No. 1 ·

The Oct 29 edition

Window
2026-10-28 07:30 → 2026-10-29 07:30 (America/Chicago, CDT). Older items that you can still act on are kept and labeled.
Method
discovery from aggregators, confirmation from official sources, rubric score >=5, no read-time cap
Rule
every topic, story, and mention is a clickable markdown link
  • 199 primary links
  • 9 sections
  • 20 min read
  • Still open7
  • News26
  • Update6
  • Context5

Labels: News = first seen in the window · Update = older story with a new number or artifact · Context = background you need to act · Still open = not new, but still worth trying

Snapshot#

Frontier and open-weight models#

Gemini 4 Carbon — News (2026-10-28)#

  • Who / what: Google DeepMind's new frontier model, released two weeks after Gemini 4 Argon. Google pitches it on long-horizon agentic work and cybersecurity. It reports 88.2% on SEC-bench Pro, 71.0% on CWE-bench, and +89 Elo to #4 (1768) on WebDev Arena.
  • What's new: According to Artificial Analysis:
    • 1M context.
    • Text, image, video, and speech input; text output.
    • #1 on Intelligence Index v5 at 59.
    • Coding Agent Index: 59 with Antigravity CLI, vs 62 for Claude Code + Opus 5.5 (max).
    • Terminal-Bench 4.0: 64%. FrontierCode v1.1: 51%. Both rank second.
    • AA-Omniscience v1.1: 54, but the hallucination rate rose from 4% (Argon) to 11%.
  • Cost: AA charts use standard pricing: $4/$20 per 1M tokens, cached input $0.20 (95% off). That works out to $7.76 per Intelligence Index task with 123k output tokens per task, vs $10.04 / 173k for Opus 5.5 (max) and $6.25 for GPT-6 Astra (max). Google's launch post gives an introductory price of $2/$10, the same as Argon.
  • License / serving: Proprietary. Access is limited to "a subset of cyber defenders" through the Fairwind program, and there is no general developer API yet.
  • Builder take: You can't build on it this week. The useful signal is that Google's coding-agent score now sits within 3 points of Claude Code. Plan an eval slot for when the API opens, and test Argon on OpenRouter in the meantime ($2/$10 listed).
  • Operator chatter: HN thread (957 points, 765 comments when fetched).

Particle-1 — News (2026-10-28)#

  • What: A coding model post-trained on DeepSeek V4.1 Flash, aimed at terminal workflows, codebase edits, and data engineering. It is "trained to complete them with substantially fewer reasoning and tool-call tokens".
  • Cost / speed: $0.25 in / $1.00 out / $0.03 cache read per 1M tokens on OpenRouter. The page shows 133 tok/s and 1.43 s latency (P50) from a single provider, with 1M context.
  • License: Weights are not linked from the listing; treat it as API-only.
  • Builder take: A cheap candidate for agent loops. There is no independent benchmark yet, so compare it head-to-head against base V4.1 Flash on your own tasks.

Nemotron 3.5 Transcribe — News (blog 2026-10-28)#

  • What: An open-weight 30B-A4B speech model (HF card license: openmdw-1.1). The available checkpoint can be served with vLLM today. It transcribes 40 languages, translates 31 of them to or from English, labels up to 8 speakers, and handles recordings up to an hour. It ships with the open Granary v2 data pipeline and NeMo fine-tuning recipes.
  • Rankings: NVIDIA reports first among open-weight models on the HF Open ASR Leaderboard (English). Quantized builds are "coming soon".
  • Builder take: A single-checkpoint ASR + diarization + translation option for voice-agent pipelines that you can host yourself.
  • What: Six open-weight models in the 2B class (text, vision, audio, omni, decision) on one LFM3-1.7B backbone, trained on 22T tokens. The text model has 256K context and fits in about 1 GB at 4-bit. Every model can be fine-tuned on a single GPU.
  • License / serving: LFM Open License (lfm1.0 on the HF card). Weights are on Hugging Face, and the models are also on the Liquid API and Playground.
  • Builder take: A good base for on-device specialist fine-tunes and a reasonable local router or classifier.

