Week of 2026-07-16

First weekly wiki after the desk came online. The underlying digest is a single backlog clear (280 curated items on 2026-07-16), not seven quiet days. Treat themes as a baseline map; week-over-week movement will be clearer once daily digests stack up.

Themes

Agent evaluation beyond one-shot scores

Tier-1 work clustered hard on whether agent gains survive new tasks and re-optimization. Continual Terminal-Bench-style protocols, deep-search stress tests, composable agent graphs, and verifier cascades all point the same way: static pass-rate screenshots are a weak teaching target. Prefer evaluation design (transfer, retention, regression control) over tool demos.

• Strong sources: Do Agent Optimizers Compound?, DeepStress, MyAG

LLM-as-judge stimulus integrity

Judge pipelines are proliferating faster than corpus hygiene. The standout mechanism paper is the test-oracle failure mode for synthetic negatives (shared decode budgets silently truncating "hallucinations"), alongside rubric generalization and audio-judge protocol shortcuts. Teaching cue: audit the stimuli before you trust the score.

• Strong sources: Test Oracle Problem in Synthetic LLM-as-Judge Corpora, Auditing Protocol-Level Shortcuts in Large Audio Language Model Judges

Coding agents and project-scale adoption

Research and practitioner notes are shifting from "can the agent write a function" to governance: agentic PRs across many repos, compiler feedback in the generation loop, and security framing that treats behavior compromise, not only resource compromise. Pair with forecast note on project-scale adoption.

• Strong sources: Early Adoption of Agentic Coding Tools by GitHub Projects, Generative Compilation

Post-training calibration and self-distillation

SFT/RL/OPD stages reshaping confidence, plus label-free self-distillation via consensus context. Useful as a durable critical-evaluation strand: when does a model's confidence track reality after post-training.

• Strong sources: Post-Training Shifts Confidence, Consensus as Privileged Context

Model releases and practitioner synthesis

Codex usage and GPT 5.6 family announcements dominated aggregator traffic. Durable teaching value is low (names churn); use as context for why frontier one-shots must stay thin, and why evaluation curricula outlast launch posts.

AI education tooling

Smaller cluster: high-order question generation and rapid upskilling frameworks. Relevant to this project's own pedagogy (quizzes/exercises), not yet a watchlist-level spike.

What moved

• First desk window, not a quiet week. One digest cleared 280 curated items; there is no prior wiki to compare.
• Forecast watchlist opened with three velocity-shaped bets: agent evaluation protocols (high), LLM-as-judge stimulus integrity (medium), agentic coding adoption at project scale (medium). See published/forecasts/2026-07-16.md.
• Curriculum already scaffolded on the two densest mechanism themes (judge corpus audit; optimizer compounding), each with lab specs drafted.

Teaching cues

• Durable: corpus integrity for LLM-as-judge evals; compounding/transfer metrics for agent optimizers; post-training calibration skepticism.
• Frontier one-shot: named model launches, specific coding-agent product workflows, trending repo spikes without a mechanism paper.
• Hold / watch: broad "Practitioner Notes" volume; multimodal workshop-style items without a clear competency map.

Digest trail

• published/digests/2026-07-16.md - 2026-07-16 → 2026-07-16 - 280 items
• Forecast companion (not a digest row): published/forecasts/2026-07-16.md