projectfile.org

org.projectfile.ai

A project’s stance on AI and LLM use, over three independent directions. INBOUND: attitude (encouraged | allowed | neutral | discouraged | prohibited) as the default for every entry of the activities map (channels — pull-requests, commit-messages, bug-reports, discussions, code-review, security-reports; content — code, code-comments, documentation, translations, prose, images, audio, video; tasks — refactoring, bug-fixing, tests, grammar-tools) — one vocabulary, so an absent activity key means exactly attitude; autonomy (any | assisted | none) for what kind of use is permitted; applies-to (everyone | contributors) for whom it binds; the disclosure trio disclose-required / disclose-trailer / disclose-details; obligations, the duties a contributor keeps when a tool was involved; the gates issue-required and excluded-labels; enforcement, the escalation ladder for a violation; and statement, free localized prose carrying the reason the enums cannot. INTERNAL: project-use, a map of the same activity keys to an autonomy token, declaring how the project itself uses AI. OUTBOUND: content-signals, consent values from contentsignals.org (IETF aipref), order-preserving and open-vocabulary. Unknown activity keys are accepted and preserved. An absent namespace means NO DECLARED POLICY and MUST NOT be rendered as a permissive default. Includes the org.projectfile.ai.skills sub-namespace for agent-runnable skill manifests. Consumed by Bridges to render the policy file — filename, default AI_POLICY.md — and to cross-reference it from README and CONTRIBUTING.

  • Namespace: org.projectfile.ai
  • Owner: projectfile.org (community-maintained)
  • Status: proposed

Shape

Section titled “Shape”

Illustrative fragment — not a complete projectfile document. Consumers MUST round-trip unknown namespaces even when no shape is published.