GPT-6.1 Terra — Still open (2026-10-26)#

  • Cost: $1/$5 per 1M tokens, cached input $0.05 (on OpenRouter and per AA). That is $0.40 per Intelligence Index task at max effort, vs $1.40 for GPT-5.6 Terra. Coding Agent Index: 48 (Sol 54, Luna 31).
  • AA hallucination rate: 38%, down 50 points. AA's xhigh and low efforts sit on the Coding Agent Index cost-efficiency frontier.
  • Builder take: The cheap mid-tier default to benchmark against. See Simon Willison's notes.

Kimi K3.1 Pro — Still open (2026-10-23)#

  • Model: 2.8T total / 104B active parameters, 1M context.
  • Price: $2/$10, $0.20 cache hit (OpenRouter).
  • AA numbers: $1.98 per task. 55 on the AA Intelligence Index at launch (pre-v5).
  • Weights: Not published yet; AA says Moonshot is expected to release them "soon after launch". Once they are out, AA says it will be the leading open-weights model. Watch for the drop. (Simon Willison's notes)

DeepSeek V4.1 Flash 1017 — Context (HN surfaced the KV-cache paper in the window)#

  • Architecture (from the paper):
    • 552B backbone, with 16B active per token at decode and 8B at prefill.
    • Global KV cache cut to 890 bytes per token (about 1/4 of V4-Flash) using CSA2 cross-layer reuse plus FP4 KV.
  • DeepSeek-reported benchmarks: Terminal-Bench 4.0 42.3 (Claude Code harness), DeepSWE v1.1 77.0.
  • Price: $0.15/$0.60 per 1M tokens, $0.003 cache read on OpenRouter.
  • Weights: Top of HF trending (3,685 likes). Community GGUF quants are trending too (ISTA-DASLab).
  • Builder take: Still the cheapest serious open-weight agent backbone, and the base for Particle-1.

Pareto frontier#

Cost per task (Artificial Analysis, Intelligence Index v5 unless noted)

Point Score $/task Output tokens/task Source
Gemini 4 Carbon (high) 59 (#1) $7.76 (standard pricing) 123k AA
Claude Opus 5.5 (max) — per AA Carbon article — $10.04 173k AA
GPT-6 Astra (max) — per AA Carbon article — $6.25 — AA
Kimi K3.1 Pro (open weights pending) 55 (pre-v5 index, 2026-10-23) $1.98 55k AA
GPT-6.1 Terra (max) — $0.40 — AA (low and medium effort sit on the Pareto line)

Coding-agent numbers (AA Coding Agent Index v2.0)

  • Claude Code + Opus 5.5 (max): 62, $11.42 per task.
  • Devin Fusion (Opus 5.5 max + GPT-6 Luna high): 62, $8.00 per task.
  • GitHub Copilot Fusion (max) in Copilot CLI: $5.59 per task, 7.2M tokens per task (about half of Claude Code + Opus 5.5). AA puts it "2 points below" Claude Code + Opus 5.5; GitHub's changelog says it scored 60. Component scores: Terminal-Bench 4.0 71 (leads), FrontierCode 51, SWE-Atlas-QnA-AA 57.
  • Gemini 4 Carbon + Antigravity CLI: 59.
  • GPT-6.1 Terra: 48. Its xhigh and low effort settings sit on the cost-efficiency frontier.

Session economics: The headline this window is routing, not new models. Fusion gets close to the frontier coding score at about half the $/task by cascading and critiquing across models, and it bills each underlying model at its standard rate, with a time-limited 30% discount on paid Copilot plans (changelog). OpenRouter's rankings now include a "Cost per session" view for coding-agent sessions; the static page didn't render values, so no number is quoted here.

Open-weight and Chinese-lab points

Where sources disagree

  • Index scores: OpenRouter's benchmark panel, which is sourced from AA, shows decimal scores such as Claude Opus 5.5 at 58.4 and Kimi K3.1 Pro at 46.0. AA's Kimi launch article gave Kimi 55, before Index v5 launched on 2026-10-27. Check which index version a number comes from before comparing.
  • GPT-6.1 Terra price: Price Per Token's model-release feed lists $2.00 in / $10.00 out, while OpenRouter and AA list $1/$5. Trust the provider page before budgeting.
  • Terminal-Bench 4.0 numbers depend on harness. Kimi K3.1 Pro's 67% (model eval, AA), Carbon's 64% (Antigravity CLI), and Copilot Fusion's 71 (Copilot CLI) are not the same setup.

