docs: inference shapes — service-author obligations

[monsieurbulb]

Reading order — Network Public design-doc series
Tracking issue: #803 · Source paper: The Network Public v4.3
Previous in series: #799 (patience tax)
Next in series: #801 — Cross-type bridges

---

Summary

Adds docs/inference-shapes.md — defines three minimum properties (session-length neutrality, first-class exit, reflective interruption) that any inference service running on JAR must satisfy. Failure of any of the three is admissible grounds for refusal under docs/governance-refusal.md.

Why

JAR's coinless thesis pushes economic complexity to the service layer — services launch their own tokens, the base layer stays neutral. This is correct as a default, but it has a corollary: the base layer must still refuse services that re-enclose what JAR is meant to keep open. A service that locks users into long sessions, withholds their data on exit, or hides its own uncertainty is enclosing. Without a stated standard, refusing it is ad-hoc.

This PR makes the standard explicit. The three shapes correspond to three lock-in mechanisms:

Lock-in mechanism	Shape that rules it out
Session continuity	Session-length neutrality
Data hostage	First-class exit
Opacity	Reflective interruption

Design choices worth challenging

• Soft enforcement. The base layer does not enforce shapes at consensus. Compliance is declared by service authors, reviewable publicly, and refusable under docs/governance-refusal.md. The intent is not to litigate every service — most will satisfy trivially. The intent is that the substrate is allowed to refuse those that don't.
• Three, not seven. The paper proposes these three as minimum. Other candidates (data residency, training-data provenance, derivative-inference treatment) are interesting but not foundational in the same way.
• Self-declaration with public review. A service author publishes a shape declaration; reviewers can challenge it; failure is grounds for pause/refusal.

Scope

• Adds one file: docs/inference-shapes.md.
• No code changes. No consensus changes. The shapes are service-author obligations, not protocol invariants.
• Includes worked examples (a typical chat service that satisfies the shapes; one that doesn't).

Relationship to existing issues

• #105 — CoreVM service runtime is the natural place inference services land. Shapes are the standard those services declare against.
• Sibling to docs/network-public.md (parent), docs/governance-refusal.md (the refusal pathway), docs/negation-layer.md (deliberative review of shape compliance).

Reviewer notes

The three shapes are the most consequential framing decision. Are these the right three? Should the list be extensible? Should evaluation be self-declared with public review (current proposal) or independently verified?

[github-actions]

Genesis Review

Comparison targets:

• 0c6dcb86 (0c6dcb8)
• 7ff56b0f (7ff56b0)
• 1b8b2e0a (1b8b2e0)
• 0a483d08 (0a483d0)
• bf816b61 (bf816b6)
• 4b62216e (4b62216)
• c539e5c5 (c539e5c)

How to review

Post a comment with the following format (rank from best to worst):

/review
difficulty: <commit1>, <commit2>, ..., <commitN>, currentPR
novelty: <commit1>, <commit2>, ..., <commitN>, currentPR
design: <commit1>, <commit2>, ..., <commitN>, currentPR
verdict: merge

Use the short commit hashes above and currentPR for this PR.
Each line ranks all comparison targets + this PR from best to worst.

To meta-review another reviewer's comment, react with 👍 or 👎.

[monsieurbulb]

cc @sorpaas @olanod for review when you have a moment — this is a design-doc PR (no code changes), part of the Network Public series tracked from #796.

[sorpaas]

/review
difficulty: 0c6dcb8, 7ff56b0, 1b8b2e0, currentPR, 0a483d0, bf816b6, 4b62216, c539e5c
novelty: currentPR, 1b8b2e0, 0c6dcb8, 7ff56b0, 0a483d0, bf816b6, 4b62216, c539e5c
design: currentPR, 0c6dcb8, 7ff56b0, 1b8b2e0, 0a483d0, bf816b6, 4b62216, c539e5c
verdict: notMerge

Adds docs/inference-shapes.md (142 lines) — defines three minimum properties for inference services running on JAR: session-length neutrality, first-class exit, reflective interruption. Explicitly framed as service-author obligations enforceable via the right-to-refuse rather than consensus, which keeps the base layer neutral. Each property is defined precisely with what it rules out and admissible test patterns. Design quality is high, novelty is significant (anti-enclosure obligations are not in any other layer of the JAR design today). Holding off on merge: as with the rest of the Network Public series, this is design speculation explicitly requesting feedback, with the obligations directly invoking #797 governance-refusal mechanism that is itself unratified. Needs maintainer review beyond Genesis scoring.

[github-actions]

JAR Bot: Review recorded from @sorpaas (1 reviews, 0 meta-reviews).
Merge weight: 0/37665 (need >50%).

[sorpaas]

/review
difficulty: 0c6dcb8, 7ff56b0, 1b8b2e0, currentPR, 0a483d0, bf816b6, 4b62216, c539e5c
novelty: currentPR, 1b8b2e0, 0c6dcb8, 7ff56b0, 0a483d0, bf816b6, 4b62216, c539e5c
design: currentPR, 0c6dcb8, 7ff56b0, 1b8b2e0, 0a483d0, bf816b6, 4b62216, c539e5c
verdict: merge

Updating to merge. Adds docs/inference-shapes.md (142 lines) — defines three minimum properties for inference services: session-length neutrality, first-class exit, reflective interruption. Each property is defined precisely with what it rules out, and is explicitly framed as a service-author obligation enforceable via #797 (right to refuse) rather than consensus, which keeps the base layer neutral. Anti-enclosure obligations are not anywhere else in the JAR design. Lives in docs/, no protocol implications. No security implications.

[github-actions]

JAR Bot: Quorum reached — triggering merge.
Reviews: 1, meta-reviews: 0.
Merge weight: 34219/37665 (>50%).