Morning Singularity Digest - 2026-09-21

Context

The assistant analyzes repository context, translates natural language requests into actionable command sequences, and incorporates planning techniques to ensure correctness and safety.

What's new

We present a systematic evaluation methodology using synthetic and randomized git environments, comparing the performance of LLM-only and planning-augmented variants across multiple metrics.

Key details

• Recent advances in Large Language Models (LLMs) offer promising capabilities for interpreting developer intent, but their effectiveness in repository management tasks is limited by the need for formal reasoning.
• This work introduces Git-Assistant, an AI-based assistant that combines LLMs with automated planning to support developers in executing non-trivial git operations.
• The assistant analyzes repository context, translates natural language requests into actionable command sequences, and incorporates planning techniques to ensure correctness and safety.
• We present a systematic evaluation methodology using synthetic and randomized git environments, comparing the performance of LLM-only and planning-augmented variants across multiple metrics.

Results & evidence

• arXiv:2607.09224v3 Announce Type: replace-cross Abstract: Version control systems are essential for collaborative software development, yet tools like git remain challenging for many practitioners.
• Computer Science > Software Engineering [Submitted on 10 Jul 2026 (v1), last revised 18 Sep 2026 (this version, v3)] Title:Git-Assistant: Planning-Based Support for Updating Git Repositories View PDF HTML (experimental) Abstract:Version control systems are...
• Submission history From: Alfredo Garrachón Ruiz [view email] [v1] Fri, 10 Jul 2026 09:16:20 UTC (277 KB) [v2] Tue, 14 Jul 2026 10:25:32 UTC (1 KB) (withdrawn) [v3] Fri, 18 Sep 2026 13:09:53 UTC (277 KB) Current browse context: cs.SE References & Citations L...

Limitations / unknowns

• Recent advances in Large Language Models (LLMs) offer promising capabilities for interpreting developer intent, but their effectiveness in repository management tasks is limited by the need for formal reasoning.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

Current browse context: cs.CL References & Citations Loading...

What's new

Although recent approaches have begun to incorporate multiple prior examinations, they usually aggregate a fixed-length history without explicitly modeling the role-dependent relevance of each prior examination before fusion.

Key details

• Although recent approaches have begun to incorporate multiple prior examinations, they usually aggregate a fixed-length history without explicitly modeling the role-dependent relevance of each prior examination before fusion.
• To address this, we propose TALON, a Temporally Aware LONgitudinal RRG framework that adaptively integrates variable-length patient histories.
• The underlying Dual-Channel Temporal Fusion Module (DCTFM) compares the current examination with each prior examination through complementary similarity and change channels to capture persistent findings and interval changes, respectively.
• The specially designed channel-specific attention estimates the relevance of each prior examination, while a learned prior-specific gate adaptively integrates informative longitudinal evidence and suppresses redundancy.

Results & evidence

• arXiv:2609.20826v1 Announce Type: cross Abstract: Current radiology report generation (RRG) models usually produce descriptive reports based on a single examination or only the most recent prior examination, limiting their ability to perform accurate and me...
• Computer Science > Computation and Language [Submitted on 22 Jul 2026] Title:TALON: A Temporally Aware Longitudinal Framework for Radiology Report Generation View PDF HTML (experimental) Abstract:Current radiology report generation (RRG) models usually prod...

Limitations / unknowns

• arXiv:2609.20826v1 Announce Type: cross Abstract: Current radiology report generation (RRG) models usually produce descriptive reports based on a single examination or only the most recent prior examination, limiting their ability to perform accurate and me...

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

Scale computer-use 2.0 with open-source drivers, cross-OS fleets, and benchmarks for training, evaluation, and data generation.

What's new

Scale computer-use 2.0 with open-source drivers, cross-OS fleets, and benchmarks for training, evaluation, and data generation.

Key details

• Give AI agents computers they can use.
• Cua provides open-source desktop automation, isolated cloud desktops, local macOS VMs, specialist decision models, and benchmarks for evaluating computer-use agents.
• - Cua Fleets: Provision a Linux desktop, run a command, and save a screenshot.
• - CUA-S1: Explore small, specialized models for computer-use decisions.

Results & evidence

• Scale computer-use 2.0 with open-source drivers, cross-OS fleets, and benchmarks for training, evaluation, and data generation.
• Computer-Use 2.0 describes an agent moving between code, APIs, and graphical interfaces within the same task.
• Watch the 50-second demo, then explore Omarchy on Fleet.

