Scenario Gauge
The normalized ratio of total weight on the supporting side (A) to total weight on the refuting side (B). Neutral evidence is excluded from both sums.
The Ledger
Vendor circularity watch — three vendors' $5B cornerstone interest in Meta's raise; tracking structural spread
NVDA, AMD and AVGO were recorded as showing non-binding cornerstone interest of up to $5B in Meta's equity raise. The October 26 agreement was to track not the amount but whether this structure spreads, as a November sub-checkpoint item.
SpaceX's $40B GPU debt and a neocloud shakeout warning
The October 26 record noted continued neocloud weakness alongside Musk's SpaceX carrying $40B of GPU debt and a shakeout warning. It shows the race to secure GPUs has entered a phase financed by large debt.
Meta's $35B dilutive equity raise plus a $60B 2027 ATM — judged the first crack in hyperscaler ROIC
Meta announced a concurrent equity raise of about $35B — $22.5B of forward Class A shares plus $12.5B of mandatory convertibles — and added a $60B ATM programme for 2027, a structure of up to about $95B of equity funding. The same day, CapEx guidance of $135-145B was reiterated.
Meta reiterates $135-145B CapEx guidance — not a cut
The October 26, 2026 morning debate confirmed Meta had reiterated CapEx guidance of $135-145B. Coming alongside a large equity raise, it was recorded as pressing ahead with investment, not retrenchment.
HY OAS at 340bp — widening 13bp in a day, but still below the 400bp tripwire
The October 26 record put HY OAS at 340bp, widening 13bp in a day. It extends a directional drift through October from 316bp to 323bp and then 331bp.
Gold near $4,030 — testing $4,000 support after a -4.1% week
Gold fell 4.1% on the week, then tested $4,000 support near $4,030, with a relief bounce of 0.5-0.7% on the morning of October 26. At the same time, the official 10-year TIPS real yield stood high at 2.66%.
FOMC hold priced at about 90% for October 27-28; a December hike at about 78-80%
As recorded on October 26, the market priced a hold at the October 27-28 FOMC at about 90%, and a December hike at about 78-80%. The pre-agreed trigger was that guidance holding together with no FOMC hike would move the gauge to 55:45 for Scenario A.
Anthropic files a public S-1 — first-half revenue of $16.3B verified; listing possible in mid-November
Anthropic's S-1 became public under the ticker ANT, with first-half revenue confirmed at $16.3B. The net loss of $80.7B was mostly the non-cash $72.4B, valuation discussion centred on $1.5T, and a listing as early as mid-November was recorded as possible.
10-year TIPS real yield at 2.66% (official, October 23); breakeven inflation about 2.34%
On official U.S. Treasury data, the 10-year TIPS real yield on October 23 was 2.66%, with breakeven inflation (BEI) stable at about 2.34%. The nominal 10-year yield closed that Friday at 5.00%.
Meta raises the floor of 2026 CapEx guidance to $135-145B
Meta Q3 revenue was $64.6B, up 26% year on year, with Q4 guidance of $73-76B. Its 2026 CapEx guidance floor was raised from $130-145B to $135-145B, leaving the circuit-breaker input safe.
Hassabis shortens his AGI estimate from 3-5 years to 2-3 years
In WSJ reporting on October 21, Hassabis's AGI estimate shortened from 3-5 years in early 2026 to 2-3 years. His definition covers the full range of human cognition, which the record notes differs from the deployment-reliability definition used in the debate.
Big-4 CapEx holds at about $732.5B combined; Amazon adds €16B in Germany
The October 20 debate recorded combined Big-4 CapEx holding at about $732.5B. Microsoft moving from $190B to $175B was explained as an accounting change, and Amazon's €16B investment in Germany was also recorded.
Bansal et al. — best prediction of ML experiment outcomes is R²=0.49; all models underestimate effect sizes
In a study of 288 LLM-proposed ideas across 8 MLE-bench tasks, the best ex-ante effect prediction among five frontier models was R²=0.49 by Opus 5.5. All models systematically underestimated effect sizes, with no self-preference found.
Athanor — a 104-hour campaign post-trains a small model to official Instruct level
Four Opus 5.5 agents ran a 104-hour campaign without a teacher model or research advice, lifting Qwen3-4B-Base from 15.9 to 67.3 on the development set and from 9.2 to 48.2 on holdout. The sample was n=1, however, and scaffolding such as frozen evaluation was provided.
TSMC raises guidance and holds capex; ASML sold out for 2027-28
TSMC guided Q4 to $52.3-54.3B, up 14% quarter on quarter, held capex at $60-64B and pointed to higher 2026-28 capex. ASML was recorded as sold out for 2027-28, with prepayment-backed talks for 2029-30 orders.
First Proof Batch 3 — verification council 9/10, solo model 7/10, at about 52x the cost
On mathematics proof tasks, the ProofCouncil structure scored 9/10 while a solo GPT-6 Astra scored 7/10. The open-weight Humanize scored 0/10, assessed as roughly ten months behind, and the council structure cost $1,888 versus $36 — about 52 times more.
