The missed dates count as losses
The month's arc
The weeklies trace a change from access controls to promises of restraint, then to disputes about who could enforce either. The financing thread moved from filed commitments to delayed infrastructure. What stalled was proof that either industry's proposed oversight or its construction schedules could keep up with the contracts already signed. SourcesCCCC
The score
The ledger at publication contains 74 calls: 58 open, 11 correct and five wrong, with no partial or void resolutions. No open call has a resolution date before October 1st. The two September 30th deadlines were already settled in the ledger before this wrap: no OpenAI September IPO, correct; a public Anthropic by month-end, wrong. This wrap leaves every claim and confidence unchanged. SourcesA
The cumulative is 0.20693125, displayed by the scorer as 0.207. The September reckoning is 0.24009. Its cohort includes the August public-filing deadline resolved on September 1st and the September deadlines resolved on October 1st; it is a resolution cohort, not a set of predictions all made in September. SourcesA
| Confidence bucket | Actual correct | Calls |
|---|---|---|
| 60% | 62% | 8 |
| 70% | 50% | 4 |
| 80% | 100% | 4 |
This reproduces predictions.py score, which rounds confidence into buckets and the displayed actual rate to whole percentages. The first bucket's exact hit rate is 62.5%. The 70% bucket is overconfident on this sample; the 80% bucket has yet to lose. Neither four-call bucket justifies a general change in confidence. Two correct calls also share the same acquisition announcement, so the book contains fewer independent tests than its row count suggests. SourcesA
Every miss, with the reason it failed
Anthropic's public S-1, twice: timing error followed by an inadequate update. Prediction 2026-08-24-F1 assigned 55% to an August 31st public filing. Prediction 2026-09-01-F1 raised that to 65% for September 30th. Both lost. A confidential reaching reporters is not a public S-1 on . Repeating the call with more confidence required evidence about the filing step itself; a growing IPO narrative did not supply it. The ledger's search settled the second deadline. Neither loss becomes partial because a listing may eventually happen. SourcesAA
Astra's release: the wrong model of the decision. Prediction 2026-08-19-T1 put 72% on no general release of a Critical-cyber-tier model before year-end. Astra reached paying ChatGPT tiers while specific cyber capabilities remained gated. Liability concerns did not produce the release prohibition we inferred. This was a reasoning error about how a lab could separate model access from feature access, not a release that merely arrived a few weeks early. SourcesB
Enflame: enthusiasm was a poor estimator of the close. Prediction 2026-09-06-F1 assigned 65% to a first-day close of at least 426.54 yuan, three times the 142.18-yuan offer. It closed at 397. Oversubscription did not establish a closing-price floor. The direction was right and the threshold wrong; the score is still wrong. The September 13th weekly went too far in treating this miss as proof that IPO frenzy had broken: a close about 2.79 times the offer still showed strong demand. A correction is appended there. SourcesBB
Sonnet pricing: a missed fact, carried into September. Prediction 2026-08-06-T3 was already resolved wrong on August 30th. Its note records that the desk took three weeks to catch cancellation of the planned price increase. It is outside the September resolution cohort but remains in the cumulative score. Worse, the technology library still described the increase as forthcoming. That stale description is now corrected. Sonnet 5.5's announcement explicitly retains Sonnet 5's $2 input and $10 output price per million . This was a data-maintenance failure as well as a failed pricing-power thesis. SourcesAA
The six correct September resolutions
| Call | What settled it |
|---|---|
| Prediction 2026-08-25-B1 | Hugging Face agreed to a sale within the call's window. |
| Prediction 2026-08-27-B1 | Nvidia announced that agreement before September ended. |
| Prediction 2026-08-27-F1 | Moonshot submitted the Hong Kong application under the call's reporting standard. |
| Prediction 2026-09-06-F2 | A named Google compute commitment cleared the existing criterion's threshold. |
| Prediction 2026-08-26-B1 | Kimi K3 became a first-party Bedrock model. |
| Prediction 2026-08-06-F2 | September ended without an OpenAI IPO. |
The permanent resolution notes retain the settling sources. The compute call's criterion explicitly permits named commitments reported by established outlets before a public S-1; scoring it correct does not imply that an S-1 appeared. SourcesA
Library pruning
Technology. Opus 5 and Sonnet 5 are now described as earlier generations, with their successors linked in the existing records. The obsolete Sonnet price-rise watch is removed. DeepSeek documents retirement of the hosted V4 Flash and Flash Vision models, whose legacy names now route to V4.1 Flash. Both records now say so. DeepSeek's announced Pro retirement was reversed, so the Pro record is retained and its expired August urgency is removed. A successor release does not by itself establish that the older API has shut down. This review does not certify every repository's maintenance status. SourcesAAAC
Completed listings. HKEX confirms Camsense's September 30th listing; its record moves from priced to listed. AIAI's September Form 8-K names its Nasdaq trading symbol and exchange, correcting a record still labeled filed. Transwarp and Ligent were already marked listed, but their descriptions still spoke about upcoming debuts; those descriptions now agree with the exchange record. These are catalog corrections, not claims that all these debuts happened on publication day. SourcesAA
| Remaining filing | Month-end disposition |
|---|---|
| OpenAI | Filed, delayed. The September IPO call settled correct; no completed listing. |
| Anthropic | Confidential filing retained. The public S-1 deadline failed. |
| SB Energy | Public filing retained; the last weekly records a confirmed delay. No withdrawal is established. |
| Moonshot | Filed; dual-venue exploration does not establish a completed offering. |
| Nscale | Public filing retained; convertible financing does not count as IPO pricing. |
| Agility | Filed; its September 29th board announcement still makes appointments conditional on merger completion. |
| Jio | Filed; its issuer page provides draft offer documents; completion is unverified. |
| Microintelligence | Filed; the notice clears an overseas issuance process; no debut is established. |
| ESWIN | Filed; the subscription timetable points to October 9th, after this coverage window. |
| REDLattice | Merger disclosure filed; the record does not establish transaction completion. |
Existing filing evidence is preserved in each venture record. Fresh issuer or regulator checks support the Agility, Jio and Microintelligence dispositions; DBS supplies ESWIN's subscription timetable. The review has not independently re-established every rumored offering or every company's current operating health. Later developments without confirmation remain open checks. SourcesAAAA
Media. The existing agent-civilizations record now links to the September 6th weekly that revisited it and to this wrap. Its note identifies it as a reconstruction to read alongside the subsequent incident reporting, with the promotional ranking removed. No episode is entered twice. The media validator checks the catalog; it is not a certification that every external player remains available. SourcesC
Still contested
The summit's announced incident channel is not yet evidence of binding operating rules; Prediction 2026-09-20-B2 stays open until its October 8th test. Anthropic's reported and revenue pace remain distinct from final IPO pricing and audited annual revenue. The SB Energy timing dispute developed into a reported delay, but that does not settle its November pricing call. The long-horizon containment calls also remain open: incidents support the thesis without proving the absence of a future architectural fix. SourcesCA
Sources
See the source register. Prediction outcomes use the permanent ledger; the monthly does not replace its resolution notes.