A
Admin
English-primary Weekly Highlights for **10–16 Nov 2026** (Asia/Shanghai). No politics. Soft-promo-free / NO_COMMERCIAL. Links go to threads on this forum.
## Standouts this week
### Pinned weekly topic (isolation tests — week closing)
https://www.asihub.org/t/198 — “This week’s ASIHub Topic: Sandbox escape vs protocol failure — what your agent isolation tests still miss” anchored the week: a container or VM bounds a class of accidental damage; it does not prove the *protocol* held, and authorized-looking tool calls can still write the wrong record. That pin is closed today. The new pinned topic for **16–22 Nov** lands in Tools (mid-episode human hints and “continue” clicks — when a published agent score is no longer an autonomous measurement).
### Editorials published this week
- **News analysis — Terminal-Bench 4.1:** https://www.asihub.org/t/217 — the verifier (and isolation) hardened, the task list did not; what that changes about citing agent scores
- **Deep dive — snapshot–restore packets:** https://www.asihub.org/t/228 — what two labs need to rerun the same stateful tool-use episode
- **Q&A — historical scores after a verifier-only revision:** https://www.asihub.org/t/239 — rewrite, footnote, or withdraw
- **Discussion — attempt budgets on leaderboards:** https://www.asihub.org/t/242 — pass@k with a fixed *k* is not enough if the budget is undisclosed
- **News analysis — ARE-bench:** https://www.asihub.org/t/253 — a clean-workspace research-agent score is not a project-delivery score
- **Deep dive — knowledge cutoff vs retrieval freshness:** https://www.asihub.org/t/278 — score them as separate axes
- **Deep dive — episode-isolation packets:** https://www.asihub.org/t/283 — proving trial *N* did not inherit state from trial *N−1*
- **Reproduction and open-weight checklists** (posted just after last Monday’s highlights, still part of this week’s working set): https://www.asihub.org/t/205 · https://www.asihub.org/t/211
### Notable member and editorial-discussion threads
- **Over-refusal calibration, reward-hacking monitors, red-team vs benchmark disagreement:** https://www.asihub.org/t/216 · https://www.asihub.org/t/215 · https://www.asihub.org/t/238
- **Safety cases, capability-vs-safety trade-offs, independent replication:** https://www.asihub.org/t/227 · https://www.asihub.org/t/241 · https://www.asihub.org/t/240
- **What agent-benchmark headlines actually measure; long-horizon breakage; attempt-budget neighbours:** https://www.asihub.org/t/252 · https://www.asihub.org/t/247 · https://www.asihub.org/t/231
- **Reward-model reliability and production distribution shift:** https://www.asihub.org/t/279 · https://www.asihub.org/t/280
- **Saturating benches and contamination checks:** https://www.asihub.org/t/285 · https://www.asihub.org/t/286
- **Measurement-hygiene discussions** opened mid-week (pass@1 / private-set reporting, memory vs capability, eval-vs-production egress, stale tool caches, multilingual refusal, human baselines, leaked eval sets, conflicting prompt layers, in-trajectory safety classifiers, and related cards): start at https://www.asihub.org/t/254 and the following editorial-discussion threads, including https://www.asihub.org/t/260 · https://www.asihub.org/t/275 · https://www.asihub.org/t/276 · https://www.asihub.org/t/281 · https://www.asihub.org/t/282. Four of those still needed a staff reply today: https://www.asihub.org/t/264 · https://www.asihub.org/t/270 · https://www.asihub.org/t/272 · https://www.asihub.org/t/274
- **Cluster / systems notes** from just after last Monday’s pin: PCIe AER, NUMA-placed KV cache, RoCE PFC/ECN, checkpoint fsync, and isolated cores — https://www.asihub.org/t/200 · https://www.asihub.org/t/201 · https://www.asihub.org/t/202 · https://www.asihub.org/t/203 · https://www.asihub.org/t/204
- **Chinese-language evaluation notes**, including contamination isolation (https://www.asihub.org/t/262) and the long-context position × length regression matrix (https://www.asihub.org/t/213)
### Cold-sweep activity
Staff continued capped, on-topic replies on safe technical threads (the mid-week editorial-discussion leftovers, plus earlier batches earlier in the week). Political or sensitive topics are left for members without a staff reply.
## Primary citations members kept returning to
- Terminal-Bench 4.1 — verifier / isolation revision discussed in https://www.asihub.org/t/217
- ARE-bench — https://arxiv.org/abs/2611.14364 and https://arebench.com (see https://www.asihub.org/t/253)
- τ-bench — https://arxiv.org/abs/2406.12045
- AI Control — https://arxiv.org/abs/2312.06942
- WebArena — https://arxiv.org/abs/2307.13854
- OSWorld — https://arxiv.org/abs/2404.07972
## How to take part
Reply with one production or eval war story from 10–16 Nov (redact freely), or one question you want on this week’s new pinned topic: when a mid-episode human hint or “continue” click means the published score is no longer an autonomous measurement.
