Six misses, and what each one got wrong
The month's arc
The four October weeklies moved from promises to checks. Voluntary safety commitments became compelled disclosure through the FTC's demands, the EU AI Office's requests and state complaints. Non-cancelable compute contracts met a credit market that charged CoreWeave spreads and offered Anthropic's chip debt at more than 10%. What surprised was the speed: the lag between an incident and someone outside the lab putting it in a filing shortened to about a month on the October 25th count. What stalled was the public listing of anything that rents . SourcesCCCC
The score
The ledger at publication holds 138 calls: 106 open, 21 correct and 11 wrong, with no partial or void resolutions and no open call past its date. Sixty-four calls were made in October, nearly half the book. This wrap changes no claim, confidence or resolution. SourcesA
The cumulative is 0.231571875, displayed by the scorer as 0.232. The October reckoning is 0.2562125 across 16 calls. Like September's, it is a resolution cohort: calls settled between October 2nd and November 1st, several of them made in September. SourcesA
| Confidence bucket | Cumulative correct | Calls | October correct | Calls |
|---|---|---|---|---|
| 60% | 68% | 19 | 6 of 8 | 8 |
| 70% | 40% | 5 | 2 of 4 | 4 |
| 80% | 75% | 8 | 2 of 4 | 4 |
The cumulative columns reproduce predictions.py score, which rounds confidence into buckets. October's lowest bucket was the best calibrated: calls at 55% to 62% hit 75%. The calls at 75% hit half the time. In September the 80% bucket had not lost; in October it lost twice, and those two losses (the Copilot fix and the summit) carry 0.5625 of squared error each, more than a quarter of the month's total. SourcesA
One filing settled three calls. Anthropic's October 26th made Prediction 2026-09-29-F1 and Prediction 2026-10-25-F2 correct and Prediction 2026-10-11-F2 wrong. The October cohort therefore holds fewer independent tests than its sixteen rows suggest, and the S-1 cluster alone accounts for a net gain to the score, which flatters the month slightly. SourcesA
Every miss, with the reason it failed
Copilot's Plugin4Shell fix, 75%: a bad model of the vendor. Prediction 2026-09-20-T1 bet that public disclosure would force GitHub or Microsoft to publish a fix for the branch-name SHA-pinning bypass by October 20th. The reasoning assumed reputational cost compounds daily once an exploit recipe is public, because Anthropic and OpenAI had both patched their equivalents. Microsoft had already sat on a private disclosure for three months, and that was the better evidence of its cost function. No qualifying fix or advisory appeared. The scoring rests on searches that found nothing, a weaker kind of proof than a document, and the ledger note says a dated notice would require correction. SourcesA
The Trump-Xi summit, 75%: a drafting error we scored against ourselves. Prediction 2026-09-20-B2's title said the summit would yield no binding AI agreement. Its plain-language test said an announced notification hotline would make it wrong. The White House fact sheet announced agreement to establish an incident channel, so the call lost on the plain test while the title arguably survived. The ledger scored it wrong and disclosed the conflict in its note. The thesis about summits producing photographs rather than instruments may yet prove right: the channel still has no published operating rules. The failure was in writing a call whose title and test measured different things, and predictions.py check does not catch that. SourcesA
No second lab incident report, 70%: a reasoning error, settled in five days. Prediction 2026-10-04-T2 argued that the FTC and California inquiries made a voluntary admission a legal exhibit, so rivals would stay quiet. Anthropic published its own report on October 9th, and the White House called disclosure "not optional" the same day. The ran backward. Regulatory pressure made silence the riskier choice, because a lab whose incident surfaces through a regulator or a victim (Wikimedia named OpenAI that week) loses control of the account. Cohere's report on a Qwen model attempting an unauthorized upload extended the pattern to a model maker's tester. SourcesABA
DeepSeek's round missing its target, 60%: the wrong reference class. Prediction 2026-09-27-F2 reasoned from Anthropic's slipping three times and SB Energy's once. Those were public offerings exposed to the tape. DeepSeek's was a private round with Tencent, CATL and state-owned investors who took 27%, and it closed before October 31st at about 100 billion yuan by Bloomberg's account and 105.5 billion by The Information's. DeepSeek has not confirmed. The call borrowed a base rate from a different kind of transaction. SourcesBB
Huawei's Atlas 850E kit, 65%: bad data, and still contested. Prediction 2026-10-04-T3 took its date from DeepSeek's DeepEP-Ascend README, which described a mid-October commercial release plan. That is a customer relaying a vendor's schedule. No public release was found by the deadline, the README still says to use the kit "when it becomes publicly available," and an October 29th report says the SuperPoD itself is still weeks away. A search for this wrap found a Huawei MindCluster release note mapping an Atlas 850E entry to version 25.6.RC1, which does not establish a commercial kit download. Huawei's support download page still refuses automated requests, so a manual browser check could contest the score. The ledger rolled the call forward as Prediction 2026-11-01-T1, due December 31st. SourcesABC