Coding CLIs, IDEs, harnesses#

  • News — GitHub Copilot Fusion GA
    • Three effort levels: Standard, High, Max.
    • Three orchestration modes: Single, Cascade (a cheap draft escalates if it fails a check), and Critique (draft → review → revise).
    • Admins control which models it can use.
    • Available in VS Code, the Copilot app, Copilot CLI, and JetBrains.
  • News — Fusion with local models
    • Managed download on Windows.
    • Copilot CLI can route to models you configure yourself through a local OpenAI-compatible server on Windows, macOS, or Linux (for example Qwen3.8 Flash Next).
    • Caveat: routing decisions still happen in the cloud, so the session isn't offline.
  • News — Copilot app Universe releases
    • Projects (public preview): multi-repo goals with background agents.
    • Computer use (public preview, macOS/Windows), with an approval prompt before it controls an app.
  • News — Copilot in VS Code Universe releases
    • Cloud sessions you can continue across VS Code, the app, Mobile, and the CLI.
    • A new Agent Host that keeps sessions running after the window closes, including on a remote machine over SSH or dev tunnels.
    • Control agents from your phone via code agent host --github-environment.
  • News — opencode v2.2.1 (2026-10-29 00:16 CDT)
    • /goal <objective> keeps a session running until a reviewer confirms the goal is met.
    • Opt-in /schedule prompts and an experimental image_generate tool.
    • Breaking: some provider entry points in @opencode/ai were removed or renamed (fal, Black Forest Labs, Replicate, Stability). The CLI, Desktop, and HTTP API are unaffected.
  • Update — Claude Code v2.1.312 (2026-10-27)
    • A subagent's autoCompactWindow now lowers CLAUDE_CODE_AUTO_COMPACT_WINDOW.
    • Prompt and agent hooks are capped at 24 concurrent per session.
    • Fixed Claude overwriting a very large CLAUDE.md after seeing only its start; Write now asks it to read the whole file or edit in place.
    • Automatic model switches after a flagged message now also cover unrecognized deployment names; CLAUDE_CODE_DISABLE_REFUSAL_FALLBACK=1 turns them off.
    • Bedrock, Vertex, and Foundry use the shorter system prompt for unrecognized deployment names (CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT=0 restores the full prompt).
  • News — Operator caveat: Claude Code agent_summary token burn (HN, 439 points)
    • The author's OTel export attributes 16.4% of Claude Code spend and 19.7% of API requests to a 3–5-word status-line summarizer for background agents.
    • The fix is an unofficial MIT-licensed Bun preload that disables that timer. It doesn't modify the binary, but an update can silently undo it.
    • Take: before you adopt it, group your own claude_code.api_request events by query_source.
  • Still open — Cursor Agent SDK (posted 2026-10-20; HN thread in window)
    • Agents are defined as code: Markdown prompts plus TypeScript tools, with built-in evals and event replay.
    • agent-sdk deploy runs the agent on Cursor's infrastructure, or you can run it yourself as a single Node process.
  • Context — OpenRouter top apps (today's tokens)
    • Hermes Agent: 1.99T
    • Claude Code: 1.4T
    • Codex: 863B
    • Kilo Code: 760B
    • OpenClaw: 646B
    • pi: 624B
    • DeepSeek Harness (new): 318B
  • Context — pi v1.4.0 (2026-10-21): compaction summaries no longer have a fixed output cap, and codemode queues parallel scripts instead of failing. Pi-based IDE JetBrains/thinkrail is trending (see §5).

Usable tools and open projects#

Stars come from the GitHub API, and "+today" from GitHub trending (daily), both read about 07:40 CDT.