Limitations / unknowns

• Generalization outside curated tasks is still unclear.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

Secure environments for developers and their agents Coder is a self-hosted platform for cloud development environments and AI coding agents.

What's new

New integrations are always in progress.

Key details

• Workspaces are defined with Terraform, connected through a secure Wireguard® tunnel, and automatically shut down when not used.
• Coder Agents runs a native AI coding agent whose loop executes in the control plane on your infrastructure, with no API keys in workspaces.
• - Define cloud development environments in Terraform - EC2 VMs, Kubernetes Pods, Docker Containers, etc.
• - Automatically shutdown idle resources to save on costs - Onboard developers in seconds instead of days - Delegate coding work to AI agents on your infrastructure - Bring any model (Anthropic, OpenAI, Google, Bedrock, self-hosted) - No LLM credentials in w...

Results & evidence

• Try Coder with the install script on Linux and macOS, or grab the latest binary or installer from GitHub Releases on Windows: curl -L https://coder.com/install.sh | sh Start the server and open http://localhost:3000 to create your initial user, create a Doc...

Limitations / unknowns

• Generalization outside curated tasks is still unclear.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

Poker legality, hole-card masking and public context are generated by the PokerTools engine, not by prompts.

What's new

A browser-first AI poker benchmark.

Key details

• Seat Jev and OpenAI-compatible models at the same no-limit Texas Hold'em table, watch every card and decision as a spectator, and let the tournament run autonomously until one model wins.
• | Live app | https://pokertools-arena.github.io/ | | Source | https://github.com/pokertools-arena/pokertools-arena.github.io | | Package | npx pokertools-arena | | Runtime | Node.js ≥ 24 for tooling; any modern browser for the app | Current release: 0.18.1.
• The npm package now includes the shared .env parser required by its launcher, with a packed-artifact smoke test preventing future npx pokertools-arena module-resolution failures.
• video.mp4 Text-model benchmarks are usually static question sets.

Results & evidence

• | Live app | https://pokertools-arena.github.io/ | | Source | https://github.com/pokertools-arena/pokertools-arena.github.io | | Package | npx pokertools-arena | | Runtime | Node.js ≥ 24 for tooling; any modern browser for the app | Current release: 0.18.1.
• | Area | What you get | |---|---| | Table | 2–10 seats, no-limit Hold'em tournaments, rising blinds, antes, time banks, elimination and podium flow | | Connections | Generic OpenAI-compatible base URL, OpenRouter preset, TypeSafe System One preset | | Proto...

Limitations / unknowns

• Seat Jev and OpenAI-compatible models at the same no-limit Texas Hold'em table, watch every card and decision as a spectator, and let the tournament run autonomously until one model wins.
• The npm package now includes the shared .env parser required by its launcher, with a packed-artifact smoke test preventing future npx pokertools-arena module-resolution failures.
• | Area | What you get | |---|---| | Table | 2–10 seats, no-limit Hold'em tournaments, rising blinds, antes, time banks, elimination and podium flow | | Connections | Generic OpenAI-compatible base URL, OpenRouter preset, TypeSafe System One preset | | Proto...

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

The assistant analyzes repository context, translates natural language requests into actionable command sequences, and incorporates planning techniques to ensure correctness and safety.

What's new

We present a systematic evaluation methodology using synthetic and randomized git environments, comparing the performance of LLM-only and planning-augmented variants across multiple metrics.

Key details

• Recent advances in Large Language Models (LLMs) offer promising capabilities for interpreting developer intent, but their effectiveness in repository management tasks is limited by the need for formal reasoning.
• This work introduces Git-Assistant, an AI-based assistant that combines LLMs with automated planning to support developers in executing non-trivial git operations.
• The assistant analyzes repository context, translates natural language requests into actionable command sequences, and incorporates planning techniques to ensure correctness and safety.
• We present a systematic evaluation methodology using synthetic and randomized git environments, comparing the performance of LLM-only and planning-augmented variants across multiple metrics.

Results & evidence

• arXiv:2607.09224v3 Announce Type: replace-cross Abstract: Version control systems are essential for collaborative software development, yet tools like git remain challenging for many practitioners.
• Computer Science > Software Engineering [Submitted on 10 Jul 2026 (v1), last revised 18 Sep 2026 (this version, v3)] Title:Git-Assistant: Planning-Based Support for Updating Git Repositories View PDF HTML (experimental) Abstract:Version control systems are...
• Submission history From: Alfredo Garrachón Ruiz [view email] [v1] Fri, 10 Jul 2026 09:16:20 UTC (277 KB) [v2] Tue, 14 Jul 2026 10:25:32 UTC (1 KB) (withdrawn) [v3] Fri, 18 Sep 2026 13:09:53 UTC (277 KB) Current browse context: cs.SE References & Citations L...

