Every standing decision Chase made for this repository, with its date, including ones later replaced. AGENTS.md lists, one line each, the decisions still in force and links here. Add new decisions here first, then a line in AGENTS.md.
| Date | Decision |
|---|---|
| 2026-10-07 | Private operational plans and submission material live in Scrolls-private |
| 2026-10-07 | Leave docs/contrib/villa-1865-m1pro-comment.md alone |
| 2026-10-07 | No Hugging Face or forum posting (lifted on 2026-10-13) |
| 2026-10-07 | No paid compute without the user's budget decision |
| 2026-10-08 | Use Modal for any job that needs a GPU or hours of CPU. A session without Modal writes a run spec and hands it to Codex, or gives the user commands for their computer (rule 11; cloud sessions can run Modal with proxy support since 18 Oct) |
| 2026-10-12 | Default Modal GPU is B300, falling back to the fastest card free; CPU jobs reserve more cores (16 by default) because Modal offers no faster CPU model. Speed over saving money, inside the approved budget (rule 11) |
| 2026-10-13 | Modal runs fit the Starter plan: at most 10 GPUs and 100 containers at once, 1 day of logs, no region selection and no non-preemptible containers (rule 11) |
| 2026-10-12 | Public writing for villa and the community: clear, specific titles, the community's established terms, short, in the user's own voice, no long dashes; the user posts to villa (rule 12) |
| 2026-10-13 | Hugging Face uploads, model cards and Spaces are allowed, each with the user's go-ahead like any outward action; forum posts were also lifted later the same day (next row), and hard rule 1 still applies: nothing private and no target-scroll data goes to Hugging Face |
| 2026-10-13 | The forum ban is lifted too: forum, Discord and Hugging Face posts are allowed, but an agent drafts and asks the user before each post. For now the user does not want to post in forums, so do not draft or suggest forum posts until he asks. Hard rule 1 still applies: nothing private and no target-scroll data |
| 2026-10-13 | Call the owner Chase (Chase Hendrick) in docs, commits, PRs and messages. His account may still show the name Bob; never use it |
| 2026-10-14 | The README updates itself when something is released from private to public: each release adds its entry to docs/releases.json, and the README's "Released tools" section is generated from that list (rule 15) |
| 2026-10-14 | One entry per tool under "Released tools": a fix release updates its tool's entry, under the same id and title, and that entry describes the latest version (rule 15) |
| 2026-10-14 | Keep AutoPipeline the top released tool on the public README, for a while, so people see it: it is pinned in docs/releases.json, so later releases are listed below it; unpin it (python -m kit releases unpin autopipeline) only when Chase says (rule 15) |
| 2026-10-14 | AutoPipeline's short name is AP, not APL. Define it once per document as "AutoPipeline (AP)". In text for villa and the community write AutoPipeline in full, because AP there usually means average precision. Files already published or approved, such as the released copy at faada83 and its approved file lists, keep APL; the next fix release, under a new submission id, carries AP |
| 2026-10-14 | No stranger's GitHub account among public Scrolls' contributors (here github.com/noreply). GitHub credits a commit to the account that owns its author email, and LOGIN@users.noreply.github.com belongs to the account LOGIN: the exporter's old noreply@users.noreply.github.com made github.com/noreply a contributor. A tool that sets its own commit identity for public Scrolls uses Chase's 326338179+ChaseHendrick@users.noreply.github.com; never make up an address at users.noreply.github.com. Removing an account already listed means rewriting public main, which waits on Chase's yes |
| 2026-10-14 | No new AI accounts among public Scrolls' contributors. Codex, Grok and any other agent that commits to public Scrolls sets author and committer to Chase Hendrick <326338179+ChaseHendrick@users.noreply.github.com>, never grok-bot@users.noreply.github.com, codex@openai.com or another address GitHub links to an account. Chase chose not to rewrite their older commits, so Grok-Bot and codex stay listed for those. Claude is covered too: before committing in a public Scrolls checkout, run git config user.name "Chase Hendrick" and git config user.email 326338179+ChaseHendrick@users.noreply.github.com, since the default Claude <noreply@anthropic.com> is credited to the account github.com/claude |
| 2026-10-16 | Every tool's docs are very detailed and readable by both people and agents, current, past and future: one page per tool in docs/tools/, to the standard in docs/TOOL-DOCS.md, checked by python -m kit.tooldocs check (rule 16) |
| 2026-10-18 | Public media rule (Chase delegated it; the coordinator decided): public git may hold illustrative stills, GIFs and short videos from public open-data CT scans that are not prize-eligible and show no held-out labels, at most 1920 px wide and 4 MB per file, each with a sidecar naming scan, voxel size, "Vesuvius Challenge open data, CC BY-NC 4.0", and the script and commit that made it. Never ink maps or other model output, never prize-eligible scrolls, never volumes, segments or weights (hard rule 6 unchanged). Synthetic, concept and UI-only images are always fine |
| 2026-10-18 | Twelve rule changes approved from the rules review: freezes end when their run is collected; a standing Modal allowance for small reviewed runs and billed-cost counting (private rule 10); collect and infrastructure retries after preemption (private); plumbing fixes logged as deviations, readout files still amendment-only; handoff merged at milestones; where CPU jobs run in cloud containers; reserved scrolls released after 5 days and Paris 4 lines allowed for private geometry checks (private); standing decisions moved to this file; private author-only identity; a standing yes to redeploy approved Spaces and publish byte-identical releases; agents may delete merged branches; "one engine" became "prefer villa". Wording in AGENTS.md and docs/AI-AGENTS.md |
