Science & Analytics
The Scientist
Microsoft's own table shows Decision-1 less calibrated than the model it beats.
Calibration comes in at 92.2 against the comparator's 93.1, and that gap matters more than the win if you are building a router or verifier, because those threshold on the probability, not the top answer. The cost case is cleaner: $0.042 per million input tokens and 85ms undercuts any LLM judge. If local deployment is a hard requirement, Liquid's d1-3B is the only model in the cohort that qualifies.
In Play
Context layers lift data agents
LinkedIn connected its analytics agent to a shared context layer. Acceptable answers rose from 25.9% to 79.5%, and median time-to-answer fell from 10.7 to 4.7 minutes, per LinkedIn's engineering write-up as summarized by TLDR Data. If you run a text-to-SQL or analytics agent, curated context is now the first thing to test. The write-up gives no test size, no rubric and no ablation of its two context sources.
Ask ClarityDecision models replace LLM judges
Microsoft-Decision-1 returns a probability for each fixed answer option instead of generating text. It charges $0.042 per million input tokens and reports 85ms median latency, per AI Breakfast. It joins Liquid d1-3B and SubQ Snap in a cohort built for routing, verification and judging. Microsoft's own table gives it a calibration score of 92.2, below the 93.1 of the comparator it beats on accuracy.
Ask ClarityFrontier upgrades cost more per point
Artificial Analysis scores Meta's Muse Wonder Full at 56, against 58 for Claude Opus 5.5, while Meta's own SWE-Marathon chart puts Full ahead 57.5 to 46.3, per AI Breakfast. The callable Muse Wonder scores 52 and costs $4.52 per benchmark task, about 2.8x the cost of Muse Spark 1.3 for 4 more points. Rank your candidates on cost per solved task on your own workload, not on leaderboard position. The Information notes that the top 10 models' share of OpenRouter usage fell from 55% to 14% between Q1 2025 and Q3 2026.
Ask ClarityAgent controls move outside the process
CSO reports that AWS agents were repeatedly tricked into revealing credentials, and that patching one attack path left others open. CyberScoop relays a report that North Korean actors wrote package documentation aimed at AI coding assistants. Any agent your team runs that reads package pages and holds credentials is inside this threat model. Microsoft's Execution Containers, shipped Oct 7, enforce limits from outside the model. AWS's Strands Box relies on OS isolation, which means more code has to be trusted.
Ask ClarityPower and optics ration compute
Bloomberg reports that electricity is now the biggest roadblock to AI data center construction, with Oracle trucking gas to sites. Lumentum says its optical components are sold out through early 2029. If your roadmap assumes reserved GPU capacity arrives on a fixed date, treat that date as less certain than it was last quarter. Microsoft claims Maia 300 delivers 40% more output per unit of power, but it names no baseline and the chip cannot be rented.
Ask Clarity
Deep Dives
- ●
A judgment call now costs a fraction of a cent. Are the probabilities any good?
Single-pass scorers easily beat generative models on latency, but the test that decides adoption is calibration against a small fine-tuned classifier.
The products in this cohort share a label and not much else. Microsoft-Decision-1 is a post-trained Qwen3.5-9B with a 32,768-token context, offered only as a hosted service on Foundry and OpenRouter. Liquid's d1-3B is the one open-weight option that runs…
2 action items
- ●
LinkedIn's 3x jump comes without the ablation that says which half to build
Generated profiles and human-reviewed notes fail in different ways, and the published result cannot say which one earned the gain. Your own golden set can.
Two context sources that fail differently The layer combines two inputs with opposite failure modes. LLM-written profiles come from schemas, query logs and code. They are cheap and go stale when schemas change. Human notes are reviewed in Git like…
2 action items
- ●
Coding agents now have attackers writing for them. Enforce controls outside the process
Prompt rules and per-exploit patches fail the same held-out test. The controls that count are the ones a hijacked agent cannot reach or rewrite.
Patching the exploit you saw is fitting to the test set CSO's account of AWS describes a sequence, not a single bug. Agents leaked credentials, one path was patched, and other paths stayed open. If you validate a mitigation only…
3 action items
The edition continues
Take the signal into the room.
Sign up or log in to read all 3 deep dives in full, plus the final take.
Read the full editionContinue with LinkedIn