Limitations / unknowns

• Recent advances in Large Language Models (LLMs) offer promising capabilities for interpreting developer intent, but their effectiveness in repository management tasks is limited by the need for formal reasoning.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

Scale computer-use 2.0 with open-source drivers, cross-OS fleets, and benchmarks for training, evaluation, and data generation.

What's new

Scale computer-use 2.0 with open-source drivers, cross-OS fleets, and benchmarks for training, evaluation, and data generation.

Key details

• Give AI agents computers they can use.
• Cua provides open-source desktop automation, isolated cloud desktops, local macOS VMs, specialist decision models, and benchmarks for evaluating computer-use agents.
• - Cua Fleets: Provision a Linux desktop, run a command, and save a screenshot.
• - CUA-S1: Explore small, specialized models for computer-use decisions.

Results & evidence

• Scale computer-use 2.0 with open-source drivers, cross-OS fleets, and benchmarks for training, evaluation, and data generation.
• Computer-Use 2.0 describes an agent moving between code, APIs, and graphical interfaces within the same task.
• Watch the 50-second demo, then explore Omarchy on Fleet.

Limitations / unknowns

• Generalization outside curated tasks is still unclear.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

Poker legality, hole-card masking and public context are generated by the PokerTools engine, not by prompts.

What's new

A browser-first AI poker benchmark.

Key details

• Seat Jev and OpenAI-compatible models at the same no-limit Texas Hold'em table, watch every card and decision as a spectator, and let the tournament run autonomously until one model wins.
• | Live app | https://pokertools-arena.github.io/ | | Source | https://github.com/pokertools-arena/pokertools-arena.github.io | | Package | npx pokertools-arena | | Runtime | Node.js ≥ 24 for tooling; any modern browser for the app | Current release: 0.18.1.
• The npm package now includes the shared .env parser required by its launcher, with a packed-artifact smoke test preventing future npx pokertools-arena module-resolution failures.
• video.mp4 Text-model benchmarks are usually static question sets.

Results & evidence

• | Live app | https://pokertools-arena.github.io/ | | Source | https://github.com/pokertools-arena/pokertools-arena.github.io | | Package | npx pokertools-arena | | Runtime | Node.js ≥ 24 for tooling; any modern browser for the app | Current release: 0.18.1.
• | Area | What you get | |---|---| | Table | 2–10 seats, no-limit Hold'em tournaments, rising blinds, antes, time banks, elimination and podium flow | | Connections | Generic OpenAI-compatible base URL, OpenRouter preset, TypeSafe System One preset | | Proto...

Limitations / unknowns

• Seat Jev and OpenAI-compatible models at the same no-limit Texas Hold'em table, watch every card and decision as a spectator, and let the tournament run autonomously until one model wins.
• The npm package now includes the shared .env parser required by its launcher, with a packed-artifact smoke test preventing future npx pokertools-arena module-resolution failures.
• | Area | What you get | |---|---| | Table | 2–10 seats, no-limit Hold'em tournaments, rising blinds, antes, time banks, elimination and podium flow | | Connections | Generic OpenAI-compatible base URL, OpenRouter preset, TypeSafe System One preset | | Proto...

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

The assistant analyzes repository context, translates natural language requests into actionable command sequences, and incorporates planning techniques to ensure correctness and safety.

What's new

We present a systematic evaluation methodology using synthetic and randomized git environments, comparing the performance of LLM-only and planning-augmented variants across multiple metrics.

Key details

• Recent advances in Large Language Models (LLMs) offer promising capabilities for interpreting developer intent, but their effectiveness in repository management tasks is limited by the need for formal reasoning.
• This work introduces Git-Assistant, an AI-based assistant that combines LLMs with automated planning to support developers in executing non-trivial git operations.
• The assistant analyzes repository context, translates natural language requests into actionable command sequences, and incorporates planning techniques to ensure correctness and safety.
• We present a systematic evaluation methodology using synthetic and randomized git environments, comparing the performance of LLM-only and planning-augmented variants across multiple metrics.

