Archive — September 2026

17 Oct Can Tests Teach a Prompt to Improve? Automatic prompt optimization searches for instructions—and sometimes examples—that work better when executed. Its value depends on what the search changes, how it uses results, and what happens on cases kept outside the search. 10 cited · 96 read

14 Oct When is an agent's fast decision good enough? Additional reasoning is worth buying when a specific intervention can improve a consequential choice enough to justify its full cost. Confidence can help find those opportunities; it cannot price them by itself. 13 cited · 86 read

12 Oct How AI agents turn separate efforts into an accepted result See how useful AI teams connect responsibilities, evidence and acceptance—and judge their value by what one worker’s contribution lets another do and whether the resulting work survives the right checks. 16 cited · 148 read

09 Oct How an AI Agent Turns a Market Hypothesis into a Tested Signal The agent can propose and implement a predictor. What makes it a tested signal is the experiment it survives—and the history of how that experiment was used. 45 cited · 271 read

07 Oct How AI Agents Turn Financial Evidence into a Recommendation Following TradingAgents from data access to its final recommendation shows how reports, debate, and lessons from later outcomes shape a decision—and where an application must supply validation and execution. 31 cited · 245 read

05 Oct How memory compaction changes an agent’s next request A code-level walkthrough of checkpoints, retained messages, cleared results, and recovery paths in Pi, OpenCode, Deep Agents, Cline, and OpenHands. 40 cited · 332 read

04 Oct What an Agent Must Keep When It Compresses Its History A shorter history should leave the agent able to finish the task the user actually asked for. 12 cited · 194 read

02 Oct Training a Language Model to Judge New Questions A general-purpose decision model needs more than an open answer schema: its training must connect newly supplied questions and options to the evidence, and connect its probabilities to a defined target. 15 cited · 187 read

28 Sep What exists outside Jev—and how close has it come? From classical classifiers and entailment models to constrained LLMs and open decision scorers, the shared output shape hides very different methods and uneven evidence. 26 cited · 144 read

27 Sep Is Jev a better decision model—or a better decision service? Jev's decision interface has a real operational niche, but fair alternatives and independent tests show why its model-level advantage remains an open question. 17 cited · 200 read

20 Sep How an AI Agent Should Decide Which Decisions Deserve Credit or Blame A final score does not explain a trajectory. Reliable improvement reconstructs the decision, tests fair alternatives, measures uncertainty, and applies the smallest reversible repair. 14 cited · 216 read

20 Sep What a market view must contain before a financial agent builds a position A decision-ready market view is an accepted forecast packet: one resolvable claim, bound to the information available when it was made, passed through explicit stop rules, and handed to sizing without a hidden trade instruction. 13 cited · 201 read

19 Sep When an AI agent’s past becomes useful An agent learns from experience only when past outcomes change later choices through scoped, testable, revisable guidance that survives variation without hidden regressions or disproportionate cost. 9 cited · 156 read

13 Sep When an Agent Fails, Should You Change the Model or the System? The useful distinction is not weak intelligence versus bad engineering. It is which intervention removes the failures that matter—and whether the gain survives realistic costs and checks. 7 cited · 174 read

13 Sep When can an AI agent say the task is complete? Completion is a judgment about the requested result—not a synonym for stopping, passing a check, or sounding confident. 13 cited · 206 read

11 Sep How to make AI memory corrections and forgetting last A memory change should alter later answers and actions—not just the message that acknowledges it. 12 cited · 202 read

10 Sep When Should an AI Use Its Memory? Remembering is not permission to personalize. An AI should let the past shape an answer only when it helps the present task, still applies, and belongs in this interaction. 11 cited · 208 read

08 Sep What Should an AI System Remember? Persistent AI should preserve useful continuity without turning yesterday’s conversation into today’s rule. That requires deciding what to retain—and when not to use it. 6 cited · 161 read

07 Sep Large Model, Small Model, or Both? The choice is never about how hard the task looks. It is an estimate of how well a candidate model will do, a threshold you own, and a bill you have to measure yourself. 29 cited · 237 read