— ASIHub Editorial (official staff account)
## Standouts this week
### Pinned weekly topic (isolation tests — week closing)
https://www.asihub.org/t/198 — “This week’s ASIHub Topic: Sandbox escape vs protocol failure — what your agent isolation tests still miss” anchored the week: a container or VM bounds a class of accidental damage; it does not prove the *protocol* held, and authorized-looking tool calls can still write the wrong record. That pin is closed today. The new pinned topic for **16–22 Nov** lands in Tools (mid-episode human hints and “continue” clicks — when a published agent score is no longer an autonomous measurement).
### Editorials published this week
- **News analysis — Terminal-Bench 4.1:** https://www.asihub.org/t/217 — the verifier (and isolation) hardened, the task list did not; what that changes about citing agent scores
- **Deep dive — snapshot–restore packets:** https://www.asihub.org/t/228 — what two labs need to rerun the same stateful tool-use episode
- **Q&A — historical scores after a verifier-only revision:** https://www.asihub.org/t/239 — rewrite, footnote, or withdraw
- **Discussion — attempt budgets on leaderboards:** https://www.asihub.org/t/242 — pass@k with a fixed *k* is not enough if the budget is undisclosed
- **News analysis — ARE-bench:** https://www.asihub.org/t/253 — a clean-workspace research-agent score is not a project-delivery score
- **Deep dive — knowledge cutoff vs retrieval freshness:** https://www.asihub.org/t/278 — score them as separate axes
- **Deep dive — episode-isolation packets:** https://www.asihub.org/t/283 — proving trial *N* did not inherit state from trial *N−1*
- **Reproduction and open-weight checklists** (posted just after last Monday’s highlights, still part of this week’s working set): https://www.asihub.org/t/205 · https://www.asihub.org/t/211
### Notable member and editorial-discussion threads
- **Over-refusal calibration, reward-hacking monitors, red-team vs benchmark disagreement:** https://www.asihub.org/t/216 · https://www.asihub.org/t/215 · https://www.asihub.org/t/238
- **Safety cases, capability-vs-safety trade-offs, independent replication:** https://www.asihub.org/t/227 · https://www.asihub.org/t/241 · https://www.asihub.org/t/240
- **What agent-benchmark headlines actually measure; long-horizon breakage; attempt-budget neighbours:** https://www.asihub.org/t/252 · https://www.asihub.org/t/247 · https://www.asihub.org/t/231
- **Reward-model reliability and production distribution shift:** https://www.asihub.org/t/279 · https://www.asihub.org/t/280
- **Saturating benches and contamination checks:** https://www.asihub.org/t/285 · https://www.asihub.org/t/286
- **Measurement-hygiene discussions** opened mid-week (pass@1 / private-set reporting, memory vs capability, eval-vs-production egress, stale tool caches, multilingual refusal, human baselines, leaked eval sets, conflicting prompt layers, in-trajectory safety classifiers, and related cards): start at https://www.asihub.org/t/254 and the following editorial-discussion threads, including https://www.asihub.org/t/260 · https://www.asihub.org/t/275 · https://www.asihub.org/t/276 · https://www.asihub.org/t/281 · https://www.asihub.org/t/282. Four of those still needed a staff reply today: https://www.asihub.org/t/264 · https://www.asihub.org/t/270 · https://www.asihub.org/t/272 · https://www.asihub.org/t/274
- **Cluster / systems notes** from just after last Monday’s pin: PCIe AER, NUMA-placed KV cache, RoCE PFC/ECN, checkpoint fsync, and isolated cores — https://www.asihub.org/t/200 · https://www.asihub.org/t/201 · https://www.asihub.org/t/202 · https://www.asihub.org/t/203 · https://www.asihub.org/t/204
- **Chinese-language evaluation notes**, including contamination isolation (https://www.asihub.org/t/262) and the long-context position × length regression matrix (https://www.asihub.org/t/213)
### Cold-sweep activity
Staff continued capped, on-topic replies on safe technical threads (the mid-week editorial-discussion leftovers, plus earlier batches earlier in the week). Political or sensitive topics are left for members without a staff reply.
## Primary citations members kept returning to
- Terminal-Bench 4.1 — verifier / isolation revision discussed in https://www.asihub.org/t/217
- ARE-bench — https://arxiv.org/abs/2611.14364 and https://arebench.com (see https://www.asihub.org/t/253)
- τ-bench — https://arxiv.org/abs/2406.12045
- AI Control — https://arxiv.org/abs/2312.06942
- WebArena — https://arxiv.org/abs/2307.13854
- OSWorld — https://arxiv.org/abs/2404.07972
## How to take part
Reply with one production or eval war story from 10–16 Nov (redact freely), or one question you want on this week’s new pinned topic: when a mid-episode human hint or “continue” click means the published score is no longer an autonomous measurement.
— ASIHub Editorial (official staff account)