Skip to content

backlog.md

Latest commit

 

History

History
244 lines (228 loc) · 19.7 KB

backlog.md

File metadata and controls

244 lines (228 loc) · 19.7 KB
Raw

RL2 Backlog

Successor to the 2026-08-01 review/decision record (fix-fable.md, now deleted). Every design decision from that cycle is executed in the normative artifacts and traceable through the commit history 0ca83bd..6165f47; this file carries only remaining work and standing records. Working discipline: discuss → decide → sub-agent executes a closed task spec → coordinator checks diffs → commit. Gate: uv run tools/validate.py stays FAIL 0 and warning-free.

1. Executable conformance (highest priority — everything else's acceptance criterion)

  • Machine-readable vector format per spec/RL2_Compilation.md §7 interchange schemas: (module | universe, Request, WorldSnapshot, EvaluationConfiguration) → expected (decision, determining SourceRefs, Duty/Promise statuses, causes, diagnostics).
  • Publish the four JSON Schema files (Request, WorldSnapshot, EvaluationConfiguration, EvaluationResult) — field tables are normative in RL2_Compilation.md §7; the schema files ship here.
  • Convert the four conformance/vectors/*.md files and ~15 highest-value use cases to vectors.
  • Negative compiler vectors: at least one per diagnostic code in RL2_Compilation.md §6.1.
  • Missing positive vectors identified during review: complementary Privilege/Prohibition pair under a missing operand; inapplicable prerequisite Duty (vacuous pass vs status-condition False); permit + indeterminate resolution under each strategy; the four exception-pattern forms; profile version negotiation and digest mismatch; relative-window resolution incl. missing anchor; sentinel-population norm with missing population fact.
  • F-01 vectors (sentinel Duty binding, rl2:consequentDuty): positive — a Privilege's consequentDuty fires on grant and the envelope carries the bound obligate atom (occurrenceOf(d, Env)), never the sentinel-bearing template; bound prerequisite-gating positive and negative — a sentinel-carrying prerequisiteDuty template is gated via dutyStatus(bind(d, requestAgent, requestAsset)), both fulfilled and unmet; a consequent Duty with an Unknown condition of its own — confirms the decision outcome is unaffected by a consequent Duty's condition value (only the Privilege's own accessResult.truth = True gates firing); UnboundStatusTarget negative compile vector — an rl2:obligationStateOperand targeting a Duty whose rl2:subject/rl2:object is a sentinel.
  • R-01/R-02/R-03 vectors (status-map and occurrence model, spec/RL2_Semantics.md §Duty Template Binding and Eval composition): a template Duty (prerequisite, consequent, or independent) reached via the widened allDuties(P) receives no entry in dutyStatuses under its own identity — only concrete(d) Duties do; a bound occurrence produced by occurrenceOf at a consequent-duty or independent-clause site contributes a dutyStatuses entry keyed by its occurrenceId, disjoint from the concrete-Duty key space (boundOccurrenceStatuses). TWO-IDENTICAL-READS — same agent, same asset, two distinct Request.id values against the same template, with Evidence carrying dischargeOf naming only the first occurrence's occurrenceId — the first occurrence is Known(Fulfilled) and the second stays Known(Pending), confirming per-grant discharge does not leak across Request.id values. Per-member fallback — two grants of the same template to the same agent/asset with no Request.id on either request coalesce onto one occurrenceId, so evidence discharging the first grant also discharges the second. Consequent-duty trigger condition consumed — an occurrenceOf-produced occurrence carries condition := None (the Privilege's own accessResult already decided applicability at grant time), so a later dutyStatus re-evaluation of that recorded occurrence cannot re-decide firing eligibility, only fulfillment. Occurrence window resolved at grant time — an occurrence's window is resolveWindow's output computed once against the granting evaluation's own snapshot, so a later evaluation supplying a different snapshot does not re-anchor a relative window to the new snapshot's clock.
  • R-07 occurrence-identity interchange repair: occurrenceId begins with sourceIdentity(d), which may be a module-scoped SourceRef for a valid blank-node Duty, while Evidence.dischargeOf currently admits only an IRI in that position. Align the abstract type and canonical wire encoding (or require an authored IRI for a PER-GRANT template), then add a blank-node template plus correlated-evidence vector.
  • R-08 prerequisite status-map scope: boundOccurrenceStatuses must include a bound prerequisite only when its Privilege is in the request's prerequisite-gating domain (at least matchesRequest(n, Env)), or the result model must explicitly become an eager universe-wide report. Add an unmatched-Privilege vector that asserts no irrelevant BoundIdentity entry or diagnostic from its template.
  • conformance/vectors/README.md overclaims: these documents are evaluator-independent boundary expectations that specify selected parts of the input tuple, not machine-readable vectors, and they do not by themselves establish evaluator conformance. Correct the framing when the executable suite lands and the two artifact kinds need distinguishing.
  • F-05 materialization vectors — action Promise with postCondition crystallizes into an Achievement Duty carrying it (positive); state Promise with postCondition rejected by PromiseShape (negative).
  • Translation runner comparing normalized ODRL fixtures to expected modules.
  • Reference harness (tools/evaluate.py or similar); two independent evaluators before any interoperability claim.
  • tools/validate.py hardening: remove the ...-means-placeholder fence skip; make the injected prefix PREAMBLE opt-in (or add a strict second parse pass); parse the SHACL report graph rather than counting rendered severity strings; validate the shapes graph itself.
  • F-10 regime fixtures — imports ignored, named graphs ignored, relative-IRI rejection, warning-does-not-reject, canonical report-projection comparison across two SHACL engines.