Repo or tool What it is Stars / +24h Runnable today? Why a tinkerer should care
github/gh-aw GitHub Agentic Workflows CLI extension (now GA) 5,516 / — Runnable: gh extension install; MIT Markdown → Actions YAML agents with sandbox, firewall, and gh aw forecast cost estimates
anomalyco/opencode Open coding-agent TUI/desktop — Runnable: v2.2.1 released 2026-10-29 /goal loops and scheduled prompts with any provider, including local ones
Infini-AI-Lab/hermit Trains "harness adapters" that adapt a frozen agent without touching model or harness 331 / — Watch: v0.1.0, Apache-2.0, Python 3.12+ Harness-side continual learning to try against your golden tasks
Bernstein Deterministic Python scheduler that runs CLI coding agents in parallel git worktrees, merging only diffs that pass lint, types, and tests — Runnable: pipx install bernstein; Apache-2.0; beta, single maintainer Policy-as-code governance and signed run receipts for multi-agent runs
JetBrains/thinkrail Lightweight IDE built around the pi agent 5,664 / +1,116 Watch: Apache-2.0, pushed today JetBrains shipping a pi-native IDE; check install notes in the README
CopilotKit/OpenBot Self-hosted "AI coworkers" (AG-UI), each with its own browser, files, and only the tools you grant; "every action decided before it happens and recorded after" 7,784 / +240 Watch: MIT A reference design for computer-use permissions and audit trails
thedotmack/claude-mem Cross-session agent memory: captures, compresses, and re-injects context 104,234 / +324 Watch: Apache-2.0 Pair with the Memory Tax paper and measure net value against plain RAG
divagr18/qwenputer A RISC-V CPU compiled into Qwen3.5-4B weights, using Gated DeltaNet state as RAM (write-up) 183 / — Runnable (learn-from): weights on HF/ModelScope per the author A clear explainer on how linear-RNN layers can act as exact memory; HN, 554 points
tractordev/webframe Runs a web page like a program: console goes to stdout and exceptions to stderr, using the OS webview (single binary) 204 / — Runnable: Go, MIT Run browser-only code (WASM, WebGPU) from scripts or CI without a browser-automation stack; see Simon Willison's evaluation
Claude Code no-agent-summary preload Unofficial Bun preload that disables a background summarizer 68 gist stars / — Watch: measure first; updates can undo it Possible 15–20% token savings on background agents (author-reported)
obra/superpowers, elayadesign/ai-design-skills, Donchitos/Claude-Code-Game-Studios Skill and prompt packs +1,037 / +1,178 / +538 Skip Trending, but mostly prompt and skill markdown with no runtime

Keepers

  • gh-aw: The most production-ready thing this window. Agents start read-only, outputs pass through "safe outputs" validation plus a threat-detection job, and GA added ARC/Docker-in-Docker runners and micro-VM isolation (changelog). Using the default Copilot CLI engine requires a Copilot license, or you can bring your own agent.
  • Hermit: If you run a fixed harness (Claude Code, pi, opencode) and want it to improve on your task family without fine-tuning, this is the open framework to read. It pairs well with the HaMo paper below.
  • Open weights trending on HF

Agent builder stack#

Protocols and runtimes

  • News — Microsoft Copilot harness in Foundry:
    • Build on "the same harness Microsoft and GitHub ship in our own products" without maintaining it (announced by Foundry and GitHub leadership).
    • 20+ frontier models plus bring-your-own-key, and supports MCP, agent skills, and plugins.
    • The runtime was recently ported to Rust to cut session overhead and TTFT.
  • Still open — Cursor Agent SDK: agents as code with evals and event replay; self-host as a Node process.

Agent evals

  • News on HN — DAYJOB: Tax (Surge AI, 2026-10-27; HN, 246 points):
    • 50 professional tax environments, 1,789 workspace files, pass@1 at a 95% rubric threshold.
    • Claude Sonnet 5.5 leads at 10.8%, then Claude Fable 5.1 at 9.6% and Claude Opus 5.5 at 9.2%.
    • It's a useful template for judgment-heavy golden tasks where the agent has to raise issues nobody asked about.
  • News — TauIndianBankBench (code): 1,000 banking tasks that check that every required procedural step happened, not just the final database state. A pattern worth copying for tool-calling evals.
  • News — What Effort Level Should I Choose? (Together; code): higher reasoning effort can raise aggregate accuracy but doesn't reliably preserve tasks that lower effort already solves, and it can overspend. Run your effort sweep per task.

Context engineering and memory

  • News — Memory Tax (148 upvotes; code): many dedicated agent-memory systems underperform naive RAG over raw history. Benchmark your memory layer against plain retrieval before you ship it.
  • News — HaMo: Harness-Model Collaboration: combines fast experience cards in the harness with slow parametric updates for web agents. Your Policy Is Secretly a Harness (code) argues a trained policy inherits its harness's preferences, so changing the harness moves both the ceiling and the mismatch.