Results & evidence

• arXiv:2607.09224v3 Announce Type: replace-cross Abstract: Version control systems are essential for collaborative software development, yet tools like git remain challenging for many practitioners.
• Computer Science > Software Engineering [Submitted on 10 Jul 2026 (v1), last revised 18 Sep 2026 (this version, v3)] Title:Git-Assistant: Planning-Based Support for Updating Git Repositories View PDF HTML (experimental) Abstract:Version control systems are...
• Submission history From: Alfredo Garrachón Ruiz [view email] [v1] Fri, 10 Jul 2026 09:16:20 UTC (277 KB) [v2] Tue, 14 Jul 2026 10:25:32 UTC (1 KB) (withdrawn) [v3] Fri, 18 Sep 2026 13:09:53 UTC (277 KB) Current browse context: cs.SE References & Citations L...

Limitations / unknowns

• Recent advances in Large Language Models (LLMs) offer promising capabilities for interpreting developer intent, but their effectiveness in repository management tasks is limited by the need for formal reasoning.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

Current browse context: cs.CL References & Citations Loading...

What's new

Although recent approaches have begun to incorporate multiple prior examinations, they usually aggregate a fixed-length history without explicitly modeling the role-dependent relevance of each prior examination before fusion.

Key details

• Although recent approaches have begun to incorporate multiple prior examinations, they usually aggregate a fixed-length history without explicitly modeling the role-dependent relevance of each prior examination before fusion.
• To address this, we propose TALON, a Temporally Aware LONgitudinal RRG framework that adaptively integrates variable-length patient histories.
• The underlying Dual-Channel Temporal Fusion Module (DCTFM) compares the current examination with each prior examination through complementary similarity and change channels to capture persistent findings and interval changes, respectively.
• The specially designed channel-specific attention estimates the relevance of each prior examination, while a learned prior-specific gate adaptively integrates informative longitudinal evidence and suppresses redundancy.

Results & evidence

• arXiv:2609.20826v1 Announce Type: cross Abstract: Current radiology report generation (RRG) models usually produce descriptive reports based on a single examination or only the most recent prior examination, limiting their ability to perform accurate and me...
• Computer Science > Computation and Language [Submitted on 22 Jul 2026] Title:TALON: A Temporally Aware Longitudinal Framework for Radiology Report Generation View PDF HTML (experimental) Abstract:Current radiology report generation (RRG) models usually prod...

Limitations / unknowns

• arXiv:2609.20826v1 Announce Type: cross Abstract: Current radiology report generation (RRG) models usually produce descriptive reports based on a single examination or only the most recent prior examination, limiting their ability to perform accurate and me...

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

arXiv:2609.18310v3 Announce Type: replace Abstract: We introduce Nemotron-SEA-LION-v4.8, a family of Southeast Asian Languages In One Network (SEA-LION) models built upon NVIDIA Nemotron 3.

What's new

arXiv:2609.18310v3 Announce Type: replace Abstract: We introduce Nemotron-SEA-LION-v4.8, a family of Southeast Asian Languages In One Network (SEA-LION) models built upon NVIDIA Nemotron 3.

Key details

• The family includes 30B-A3B and 120B-A12B models, with both continued-pretrained base checkpoints and post-trained variants.
• We adapt the models using Southeast Asian, reasoning, code, and multilingual parallel data, followed by post-training with supervised fine-tuning and online on-policy distillation.
• On SEA-HELM, the 30B-A3B model improves the overall SEA score from 46.06 to 51.57, while the 120B-A12B model improves from 49.30 to 63.44.
• Across seven Southeast Asian languages, we observe broad capability gains with the 120B-A12B model showing broader and more consistent improvements across tasks.

Results & evidence

• arXiv:2609.18310v3 Announce Type: replace Abstract: We introduce Nemotron-SEA-LION-v4.8, a family of Southeast Asian Languages In One Network (SEA-LION) models built upon NVIDIA Nemotron 3.
• On SEA-HELM, the 30B-A3B model improves the overall SEA score from 46.06 to 51.57, while the 120B-A12B model improves from 49.30 to 63.44.
• Computer Science > Computation and Language [Submitted on 16 Sep 2026 (v1), last revised 18 Sep 2026 (this version, v3)] Title:SEA-LION-v4.8: A Technical Report View PDF HTML (experimental) Abstract:We introduce Nemotron-SEA-LION-v4.8, a family of Southeast...

Limitations / unknowns

• Generalization outside curated tasks is still unclear.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

arXiv:2604.23478v3 Announce Type: replace Abstract: Large language models are widely used to judge the output of other language models, yet whether a judge returns the same verdict when the same request is worded differently remains largely unexamined.

What's new

arXiv:2604.23478v3 Announce Type: replace Abstract: Large language models are widely used to judge the output of other language models, yet whether a judge returns the same verdict when the same request is worded differently remains largely unexamined.