---# SPDX-FileCopyrightText: 2026 Damián Búho <damian.buho@proton.me>## SPDX-License-Identifier: CC-BY-4.0# SPDX-License-Identifier: MIT
# Shape: org.projectfile.ai extension namespace# Illustrative fragment — not a complete projectfile document.# See registry.yaml for ownership and status.## Declares a project’s stance on AI and LLM use. Bridges consume it to render# the policy file (AI_POLICY.md by default) and to cross-reference that file# from README and CONTRIBUTING.## ---------------------------------------------------------------------------# Absence is not permission# ---------------------------------------------------------------------------## A document with no org.projectfile.ai namespace has NO DECLARED POLICY.# Consumers MUST NOT read that as a permissive default and MUST NOT render a# stance the project never took. Presence of the namespace IS the opt-in —# there is no `enabled` gate, because a policy namespace with no policy in it# has no other meaning.## ---------------------------------------------------------------------------# Three directions# ---------------------------------------------------------------------------## The namespace answers three independent questions. Consumers MUST NOT derive# any one of them from another:##   INBOUND   what may come INTO the project — `attitude`, `autonomy`,#             `applies-to`, `activities`, `obligations`, the disclosure keys,#             the contribution gates, `enforcement`. Governs contributors.#   INTERNAL  how the project itself uses AI — `project-use`. Governs#             maintainers.#   OUTBOUND  how the project’s own content may be consumed BY AI systems —#             `content-signals`. Governs crawlers and trainers.## A project may welcome LLM-assisted patches, refuse to be training data, and# hand-run every release. No answer constrains the others.## ---------------------------------------------------------------------------# Document name# ---------------------------------------------------------------------------##   filename  OPTIONAL (default: AI_POLICY.md) — the basename a renderer#             writes. No standard exists: AI_POLICY.md, AI-POLICY.md, AI.md#             and LLM.md are all in the wild, so the project names its own#             file and every consumer finds it at one declared place.##             MUST be a basename. A value carrying a path separator, a `..`#             segment or an absolute path MUST be rejected, never sanitized —#             this field names a file a tool writes.## ---------------------------------------------------------------------------# Permission vocabulary# ---------------------------------------------------------------------------## ONE closed enum serves `attitude` and every activity:##   encouraged | allowed | neutral | discouraged | prohibited## `attitude` is REQUIRED and sets the project default. An activity overrides# it for that activity alone; an absent activity means "same as attitude". The# derivation is therefore the identity function — no attitude is ever# strengthened or weakened on its way to an activity.## `neutral` means no stated position: ask first. It is NOT a synonym for# `allowed`, and `discouraged` is NOT a synonym for `prohibited`; consumers# MUST NOT render either pair as equivalent.## ---------------------------------------------------------------------------# Autonomy# ---------------------------------------------------------------------------##   autonomy  OPTIONAL (default: any) — WHAT KIND of use is permitted, which#             is a separate question from how much:##               any       autonomous agents may act on their own#               assisted  a human drives, reviews and submits#               none      no LLM anywhere in the loop##             Orthogonal to the permission vocabulary, and it applies to every#             activity rather than to pull requests alone: `bug-reports:#             allowed` with `autonomy: assisted` accepts a human-filed report#             written with an LLM and refuses one filed by a bot.## ---------------------------------------------------------------------------# Who the policy binds# ---------------------------------------------------------------------------##   applies-to  OPTIONAL (default: everyone) — everyone | contributors.##               `contributors` binds outside contributions only and exempts#               the maintainers, who answer for their own practice under#               `project-use`. Projects take this shape when the rules exist#               to protect reviewer time rather than to state a position on#               the tools; consumers MUST render the exemption, because an#               unstated one reads as hypocrisy the first time a reader finds#               an AI-assisted maintainer commit.## ---------------------------------------------------------------------------# Disclosure# ---------------------------------------------------------------------------##   disclose-required  OPTIONAL (default: false) — whether contributors must#                      disclose LLM assistance.##   disclose-trailer   OPTIONAL — the trailer key contributors write, e.g.#                      `Assisted-by`. With it a consumer renders the exact#                      line to paste into a commit message; without it the#                      rendered prose stays generic.##   disclose-details   OPTIONAL — sequence saying WHAT the disclosure must#                      carry. Open vocabulary, declared order preserved.#                      Known values:##                        tool    name the tool (Claude Code, Cursor, Amp …)#                        extent  how much of the work it did#                        prompt  include the prompt that produced the change##   All three are ignored when disclose-required is false.## ---------------------------------------------------------------------------# Obligations# ---------------------------------------------------------------------------##   obligations  OPTIONAL — sequence of duties a contributor takes on when AI#                was involved. Open vocabulary; declared order preserved.#                Known values:##                  understand       able to explain the change, and how it#                                   meets the rest of the system, without the#                                   tool#                  review           self-review before asking anyone else to#                                   read it#                  test             exercised on a platform the submitter can#                                   actually run, never “hypothetically#                                   correct”#                  edit             generated text is trimmed by a human — no#                                   verbatim dumps, no comments describing#                                   what the model did#                  human-reply      a human answers human feedback; replies to#                                   maintainers are never automated#                  license-cleared  no third-party code unless its licence#                                   permits the use and its terms are met#                  attest-original  the submitter states the work is their own##                This is where the quality bar lives. `attitude` says whether#                the tool may be used; obligations say what the human still#                owes, which is the half most policies are actually about.## ---------------------------------------------------------------------------# Contribution gates# ---------------------------------------------------------------------------##   issue-required   OPTIONAL (default: false) — an AI-assisted contribution#                    MUST reference an issue the project has already accepted.#                    Drive-by patches against no agreed problem are closed:#                    the cost of a generated patch is paid by the reviewer,#                    and this is the field that moves the decision earlier.##   excluded-labels  OPTIONAL — sequence of issue labels off-limits to AI#                    whatever `attitude` says, e.g. “good first issue”: an#                    issue that exists to teach a newcomer is spent, not#                    solved, when a model closes it.