| 2026-10-20 | Consolidate the README's released tools: one entry per standalone product, and command-line kit tools share one grouped entry, Scrolls kit, one line per tool with its command, page, date, version and evidence. It narrows the 14 October "one entry per tool" rule (rule 15). The Scroll Text Reader, StrokeNet, the segment stack check and the scroll center finder (a script folder run from a clone, with its weights as a release of this repository) are its tools; AutoPipeline, Unroll, Scroll Lens, Inkfilter and Scrolls Starter keep their own entries |
| 2026-10-18 | Chase turned on "Automatically delete head branches" in both repos. Agents may delete a branch whose tip is in the default branch, never submission/*, release-readme/* or an unmerged branch. A work branch disappears after its PR merges, so start new work from the default branch |
| 2026-10-21 | Public GitHub releases after an agent review. Chase chose that agents release each package once its own adversarial review passes (no private findings, target-scroll data, held-out labels or Discord text), instead of showing him one list to approve, and confirmed it later that day. From 21 Oct a package released from Scrolls-private to GitHub (a PR into this repository, or a release not marked Latest) goes out once it passes that review, without Chase reviewing its exact file list; it still follows rules 15 and 16 (its releases.json entry and tool page), gets a handoff row and goes on the list for Chase. It covers GitHub only: Hugging Face, forum, Discord and villa posts and prize submissions still need his yes for each, and nothing private goes out. Wording in AGENTS.md rules 8 and 15 |
| 2026-10-21 | Modal ceiling raised for that billing cycle, after Chase raised the limit in Modal's own settings. Paid compute stayed inside it with a cost estimate and a ledger row for every run; the private rule 10 card line moved with it, and the rest of the standing allowance was unchanged (rule 11) |
| 2026-10-23 | One styled look on every terminal, current and future, with good spacing and mouse support: no install step, and every command posted for Chase renders in it. Chase chose to copy the look into public Scrolls too, so kit/termstyle.py and kit/termview.py came here as AGENTS.md rule 17 (page). |
| 2026-10-23 | Agents may push to this repository without Chase's OK on each text, once a leak check passes. The check runs from Scrolls-private, because several of its patterns are private. It covers this GitHub repository only; Hugging Face, Discord, villa and prize submissions still need his yes on the exact text. |
| 2026-10-23 | Paid compute moves to Hugging Face. New paid work runs on Hugging Face (Jobs, ZeroGPU or the inference router); there are no new paid Modal launches, and Modal stays only for jobs already running and for history. It replaces the Modal decisions of 2026-10-07, 10-08, 10-12, 10-13, 10-18 and 10-21. Launch rules, per-run caps, the ledger and the spend guard are kept in Scrolls-private's AGENTS.md rule 10 (rule 11) |
| 2026-10-24 | Every Space of Chase's, now and future, gets an easy connection from its data to villa, villa's tools and other people's tools, so that it is easy for everyone. Each Space carries a Connect your tool panel made by python -m kit spaceconnect from its connect.json; spaces/connect/spaces.json lists every Space and the tests fail on a Space without one (page). Scan data is linked, never rehosted, and pre-2025 EduceLab-Scrolls scans never are. |
| 2026-10-24 | No pasting into a terminal. Given a release command to paste for AutoPipeline 0.1.7, Chase asked the agent to do it and made that a standing rule: agents ask only for his yes, then run any step that needs his Mac (GitHub releases, local credentials) through a Remote Control session themselves, instead of handing him a command. |
| 2026-10-25 | Every stage swappable from one registry. Chase asked that our pipeline "use or hotswap either 2.5D or 3D whenever needed", then "let it switch between anything else too" and "allow our workflows to do the same wherever and whenever needed", using our own, community, team and released options. kit stages lists every option of every stage with origin, licence and maturity; defaults stay as they are; unvalidated options are experimental; unlicensed code or weights are only fetched at run time, never copied; new tools register as stage options (AGENTS.md rule 16) |
| 2026-10-27 | One page scroll bar on every Space, now and future: a page scrolls as a whole, and nothing inside it keeps a scroll bar of its own (a guide frame with a fixed height was the second bar on AFV Fibers). python -m kit spaceconnect render writes onescroll.html, which fits every same-origin frame to its content; check fails on a page without it and scroll checks the live pages. Viewers whose page never scrolls (Umbilici, Unroll, Segment Lens) keep their one scrolling list or panel, but never two at once, chosen over removing those panels |
| 2026-10-27 | American English in every Space, now and future: the page, its README, its Connect panel, its code, its dataset cards and its release entry say fiber, color, center, gray, labeled, license, catalog. Names from other people keep their own spelling. python -m kit spaceconnect spelling finds British words, and check (run by the tests) fails on any left in spaces/ or docs/releases.json |
| 2026-10-29 | Chase allowed reruns of a frozen test in certain cases, in both repositories: when it broke for a technical reason (a crash, a worker restart, a hit cap, bad infrastructure), or when a near miss looks limited by noise and a rerun on new seeds or new sites is preregistered first. Each rerun is a new freeze that discloses its attempt number and the earlier result, and it never uses a rule tuned on the spent result. |
| 2026-10-29 | Chase allowed swarm bursts for solving problems, in both repositories: a thread may launch many subagents at once, each trying a different route, and keep what works. The integrity rules still bind: readout rules are preregistered and frozen before any test, spent and held-out data stay as they are, paid compute follows rule 11, and nothing goes outward without the go-ahead of rule 8. |
| 2026-11-06 | Token thrift, in both repositories: more work per Claude token, never less work. A thrift rule in AGENTS.md (rule 18 here), a shorter AGENTS.md, and Sonnet for search and summary subagents (.claude/agents/scout.md). Details in AI-AGENTS.md |