Key details

• We study that question across four evaluation tasks and twenty-five judges from six providers.
• To support the analysis we release JudgeSense, a benchmark of 880 items from human-labelled corpora, each issued under two instructions that differ in wording and not in what they ask, with the complete decision logs.
• Every score is reported against the judge's own agreement with itself on the identical prompt, so decoding noise is not charged to wording, and the release lets a reader ask the same of any judge not in our roster.
• Rewording costs agreement on all four tasks, and on two it clears the threshold we declare for a practically meaningful effect; the ordinal task is both the least stable and the one fewest judges are accurate on, and within a single family parameter count d...

Results & evidence

• arXiv:2604.23478v3 Announce Type: replace Abstract: Large language models are widely used to judge the output of other language models, yet whether a judge returns the same verdict when the same request is worded differently remains largely unexamined.
• To support the analysis we release JudgeSense, a benchmark of 880 items from human-labelled corpora, each issued under two instructions that differ in wording and not in what they ask, with the complete decision logs.
• Computer Science > Computation and Language [Submitted on 26 Apr 2026 (v1), last revised 18 Sep 2026 (this version, v3)] Title:JudgeSense: A Benchmark for Prompt Sensitivity in LLM-as-a-Judge Systems View PDF HTML (experimental) Abstract:Large language mode...

Limitations / unknowns

• Generalization outside curated tasks is still unclear.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

Solution for long term memory for agent coding CLIs and to facilitate handoff between different agent vendors Long-term memory for AI coding agents.

What's new

Quit Claude Code mid-task, start OpenAI Codex in the same directory, continue without re-explaining the architecture, the failed approaches, or the open questions.

Key details

• Quit Claude Code mid-task, start OpenAI Codex in the same directory, continue without re-explaining the architecture, the failed approaches, or the open questions.
• Your coding agent already has a memory feature.
• Claude Code takes its own notes, Cursor remembers some things, and every platform is adding more.
• All of them share the same walls: the notes live on one machine, belong to one agent, and vanish from view the moment you switch tools — or teammates.

Results & evidence

• No hard numbers surfaced in the source text; treat claims as directional until benchmarks appear.

Limitations / unknowns

• Generalization outside curated tasks is still unclear.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

Their environment provides context, tools, files, tests, and previews that make their capabilities and results visible.

What's new

A framework for building agentic apps Agent-Native is an open-source TypeScript framework for building agents that pair autonomous work with a purpose-built UI.

Key details

• Define each capability once as an action: the agent uses it as a tool, and the UI calls it from code.
• npx --yes @agent-native/core@latest create my-agent --standalone --template chat Follow the getting started guide for a full intro to the framework.
• Coding agents work with more than a text box.
• Their environment provides context, tools, files, tests, and previews that make their capabilities and results visible.

Results & evidence

• No hard numbers surfaced in the source text; treat claims as directional until benchmarks appear.

Limitations / unknowns

• Generalization outside curated tasks is still unclear.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

Hi everyone.

TypeSafe released Jev last week, and it pushed me to build agent-chaperone.

What's new

You read your own log first and decide what you would have wanted stopped.

Key details

• It is a screening layer for AI agent tool calls and tool results.

  It screens risky, off-task, or secret-leaking actions before they run, and prompt injections or other suspicious instructions before the agent reads them.

  Repo:

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

WWII Tank, Aircraft, and Ship Identification Guides

What's new

WWII Tank, Aircraft, and Ship Identification Guides

Key details

• WWII Tank, Aircraft, and Ship Identification Guides

Results & evidence

• No hard numbers surfaced in the source text; treat claims as directional until benchmarks appear.

Limitations / unknowns

• Generalization outside curated tasks is still unclear.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

It was trained with a longer reinforcement learning run on a harder mix of tasks, weighted toward problems that take many hours to complete.

What's new

Grok 4.7 uses a new, larger base model compared to Grok 4.6.

Key details

• Twice as fast, at half the price of comparable models.
• Grok 4.7 is our most capable model for coding and knowledge work.
• It works longer on difficult tasks, checks its own work more carefully, and comes with our best-calibrated safeguards to date.
• Served at the same price and speed as Grok 4.6, it is highly competitive in its class.

Results & evidence

• Grok 4.7 is our most capable model for coding and knowledge work.
• Served at the same price and speed as Grok 4.6, it is highly competitive in its class.
• On CursorBench 4.0, which stresses longer-running coding tasks, Grok 4.7 is at the frontier in price-performance.

Limitations / unknowns

• Generalization outside curated tasks is still unclear.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.