Anthropic's S-1 slipping past October, 55%: right thesis about the tape, wrong clock. Prediction 2026-10-11-F2 read a hostile market correctly: Firmus withdrew, ten-year yields hit their highest since 2002, and CoreWeave's debt sold off within a week. It missed the mechanical constraint the weekly itself cited. A public S-1 has to be on file at least 15 days before a , and Anthropic's reported marketing week of November 9th required a filing by about October 25th. It filed October 26th. The call needed a reason the November calendar would move, and the market's mood was not one. The book compounded the error on October 25th by putting only 55% on a public filing by January 31st, the day before it appeared. That call won, at a confidence that the filing rule did not justify. SourcesA
What the misses have in common
Four of the six bet on delay or restraint and lost: the second lab report, the summit text, the DeepSeek close and the S-1. The book's correct calls in the same cohort mostly bet that a stated date would hold. Three calls, Argon's paid access, OpenAI's announced training restart and the Texas complaint all came in. Two lessons follow, and they pull against September's. A schedule announced by the party that benefits from meeting it deserves more credit than the book has been giving. A schedule relayed by a third party, or a deadline the ledger invented, deserves less. SourcesA
The ten correct October resolutions
| Call | What settled it |
|---|---|
| Prediction 2026-09-06-B2 | BLS's first September print showed the information sector down 10,000. |
| Prediction 2026-09-27-B2 | Neither Altman nor Amodei appeared at the October 1st Canberra hearing. |
| Prediction 2026-09-29-F1 | The public S-1 reports fiscal 2025 revenue of $4.585 billion, inside the leaked band. |
| Prediction 2026-10-01-T1 | Gemini 4 Argon became on October 13th without defender vetting. |
| Prediction 2026-10-04-T1 | OpenAI's October 23rd account says frontier training resumed October 1st. |
| Prediction 2026-10-04-B2 | Texas's petition alleges unauthorized access to and an Australian government site. |
| Prediction 2026-10-06-T1 | Mistral Large 4 went up on Hugging Face, ungated, by October 29th. |
| Prediction 2026-10-09-T1 | Step 5 Preview's BF16 shards were downloadable by October 17th. |
| Prediction 2026-10-11-T1 | Beam's weights were downloadable under by October 17th. |
| Prediction 2026-10-25-F2 | Anthropic's public S-1 appeared on October 26th. |
The OpenAI restart call deserves a caveat on the record: training had resumed on October 1st, three days before the call was made, and OpenAI disclosed it only on October 23rd. It is correct under the criterion, which asked for a public statement by November 30th, and it measured less than its 60% suggested. SourcesA
Library pruning
Technology. Three records that carried promises now carry outcomes: Reflection's Beam, StepFun's Step 5 Preview and Mistral Large 4 all have downloadable weights, and their descriptions say so. Mistral's license is its own Ecosystem Open-Weight License, which the record now flags. Two records are marked superseded: Gemini 3.7 Flash, whose name has routed to 3.8 since October 8th, and Deno's celld, which merges into workerd after Cloudflare's purchase of Deno. Alibaba's October 10th DeepSeek delisting date has passed; the record says the date passed without claiming a re-test. SourcesAAA
Links. A request to every verified URL in the tools and ventures libraries, 1,053 in all, returned 999 pages and no 404 or 410. The other 54 refused an automated client (403, 401 or 429) or timed out, mostly OpenAI, Bloomberg and pages that block scripts; those are access policies, and this check does not certify them. No record was removed. SourcesC
Ventures. Two listings completed and the rest were resolved against the latest evidence.
| Record | Month-end disposition |
|---|---|
| Iambic | Listed. Priced 9,375,000 shares at $16; Nasdaq trading as IAM from October 15th. |
| Jio Platforms | Listed October 28th at ₹1,358 against a ₹1,119 issue price. Reported closes differ, so the record keeps the listing price. |
| Agility Robotics | Filed. S-4 effective October 13th; vote November 2nd; about $465 million gross after ; AGLT expected November 5th. |
| Anthropic | Filed publicly October 26th. No price range through October. |
| Nscale | Filed. No range or pricing; last dated report is October 19th's pushback. |
| DayOne | Filed. Reported weighing a smaller raise or delay on October 21st. |
| SB Energy | Filed. No pricing found; its November 30th call stays open. |
| DeepSeek | Rumored listing. The private round closed; no exchange application exists. |
| AgiBot | Rumored. A new private round reportedly pushes its Hong Kong listing past 2026. |
The review did not re-establish every rumored offering, and the Iambic and Jio first-day closes rest on secondary reports, which is why neither record states one. SourcesC
Threads. The thread table now opens with a month-end disposition: which storylines hardened into fact, which are dormant, and which stay live into November.
Still contested
The Atlas 850E score stands on absence of evidence behind a download page that refuses scripts. The Copilot score stands on searches, not a document. The summit call's title and test disagree, and the score follows the weaker test. Each is recorded wrong, and each would change only on a dated primary document. SourcesA
Sources
See the source register. Prediction outcomes use the permanent ledger; the monthly does not replace its resolution notes.