Sandboxes, identity, permissions

  • Update — Vercel fx adapter (2026-10-27):
    • The AI SDK HarnessAgent now runs the harness and model calls on the host, so model credentials stay out of the sandbox.
    • Filesystem and shell tools stay sandboxed.
    • Also adds tool filtering, mid-turn steering, and resumable sessions.
  • News — exe.dev: Claude and GPT without API keys: VMs authenticate to Anthropic and OpenAI with Workload Identity Federation, so there are no long-lived keys on the box.
  • Context — Formal: cloud agents get production access via OIDC (2026-10-23):
    • Cursor and Devin act as OIDC issuers, and the agent gets read-only, policy-checked access.
    • The caveat: any human identity in the token is asserted by the platform, not independently authenticated.

Computer use: Copilot app computer use is in public preview with per-app approval and admin disable. No independent reliability numbers yet, so treat it as a demo until measured.

Infrastructure dash#

  • Update — Cloudflare: cross-request KV-cache exposure in GLM-5.3 on Workers AI (2026-10-26)
    • What happened: 155 responses to prompts longer than 681,280 tokens contained data from other customers' requests (2026-08-27 → 09-09).
    • Root cause: a 1,048,576-token context setting over a 681,280-token prefill KV pool, combined with an SGLang disaggregated prefill/decode bug that sent placeholder KV to decode instead of rejecting the request.
    • Fixes: sglang#36988, included in v0.5.22 or later per Cloudflare. SGLang published a coordinated security update covering related cache-lifetime fixes and recommended operator actions (HN, 181 points).
  • News — DeepSeek-V4.1-Flash KV-cache paper (HN): 890 bytes per token of global KV via cross-layer reuse plus FP4 KV, and "SWA Bounded Replay" cuts persistent KV on SSD or host memory to about 1/8 of V4-Flash. Worth reading if you budget HBM for long-context agents.
  • News — TAO: tool-aware scheduling for agentic serving: across 12 settings, it reports up to 26.8% lower latency, 36.3% lower mean TTFT, and 29.5% higher decode throughput vs the best baseline by factoring tool-return readiness into admission and placement.
  • Still open — Cloudflare Web Search API (beta since 2026-10-02): one endpoint over Ceramic.ai, Exa, and Linkup, routed through AI Gateway for logs and billing at provider list price.
  • News — OpenAI API credits for ChatGPT subscribers (changelog 2026-10-28):
    • Pro tiers get $100, $200, or $500 a month in API credits.
    • Business gets $20 per Standard seat or $100 per Premium seat, capped at $500 per workspace.
    • Credits expire at the end of each billing cycle and don't cover Codex API-key usage.
  • Engine releases: not nominated beyond the SGLang item above.

Builder actions#

  1. Audit SGLang PD-disaggregation configs. Compare the KV pool size in each node's startup log against your configured context length. Lower the context length if the pool is smaller, and upgrade to SGLang v0.5.22 or later (Cloudflare guidance).
  2. Measure Claude Code background-agent overhead. Enable OpenTelemetry export, group claude_code.api_request by query_source, and check the agent_summary share before trying the preload. The gist says it was tested on Claude Code 2.1.257–2.1.312, and an update can silently undo it.
  3. Try one agentic workflow. Install gh-aw, run the quickstart on a test repo using one of the prebuilt workflows, and run gh aw forecast to see projected credit spend. The default Copilot CLI engine needs a Copilot license.
  4. A/B a routed vs single-model session. In Copilot CLI, run the same three tasks with Copilot Fusion Standard and Max vs a fixed model, and record tokens and $ per task. Compare against AA's $5.59 vs $11.42.
  5. Run a cheap-model bake-off on OpenRouter. Put the same 10-task coding set through particle-ai/particle-1, deepseek/deepseek-v4.1-flash-1017, and openai/gpt-6.1-terra (low effort). Track tool calls and output tokens per task.
  6. Test memory against a RAG baseline. Clone MemoryTax and compare your agent memory layer (or claude-mem) against plain retrieval over raw history.
  7. Try opencode /goal. Upgrade to opencode v2.2.1 and run /goal on a bounded refactor with tests as the reviewer's success signal.
  8. Run the effort sweep. Using ReasoningEffortLLM, compare low vs high effort on your own golden tasks and flag tasks that higher effort breaks.
  9. Remove long-lived LLM keys from agent VMs. Where your platform supports it, follow exe.dev's WIF write-up or Formal's OIDC pattern, and make issuer and claims explicit policy.
  10. Pull LFM3. Download LFM3-1.7B and try a single-GPU fine-tune for a narrow classifier or router.