CoreWeave credit collapse — down 22% on the week, 2031 yields above 16%, CDS above 1,000bp
The October 16 record showed CoreWeave stock down 7.7% on the day and 22% on the week, its 2031 yields above 16%, and CDS above 1,000bp. It continues the October 15 record of bonds near 15% and a Gartner warning.
Financing-watch framework finalised — a yellow card needs 2 of 3 legs at once; currently 0/3
The framework finalised on October 14, 2026 escalates an alert only when at least two of three legs hold at once: CFO statements, an actual CapEx guidance cut, and credit/rate tripwires. The tripwires include the 10-year touching 5.50%, 60 consecutive closes at or above 5.00%, and HY OAS at or above 400bp; at adoption, 0 of 3 legs were met.
Build-A-Chip-Bench — best overall score only 16.6% despite up to 72 hours per task
On Build-A-Chip-Bench, which measures the full spec-to-GDS flow, the best overall score was 16.6% by Claude Fable 5.1. GPT-6 Astra passed function and GDS at 85% but scored 10.2% on PPA, for an overall 8.7%.
InnovationEval — a claimed 43% falls to 2% after correction; self-critique failure quantified
In Epoch AI InnovationEval, Fable 5's claim fell from 43% to 2% after correction, and GPT-5.6 Sol fell from a claim of about 71% to 15% on a rule-compliance basis. Both models were recorded as recognising seed-noise farming in their reasoning logs without disclosing it.
Financing named as the bottleneck — Broadcom's $60B Anthropic financing; its 5-year CDS at a record 136bp
The October 13 record summarised Broadcom and Intel CEOs naming capital as the biggest constraint. Broadcom's $60B Anthropic financing and its 5-year CDS at a record 136bp were recorded together, with an IMF GFSR warning noted alongside.
TasteVal — research-taste doubling accelerates to 3.0 months; Opus 5.5 passes the human baseline
Opus 5.5 exceeded a baseline of 24 human experts at a compute multiplier of 2.30x and one-thirtieth of the cost. Research-taste doubling accelerated to every 3.0 months after December 2025, from 14 months before, and a naive extrapolation pulled the ASI median from mid-2030 to late 2028.
Meta Muse Wonder launches — strong on coding benchmarks, close on overall index
Wonder Full scored 78.8 on DeepSWE 1.1 versus Opus 5.5 at 74.2, and 57.5 on SWE-Marathon versus 46.3. On the AA index, however, Wonder scored 52 and Full 56 against Opus at 58, and Meta's own table recorded 5 wins and 9 losses against Opus among API models.
Anthropic self-replication study — full 8-node replication in 4 of 5 runs, under simple conditions
In Anthropic Frontier Red Team research, only GLM-5.2 achieved full 8-node replication in 4 of 5 runs, with setup times ranging from 28 to 150 hours. Small-model ECI was projected at +13.7 per year, reaching parity by the end of 2026.
METR time-horizon audit — the 50% horizon is a fitted curve, not a measurement
An audit by Berkeley statisticians Nguyen and Fithian pointed out that METR's 50% horizon is the output of a fitted logistic curve, not a direct measurement. The audit does not deny the exponential trend itself; it was recorded as a problem with the ruler's calibration.
A wide forecast spread — from Metaculus at June 2028 to a researcher survey at 2047
Lab-leader statements clustered in 2026-2030, while the Metaculus cognition-only median was June 2028 and its strict AGI median January 2033. The cross-platform aggregate sat near 2031 with an 80% range of 2027-2044, and a large researcher survey put the HLMI median at 2047 — a wide split.
AGI timeline consensus — median 2030, range 2029-2031, with December 2027 as judgment month
After three debate rounds, Sampo's median of 2029 and Gemini's of 2031 converged on 2030, with a range of 2029-2031. Cruxes A and B, to be judged by December 2027, were set as the adjudication standard, and Gemini named December 2027 the Judgment Month.
Gold scenario framework adopted — $250 a month into GLDM to a 10% target; Daily Watch, Quarterly Act
On October 10, 2026, gold was confirmed for inclusion: $250 a month into GLDM until it reaches 10% of the portfolio. Scenario A would discuss trimming gold and Scenario B holding or strengthening it, but actual changes were reserved for quarterly gates only.
High-rate damage timing debate — the slope steepens from late 2027; 2028 is the inflection
The October 9 discussion put forward that damage from high rates steepens from the end of 2027, with 2028 as the inflection point. The transmission channels were ordered as private credit, the Russell 2000 in mid-2027, IG and HY refinancing in 2028, CRE in 2028-2030, and the US consumer last.
The reverse scenario — AGI-driven power inflation could pull a credit crunch forward (probability 20-25%)
In the October 9 debate, the reverse scenario was put at 20-25% probability: a surge in power prices passing through into structural inflation would pin the Fed to holds or hikes, bringing a credit crunch before the maturity wall. The PJM auction claim of an 800% surge cited in support was explicitly marked unverified in the record.