## ---------------------------------------------------------------------------# Enforcement# ---------------------------------------------------------------------------##   enforcement  OPTIONAL — sequence of consequences for violating this#                policy, in escalation order. Open vocabulary; declared order#                preserved and rendered as the ladder it is. Known values:##                  warn     a first violation gets a warning#                  close    the contribution is closed unreviewed#                  revoke   commit or review privileges are withdrawn#                  ban      the account is barred from future contributions#                  publish  the violator is named on a public list; the list#                           URL goes in links[type=ai-blocklist]##                A policy that states rules and no consequence is a request.#                Absence means the project has declared none — consumers MUST#                NOT invent one.## ---------------------------------------------------------------------------# Statement# ---------------------------------------------------------------------------##   statement  OPTIONAL localized-string (§7) — free prose, rendered verbatim#              above the derived sections. Carries the one thing the enums#              cannot: WHY. A policy that states rules and no reason reads as#              bureaucracy, and readers discount it accordingly.## ---------------------------------------------------------------------------# Content signals# ---------------------------------------------------------------------------##   content-signals  OPTIONAL — sequence of consent signal values from#                    contentsignals.org (IETF aipref). Open vocabulary;#                    consumers MUST accept unknown values and MUST preserve#                    declared order. Known values:##                      ai-train  content may be used for AI/ML training#                      search    content may appear in search results#                      ai-input  content may be used as AI inference input##                    Absence means "no consent signals declared" — neither a#                    prohibition nor a permission. The licence governs reuse#                    independently, and most licences say nothing about#                    training, which is the gap these signals address.## ---------------------------------------------------------------------------# Activities# ---------------------------------------------------------------------------## `activities` is a MAP from activity key to a token of the permission enum.# It is a map, and not a set of keys spread over the namespace, so that the# namespace’s own key space stays closed: a policy key added in a later# revision can never be mistaken for an activity a project declared.## Open at the KEY level, closed at the VALUE level: consumers MUST accept an# unknown activity key, MUST preserve it on round-trip, and SHOULD render it# under its own name. Known keys fall in three families a project mixes# freely:##   channel  pull-requests, commit-messages, bug-reports, discussions,#            code-review, security-reports#   content  code, code-comments, documentation, translations, prose,#            images, audio, video#   task     refactoring, bug-fixing, tests, grammar-tools## The content family is what lets a project split the media question that no# single "no AI-generated media" flag can answer: `images: prohibited` beside# `audio: allowed` is one project’s real position, not an edge case.## Navigational URLs go in top-level links[]:#   - { type: ai-policy,    url: "https://..." }   # authoritative policy page#   - { type: ai-blocklist, url: "https://..." }   # public list, `publish`## A project whose authoritative policy is an external governance page declares# it there. The fields here stay the machine-readable form of that same# stance, so a consumer never has to fetch a page to answer a question.## ---------------------------------------------------------------------------# Project use (INTERNAL)# ---------------------------------------------------------------------------##   project-use  OPTIONAL — map from an activity key (same vocabulary as#                `activities`) to an AUTONOMY token — any | assisted | none —#                declaring how the project itself uses AI. No new vocabulary:#                the axis is the one `autonomy` already applies to#                contributors, pointed the other way.##                  merges: none              no model approves or merges#                  releases: none            a human tags, builds, publishes#                  pull-requests: assisted   maintainers draft with a tool##                Absence means the project has not said. A project that#                demands disclosure from contributors and says nothing about#                its own bench has answered half the question, and it is the#                half a reader trusts least.## ---------------------------------------------------------------------------# Skills (sub-namespace: org.projectfile.ai.skills)# ---------------------------------------------------------------------------## Describes agent-runnable skills and commands tied to the project.# Shape not yet specified — a future revision will define the manifest# schema (skill name, description, trigger, tooling, constraints).# Consumers MUST preserve this subtree on round-trip.
org:  projectfile:    ai:
      # --- Rendering ---
      filename: AI_POLICY.md      # OPTIONAL  basename only                    (default: AI_POLICY.md)
      # --- Top-level policy (INBOUND) ---
      attitude: allowed      # REQUIRED  encouraged | allowed | neutral | discouraged | prohibited
      autonomy: assisted      # OPTIONAL  any | assisted | none                        (default: any)
      applies-to: contributors      # OPTIONAL  everyone | contributors                 (default: everyone)
      # OPTIONAL  localized-string (§7); rendered verbatim, never summarized      statement:        en: >-          We accept LLM-assisted contributions because reviewing a patch is          our job either way. What we will not accept is a patch the          submitter cannot explain.        es: >-          Aceptamos contribuciones asistidas por LLM porque revisar un parche          es nuestro trabajo de todos modos. Lo que no aceptaremos es un          parche que quien lo envía no sepa explicar.
      disclose-required: true      # OPTIONAL  (default: false)
      disclose-trailer: Assisted-by      # OPTIONAL  trailer key; ignored when disclose-required is false
      disclose-details: [tool, extent]      # OPTIONAL  tool | extent | prompt; open vocabulary, order preserved
      obligations: [understand, review, test, edit, human-reply]      # OPTIONAL  open vocabulary, order preserved
      issue-required: true      # OPTIONAL  (default: false)  AI work needs an already-accepted issue
      excluded-labels: ["good first issue"]      # OPTIONAL  labels off-limits to AI whatever attitude says
      enforcement: [warn, close, ban]      # OPTIONAL  escalation order; list URL in links[type=ai-blocklist]
      content-signals: [search, ai-input]      # OPTIONAL  open vocabulary from contentsignals.org; order preserved
      # --- Activity overrides (INBOUND) ---      # OPTIONAL map; an absent activity means "same as attitude".      # Same enum as attitude; unknown KEYS are accepted and preserved.
      activities:        translations: encouraged        # a small project ships more locales this way than it ever would by hand
        grammar-tools: encouraged
        security-reports: prohibited        # an unverified LLM-written report costs a maintainer more than it saves
        prose: discouraged        images: prohibited        video: prohibited        audio: allowed        # the media split one flag could not express: generated cover art is        # not the same decision as a generated test tone
      # --- Project’s own use (INTERNAL) ---      # OPTIONAL map; activity key → autonomy token (any | assisted | none).
      project-use:        pull-requests: assisted        code-review: assisted        merges: none        releases: none
      # --- Skills sub-namespace ---      # Shape TBD — consumers preserve on round-trip.
      skills: {}

.onion.i2p