Checkmarks stay in your browser only.

Source log#

Show the full source logEvery discovery source checked, hits and misses, and every confirm open.

Primary discovery

Source Result Note
Artificial Analysis home Hit Changelog: Carbon eval, Copilot Fusion, Index v5
AA articles Hit Opened Carbon, Copilot Fusion, GPT-6.1 Terra, Kimi K3.1 Pro, Index v5
OpenRouter rankings Hit Top apps and AA-sourced benchmark panel; cost-per-session values not in static HTML
OpenRouter models API Hit New IDs and list prices (Particle-1, GPT-6.1 Terra, Kimi K3.1 Pro)
HN front page Hit Gemini 4 Carbon, Qwenputer, Claude Code agent_summary
HN Algolia last 24h (AI/LLM/agent) Hit The OR query returned 1 hit, so it was re-run as per-keyword queries and merged
HN Algolia inference slice (vllm, sglang, trt-llm, tensorrt-llm, mooncake, "kv cache", inference) Hit Low yield: DeepSeek KV-cache paper, exe.dev WIF, SGLang advisory
GitHub trending daily Hit Stars-today confirmed against GitHub API totals
HF Daily Papers 2026-10-29 Hit Thin (10 papers); TAO, TauIndianBankBench
HF Daily Papers 2026-10-28 Hit Memory Tax, Effort Level, HaMo, Harness geometry
HF models trending Hit Via API
Price Per Token news + /agents, /coding, /open-source, /model-releases Hit Nominated Particle-1, Copilot harness, Carbon; GPT-6.1 Terra price disagreement noted
news.learnprompt.pro Miss SPA rendered empty
learnprompt.github.io fallback Miss SPA rendered empty
LearnPrompt/ai-news-radar data JSON (data/latest-24h.json) Hit Generated 2026-10-29 11:31 UTC; GitHub Universe changelogs, Hermit, Bernstein, Octoverse
RadarAI Hit Claude Code agent_summary, Cloudflare Web Search API, Bernstein
The Agentic Digest Hit 2026-10-28 issue: Formal OIDC, Vercel fx adapter
Simon Willison 2026-10-28 Hit Octoverse, OpenAI naming; webframe evaluation (10-26)
Simon Willison 2026-10-29 Miss Day archive 404, no posts yet
X (narrow) Not used Optional

Dailies in rank order

  1. Latent Space AINews: Partial. Index titles only (10-29 Gemini 4 Carbon issue); body not read.
  2. AlphaSignal: Partial. Headlines; several Pro-gated.
  3. Code4U AI Digest: Hit. Latest date 2026-10-28, 10 items with outbound links.
  4. TLDR AI archives: Partial. Archive index loaded (latest 2026-10-28); issue body not opened.
  5. TLDR flagship: Not checked. Only used if TLDR AI is blocked, and it wasn't.
  6. The Neuron: Partial. Titles; no builder-specific ships.
  7. AI Daily Brief archive: Partial. Titles; latest 2026-10-28 (Genie 4), nothing printed.

Confirm opens: nominated → confirmed

Dropped

  • Poolside "Laguna M.1" (AlphaSignal): nominated, but the primary shows the weights shipped in mid-2026. A stale recap, dropped.
  • "DeepSeek V5": nominated as a rumor (juejin / TechNode, test identifiers in a harness repo); no official release; dropped.
  • Gemini API request-storage default (Agentic Digest quick hit): no primary, dropped.
  • Cohere's report of a Qwen model contacting an outside server during a security test (Code4U): no vendor primary opened, dropped.
  • "Claude Code adds your email to the prompt" (Tell HN): no primary URL, dropped.

Engine blogs: SGLang was nominated (Code4U → Cloudflare postmortem, which names a coordinated SGLang post; also on the HN inference slice) and confirmed by the LMSYS/SGLang security update. vLLM, TRT-LLM, InferenceX: not nominated (a vLLM TorchTPU post surfaced on HN with 2 points; it falls below the rubric and isn't printed).