-08-12 | Opinion hook on the Altman page The second-most-cited page now asks where you land versus Altman's "few thousand days" — one tap onto the forecaster spectrum. |
| 2026-08-11 | More English stock pages Rocket Lab and SK hynix ADR added to the English invest pages (five now), with a fix so Chinese-named tickers render properly in English. |
| 2026-08-10 | Traffic reporting corrected Server-side "human" pageviews turned out to include an unrecognised crawler fleet; real-reader counts now come from JS-executed pageviews, and a UA audit records who is actually knocking. |
| 2026-08-09 | English stock pages on the invest console SunWatch's per-ticker analysis now has English pages (SanDisk, Micron, SpaceX to start), with reciprocal hreflang. More each run. |
| 2026-08-08 | MCP server live at /mcp AI agents can now call get_thesis_tracker, get_verdicts and search_site natively over Model Context Protocol — free, no auth. |
| 2026-08-08 | Calibration page published Every probability-shaped claim inventoried; Brier score pre-committed at n≥20. Misses stay published. |
| 2026-08-08 | AGI-2027 resolution page + odds-vs-evidence series Resolution criteria locked 17 months before the deadline; issue #1 comparing prediction-market odds with the evidence read. |
| 2026-08-08 | Site search with demand logging Full-site search (EN+中文) at /search; queries and zero-result gaps now drive what gets built next. |
| 2026-08-08 | Red-team desk on the invest console Survival odds (52–65%) for every open market call, strongest counter-case included; space-top confidence cut 60%→52% same day. |
| 2026-08-08 | Investing sub-sites merged under one domain SunWatch → invest.agiscorecard.com, AI Investing Compass → compass.agiscorecard.com; English landing + language negotiation added. |
| 2026-08-08 | One-tap flip poll on the Thesis Tracker Pick which open prediction flips first; per-prediction email alert stores your pick as the topic. |
| 2026-08-06 | On-site signup, no redirect All subscribe CTAs now open an inline form; addresses stored first-party before forwarding, per-context promises (e.g. 'one email, only if this verdict flips'). |
| 2026-08-05 | First-party analytics + shareable baskets Server-side pageviews (counts AI crawlers ad blockers miss); /ai-stock-exposure baskets permalinked via ?b=NVDA-AMD-TSM. |
| 2026-07-26 | /ai-tools hub All free tools in one place; four embeddable via iframe. |
| 2026-07-25 | Site restructured: Research · Play · Invest Three-section navigation and directory; Chinese pages moved to a unified light design system. |
| 2026-07-12 | AGI-2027 Thesis Tracker launched One auditable 0–100 score recomputed from the 8 verdicts; history at /index-history.json. The AGI type test (/agi-test) shipped the same day. |
| 2026-06-30 | Scorecard launched 8 Situational Awareness predictions graded with pre-registered flip conditions; machine-readable at /data.json (CC BY 4.0). |
Thesis Tracker score history
| Date | Score |
|---|---|
| 2026-09-06 | 62.5/100 |
| 2026-08-08 | 62.5/100 |
| 2026-07-12 | 62.5/100 |
| 2026-06-30 | 62.5/100 |
Score points are stamped in /index-history.json (CC BY 4.0); the score only moves when a verdict changes.
Frequently asked questions
How often does this page update?
Whenever something real ships: a verdict change, a score move, a new tool or dataset. Entries are dated and correspond to public commits — nothing is backfilled or invented.
What counts as a change worth logging?
Verdict flips, Thesis Tracker score moves, new tools and datasets, and structural site changes. Routine copy edits do not qualify.
The live scorecard updates as models ship and verdicts change.
View the live scorecard →