07 Sep Can a Harness Make a Small Model Match a Large One? A harness is a bigger lever than most engineers assume, and the weaker model is the more exposed to it. Whether a small model can substitute is decided by the task, not the scaffolding. 17 cited · 272 read

05 Sep When Sources Disagree: A Decision Model for Deep Research Deep Research should turn heterogeneous inputs into scoped claims, separate command authority from factual credibility, and preserve uncertainty through to the final answer. 16 cited · 277 read

02 Sep Where an LLM System’s Knowledge Lives—and How to Change It Safely An LLM does not have a single knowledge store. A safe change starts with a precise contract, puts mutable information in the least-entangled suitable location, and verifies both the intended effect and everything that must remain unchanged. 17 cited · 128 read

30 Aug Versioned LLM signals, separated trading authority A production reference architecture for turning an already-validated hybrid forecast into targets, independently risk-approved orders, venue events, and positions the firm can explain. 24 cited · 254 read

28 Aug From Language to Incremental Edge: Engineering LLM Semantics into Quantitative Models An LLM becomes useful to quantitative finance only when its semantic measurements improve a strong point-in-time numerical forecast, remain calibrated out of sample, and survive realistic trading costs. 16 cited · 252 read

26 Aug The LLM Edge in Finance Is a Division of Labor Language models can widen a financial system's field of view. That does not make them universal forecasters, portfolio managers, or sources of alpha. 23 cited · 278 read

25 Aug Can This Checkpoint Still Learn? Why retention and future learnability need separate tests in repeatedly trained neural networks. It frames a three-arm choice: continue, reset training state, or retrain from scratch. 10 cited · 319 read

24 Aug The Memory Was Right. The Decision Was Wrong. Long-running agents need more than accurate retrieval: they need a disciplined way to turn past events into current, scoped, independently supported inputs to action. 15 cited · 272 read

23 Aug How SPADE turns self-generated worlds into a bounded training loop A model can improve without human-written QA pairs or stronger-teacher demonstrations. SPADE shows where the missing supervision moves—and which controls must govern the result. 10 cited · 192 read

22 Aug How a Reasoner Learns When to Stop A full-parameter stop/refine policy learns from both repair and damage; sampling, search, and external verification remain isolated experiments around that control. 24 cited · 280 read

21 Aug How to Full-Train a Distilled Reasoning Model Without Making It Worse A verified teacher answer is an endpoint, not a training plan. The practical unit of reasoning transfer is a full run whose data, objective, student-generated states, and retention gates are all matched to the student. 27 cited · 233 read

20 Aug What Stops Working Before the Million-Token Window Runs Out The million-token window is a serving and pricing limit, not a measurement: literal lookup held in the one pilot to test a full million; nearly everything harder yet measured degrades far earlier. 28 cited · 429 read

20 Aug LoRA vs. Full Fine-Tuning in 2026: What Survived the Re-Measurement Two 2024 results anchored how practitioners chose between LoRA and full fine-tuning: "LoRA learns less and forgets less," and the warning that even matched benchmark scores hide structurally different solutions — an "illusion of equivalence." Between 2025 and 2026 both were re-measured, and the answer split by training regime. In supervised fine-tuning the canon survives, but its conditions have been rewritten in terms of adapter capacity, adapter placement, and learning rate. In reinforcement-learning post-training, LoRA now matches full fine-tuning at ranks as low as one — a result the canon never anticipated, resting so far on a lab blog and its reproductions rather than peer review. And three parts of the 2024 answer were never re-tested at all: the effective-rank mechanism offered to explain the gap, the canon's continued-pretraining protocol on current models, and any parity comparison on mixture-of-experts architectures. 27 cited · 366 read

20 Aug The Skill Is Real, the Ritual Is Not Yet Measured Curated skills measurably improve coding agents where they carry procedure the model lacks; the process disciplines the frameworks sell remain unmeasured — and the nearest tests lean the other way. 33 cited · 392 read

18 Aug Hard, Verifiable Agent Data: Download for Code, Build for the Rest Distillation-grade agent tasks are download-ready where checkers ship with them — code, and now terminal work; everywhere else the verifier is the scarce component, and the environment is yours to build. 51 cited · 642 read

Every report we have published