TAS-098: Deterministic import/effect taxonomy, receipts, and replay limits

[AtlantisPleb]

Summary

Publish a deterministic import and effect taxonomy with effect receipts and replay limits so broader real programs can be admitted without collapsing internal compute into opaque host behavior.

Problem

The current import story is honest, but still too narrow for broader real-program support.

Current starting point:

• fixtures/tassadar/reports/tassadar_import_policy_matrix_report.json freezes exactly 1 allowed internal deterministic stub case, 1 delegated sandbox case, and 2 refused cases.
• The current matrix distinguishes env.clock_stub as a deterministic internal stub, sandbox.math_eval as explicit sandbox delegation, and host.fs_write as a refused unsafe side effect.
• The claim boundary is already correct: deterministic internal stubs, sandbox delegation, and refused side effects stay distinct.
• The missing piece is breadth: there is no reusable deterministic vs nondeterministic effect taxonomy, no effect receipt family, and no replay-limit contract for richer import-mediated workloads.

Without that, richer programs either remain unsupported forever or silently widen into ambient tool use.

Hypothesis

A typed import/effect taxonomy with deterministic vs nondeterministic classes, effect receipts, and replay limits will preserve claim boundaries while admitting useful real programs.

Source Papers / Docs

• Can LLMs Be Computers?
• MuMath-Code
• Learning Transformer Programs
• docs/ROADMAP_TASSADAR.md
• docs/TASSADAR_WASM_RUNBOOK.md

Target Surfaces

• psionic-runtime
• psionic-sandbox
• psionic-provider
• psionic-eval
• psionic-router
• fixtures/tassadar/reports/

Dependency markers that must remain explicit outside standalone psionic:

• kernel-policy
• world-mounts

Initial Scope

• Define import/effect classes for deterministic internal stubs, deterministic host-backed state, explicit sandbox delegation, bounded nondeterministic inputs, and refused side effects.
• Publish an effect-receipt schema that records which effect class ran, with what evidence and replay posture.
• Define replay limits and refusal rules for nondeterministic or stateful imports.
• Add one bounded durable-state import profile instead of treating all stateful imports as either impossible or silently internal.
• Keep external delegation explicit and challengeable rather than rebranding it as internal compute.

Benchmark / Tests

• Deterministic-vs-nondeterministic import conformance tests.
• Effect-receipt schema and serialization tests.
• Replay-limit tests across deterministic, stateful, and delegated effect classes.
• Refusal tests for disallowed effect classes, policy-denied delegation, and missing effect evidence.
• Regression tests preserving the current matrix’s explicit internal/delegated/refused split.

Claim Discipline

This issue widens the effect model. It does not collapse sandbox delegation into internal exact compute, and it does not grant general side-effect authority inside standalone psionic.

Done When

• Tassadar publishes a reusable import/effect taxonomy instead of one narrow matrix.
• Effectful executions emit receipts with explicit replay posture.
• Durable-state and delegated imports stay distinct from internal exact execution.
• Unsupported or unsafe effect classes remain typed refusals.

[AtlantisPleb]

Landed on main in 57f6af2 (Add Tassadar effect taxonomy and replay policy).

What changed:

• psionic-runtime now owns a reusable TassadarEffectTaxonomy, TassadarEffectReceipt, and replay-limit negotiation over deterministic internal stubs, deterministic host-backed state, sandbox delegation, receipt-bound nondeterministic inputs, and refused unsafe side effects.
• psionic-sandbox now publishes the widened sandbox-facing effect boundary with one admitted durable-state profile instead of treating all stateful imports as impossible.
• psionic-router now publishes fixtures/tassadar/reports/tassadar_effect_route_policy_report.json, keeping internal exact, host-state replay-bound, sandbox delegation, receipt-bound input, and refused routes explicit.
• psionic-eval now publishes fixtures/tassadar/reports/tassadar_effect_taxonomy_report.json, covering admitted and refused negotiated cases with explicit receipts and typed refusals.
• psionic-provider now projects that eval report into TassadarEffectTaxonomyReceipt.
• docs/TASSADAR_WASM_RUNBOOK.md, docs/ROADMAP_TASSADAR_INDEX.md, and docs/ROADMAP_TASSADAR_TAS_SYNC.md now index the widened effect lane.

Validation:

• cargo test -p psionic-runtime effect_taxonomy -- --nocapture
• cargo test -p psionic-sandbox effect_boundary -- --nocapture
• cargo test -p psionic-router effect_route_policy -- --nocapture
• cargo run -p psionic-router --example tassadar_effect_route_policy_report
• cargo test -p psionic-eval effect_taxonomy -- --nocapture
• cargo run -p psionic-eval --example tassadar_effect_taxonomy_report
• cargo check -p psionic-provider

Note: cargo test -p psionic-provider effect_taxonomy -- --nocapture hit a linker-only environment failure in this checkout after the crate compiled; the provider surface itself passes cargo check and the new receipt is exercised through the shared eval report path.