meta/CURATION.md

Resource Selection Guide

Include only recurring AI-agent and coding-agent systems with explicit triggers, verification, durable state, budgets, escalation, and exit. General agent links belong elsewhere.

Acceptance Test

A resource belongs when it passes all three checks:

1. Scope fit: It covers AI/coding-agent loops or a direct foundation for repeated runs that discover work, delegate, verify, persist state, decide next actions, and escalate.
2. Builder value: It helps someone design, run, verify, evaluate, operate, or critique recurring agent systems.
3. Inspectable evidence: It is public, specific, and stable enough for readers to verify.

Strong Signals

Prefer resources with one or more of these properties:

• Primary source from the author, project, vendor, or research team.
• Official docs for an agent runtime, SDK, workflow system, benchmark, or eval framework.
• Concrete implementation detail: commands, architecture, traces, code, loop instructions, automation configs, hooks, schedules, state files, worktrees, checks, or failure modes.
• Durable research foundation: ReAct, reflection, self-correction, planning, memory, tool use, evaluation, or state.
• Operational value: CI repair, PR babysitting, deploy verification, docs drift, feedback clustering, cost control, or escalation.

Weak Signals

Usually reject resources that are mostly:

• broad AI-agent trend commentary;
• generic prompt tips with no state, tools, verification, scheduling, or retry loop;
• generic context or harness resources with no repeated-run, state, verification, trigger, or escalation angle;
• pure vendor marketing without technical substance;
• unrelated event loops, growth loops, control theory, game loops, or generic automation;
• private, unstable, paywalled, or hard-to-verify sources;
• duplicated coverage where a primary source is already listed.

Evidence Tiers

Tier	Meaning	Typical examples
A	Primary or official source	Paper, official docs, project README, author write-up
B	Practitioner source with implementation detail	Field note, runbook, postmortem, architecture note
C	Synthesis or high-quality explainer	Taxonomy, comparison, tutorial
D	Commentary or news coverage	Useful only for origin, adoption, quotes, or debate

Prefer the highest-tier source that explains the same idea clearly.

How Summaries And Metadata Are Built

Every released entry links to evidence that readers can inspect and exposes its category, summary, and source details for correction. Automation may assist discovery, URL resolution, duplicate detection, bibliographic extraction, repository statistics, and draft normalization. Automated output is not evidence; original or official sources, primary metadata, and stated limitations control the final entry.

The public dataset keeps those layers separate:

• annotation, key_contribution, novelty, and impact are original summaries;
• source_title, source_description, authorship, date, venue, DOI, and repository statistics come from the source check and record where the metadata came from;
• evidence_tier records whether the row is a primary source, implementation-grounded practitioner source, or synthesis and discovery index;
• loop_layer records whether recurrence lives inside the model, agent, harness, workflow, operations, evaluation, or spans layers;
• scope_fit records whether a resource directly describes operational Loop Engineering, enables it, or is an explicitly adjacent foundation;
• signal_strength describes the kind of evidence available, not whether the work is recommended or scientifically valid;
• source_status records reachability at one point in time and is not a guarantee of future availability.

No author, company, venue, star count, or citation count can buy or guarantee placement. Vendor inclusion does not imply endorsement, and inclusion of a critique does not imply that every claim in it is accepted.

Corrections And Disputes

Accuracy corrections take priority over expansion. Authors and readers can submit the annotation-correction form or a pull request with the current entry, the disputed text, the proposed correction, and primary-source evidence.

When evidence conflicts:

1. prefer the original publishing venue, official documentation, paper record, or repository over secondary coverage;
2. distinguish a verified fact from editorial interpretation;
3. preserve material caveats, negative results, and stated evaluation limits;
4. omit a date, venue, author, or metric rather than infer it from an unreliable source;
5. record the correction in the next release notes when it changes the meaning of an entry.

Source Checks And Versioning

data/resource_source_audit.csv records a point-in-time source check. data/resources.csv and data/resources.jsonl are deterministic exports of the full English guide enriched with those results. GitHub Releases identify versioned snapshots; downstream users should cite a release or commit when reproducibility matters.

Annotation Rules

Each annotation should answer: why does this matter for Loop Engineering? Lead with the mechanism, result, use case, or limitation. Avoid self-referential openings such as "this paper explores" or "this page provides."

Good:

| 🔁 **[Autonomous Loops](https://example.com)**<br><sub>Pattern</sub> | **2026** · Example publisher | Shows how task files, stop hooks, hard limits, and a kill switch form a self-continuing agent loop. |

Weak:

| 📝 **[Cool Agent Article](https://example.com)**<br><sub>Blog</sub> | **Example publisher** | Interesting article about agents. |

Resource Type Labels

Every resource entry in README.md should use one visible type label:

• 📄 Paper
• 📝 Blog
• 📚 Docs
• 🧰 Tool
• 🧪 Benchmark
• 🔁 Pattern
• 🧾 Template
• 🧭 List
• ⚠️ Critique

Pattern Quality Bar

Pattern entries should be concrete enough to adapt to a real agent runtime. A good pattern states:

• objective;
• trigger or cadence;
• intake source;
• agents and roles;
• workspace and permissions;
• durable state;
• verification gates;
• retry budget;
• exit condition;
• escalation path;
• loop instruction, automation spec, hook config, or scheduled command;
• failure modes.

License Scope

Only original repository summaries, templates, pattern documents, and metadata are released under CC0-1.0. Linked third-party resources keep their own licenses and terms.