Opportunity — worth a try?

A first-hand, publishable test of the position the book rests on: that a vendor's AI will never be permitted to tell a customer not to buy. Until this week that claim had to be argued from principle; it can now be run against shipped software. It exercises both halves of the owner's profile at once — GTM org diagnosis and hands-on evaluation — which is the stated sweet spot and the hardest thing for a commentator to copy.

Cheapest way to test it: Get a Salesforce developer org or trial with the Claudeforce plugin and run the 37 sales skills through a scripted qualification scenario where the honest answer is that the customer should not buy. Record what each skill does — qualify out, hedge, escalate to a human, or push on. One afternoon, no spend beyond the trial. The transcript is the artifact whether the answer is yes or no.

No pressure — it quietly expires 2026-11-30 if you ignore it · full reasoning on Signal

Opportunity — worth a try?

A publishable case study with everything on the record — named company, named CEO, a revenue miss given as the reason, employees saying security review got skipped, and a reversal that changed the mechanism and not the posture. Most AI-failure writing runs on anonymous anecdote or survey percentages. This one has the memo, the Slack messages and the all-hands audio behind it, which makes it citable in the manuscript rather than merely illustrative. It is organisational diagnosis rather than tool commentary, which is the half almost no AI commentator can copy.

Cheapest way to test it: Write 1,200 words on the public record only: what Proton was actually missing, which failure fired first, and the single structural question to put to the reorg when it lands next month. Publish on Work That Holds. If it gets picked up by anyone in the CIO or governance audience, that is the cheap read on whether this argument pulls advisory conversations — and if it doesn't, you have spent a day and gained a book chapter.

No pressure — it quietly expires 2026-12-15 if you ignore it · full reasoning on Signal

Opportunity — worth a try?

A short buyer-side instrument: the questions a company should be able to answer before letting an agent touch anything real, with Satya Nadella's own words as the spine and this month's incidents as the reason anyone reads it. Vendors are selling agent platforms and startups are selling agent monitoring. Nobody is selling the buyer the honest read on whether their organization can hold an agent accountable at all — which is the half you have spent thirty years on. It exercises the org diagnosis and the hands-on evaluation in the same artifact, which is the pairing almost no commentator can copy.

Cheapest way to test it: Turn Nadella's five controls into a one-page questionnaire. Then answer it yourself, from public documentation alone, against one shipped agent product — Anthropic's Managed Agents with the new agent-authored dynamic workflows is a good target, since it fans work across phases and the docs are public. Count how many of the five you can answer without talking to a salesperson. Publish the count. That is a weekend, it needs no client, and it is a dated original measurement rather than commentary.

No pressure — it quietly expires 2026-12-15 if you ignore it · full reasoning on Signal

Evidence refreshed — Agentic coding · 2026-11-04

Evidence refreshed. Opus 4.8 still looks right on what's in hand, but fast mode for it goes away Nov 30 and nobody has run SWE-bench Verified or Terminal-Bench on Opus 5.5 against it yet. That comparison needs to happen before the deadline — today's material has no new benchmark numbers to justify a promotion either way.

Evidence refreshed — Default assistant for daily work · 2026-11-04

Evidence refreshed. Same gap as the coding-agent pick: Opus 5.5 has been out since September but the daily-driver rationale was written in June and never weighed it. Pull fresh reasoning, speed, and price numbers before fast mode for 4.8 disappears Nov 30.

Now tracking: rsi-claims-credibility · 2026-11-04

Adopted as thesis rsi-claims-credibility. This is a genuinely different question from the already-tracked OpenAI AGI-timeline thesis: it's about whether real self-improvement capability is showing up across multiple labs — Anthropic's own statement plus independent reproduction of weight-editing on Qwen and Kimi, not just one company's model. Worth its own tracked claim with a concrete falsifier tied to independent reproduction, separate from GPT-6 Astra-specific claims.

Evidence refreshed — Hardest reasoning problems, cost no object · 2026-11-03

Evidence refreshed. GPT-5.2 no longer appears on the rebuilt Artificial Analysis Intelligence Index (v5, Oct 28), and the two new leaders (Opus 5.5, Gemini 4 Carbon) don't yet have entity files, so this can't be a one-tap promotion, and promotions are weekly-only in any case. The pick stands for now, but the Aug 18 evidence citation is stale and should be refreshed once contender entity files exist.

source flag no_change — artificial-analysis-intelligence-index · 2026-11-03

The methodology shift is real and well-documented (Oct 28 rebuild toward agentic-collaborator weighting, wider open/closed gap under v5), but deciding whether to re-source the reasoning axis to Epoch's Capabilities Index is a code-level config decision reserved for the owner, not something this pass can action. Flagging it as understood rather than resolved.

Opportunity archived: A first-hand, dated test of the book's substitution claim on the exact mechanics · 2026-11-02

Open for ten weeks with two left, and the condition it was timed against has now happened: Google shipped its Computer Use tool to general availability on October 30, so screen-driving agents are broadly available from all three major labs. A first-hand test now reads as a review of a crowded field rather than an early look. The underlying idea — testing the substitution claim on presales mechanics yourself — is still good; it just isn't dated intelligence any more, so it belongs in the manuscript plan rather than the opportunity queue.

Opportunity archived: A timely, well-sourced hook to publish a short reaction essay connecting this na · 2026-11-02

Its own expiry date is today, and the stronger version of the same essay is already open. The October 5 item carries Gartner's forecast that most vendor embedded-engineer programmes get abandoned by 2028, plus Airbnb's counter-example — better sourcing than a competing CEO's op-ed, and it runs to December 15. Keeping both would mean writing the same piece twice.

Pick kept — Fast local model for drafts and tooling · 2026-10-31

Qwen3.5 9B's circulating benchmarks and the offline coding-agent demo aren't yet compared head-to-head against qwen3-8b on coding/tool-calling at matched speed and memory footprint, so the current pick stands until a direct comparison is run at the next local-fast review.

Pick kept — Hardest reasoning problems, cost no object · 2026-10-31

Gemini 4 Carbon's #1 Artificial Analysis Intelligence Index score is real, but the model is still restricted to cyber-defenders-only access, so it can't actually be used for hard problems yet; GPT-5.2 stays the pick until Carbon gets broader API access to test.

method flag no_change — artificial-analysis-intelligence-index · 2026-10-31

The Intelligence Index is still the best available independent reasoning benchmark; the v4-to-v5 reweighting is a real comparability gap, but it's a citation-discipline issue (always note which version a cited score is from) rather than a reason to drop or replace the instrument.

Passed on benchmark: AA-Omniscience v1.1 · 2026-10-31

AA-Omniscience already feeds into the Artificial Analysis Intelligence Index, which we already track as a method; tracking it separately would double-count the same underlying signal without giving the verdicts any measurement they currently lack.

Passed on benchmark: HLE-Diamond · 2026-10-31

HLE-Diamond is a sub-benchmark inside the Artificial Analysis Intelligence Index's General pillar, which is already a tracked method; adding it separately would duplicate signal already captured through the composite index.