2. ODRL migration profile ("ODRL ±" execution)

  • Author the migration-profile document: the supported ODRL 2.2 fragment, term-by-term compatibility matrix, each row carrying its three-test verdict (demand / determinism / runtime-independence) and a fixture.
  • Consequence mapping rule: odrl:remedy/consequence → Duty with applicability condition obligationStateOperand(targetNorm) = Violated (profile-dependent disposition; the machinery already exists in core).
  • New migration fixtures: odrl:xone; odrl:inheritFrom (multiple parents, missing parent, cycle); ConflictingCompactValue.
  • conformance/migration/attached-duty.md is stale against the committed mapping table: it is marked normalized and asserts that an attached odrl:duty becomes a prerequisite Duty, whereas RL2_ODRL_Mapping.md now maps odrl:duty to rl2:prerequisiteDuty or rl2:consequentDuty with disposition clarified. Restate the fixture as recording one explicitly declared prerequisite interpretation, and say that it does not establish that interpretation as ODRL's default. Same correction in conformance/migration/README.md.
  • Reclassifications: ODRL invalid conflict → clarified (whole-policy void vs request-scoped Indeterminate); Set-rule assigner preservation caveat; rightOperandReference → pre-resolve into an attributed snapshot fact or reject.
  • Refinement compile-down rule in the transpiler (component-scoped refinements → rule-level conditions; not an RL2 authoring form).
  • Crosswalk to the W3C ODRL Formal Semantics draft (Behaviour ↔ defaultDecision; state taxonomy; report model) and to XACML combining algorithms — expand mapping §8/§11.

3. Shipped profiles (the ODRL + layer)

  1. Remedies/Contracts — expression-only (decided): vocabulary + the §2 consequence shape; document the assembler's violation-evidence obligation in RL2_ExternalData.md. Violation-evidence vocabulary deferred until cross-party interop is live.
  2. Privacy/Consent — align with pod/Solid practice (Esteves et al. as reference); rewrite 4–6 data-governance use cases onto spec/profiles/rl2-privacy-profile.ttl (it currently has zero conformance coverage).
  3. AI Vocabulary — adopt the W3C draft actions; align no-ml-training.md / derived-data-restriction.md action names.
  4. Commercial/Market-Data — Quantity(Numeric, Unit) value type (same-unit comparison only, no conversion — decided); metering patterns via request.parameters; first rl2:ProfileOperator exercises.

4. Infrastructure

  • w3id.org registration PR for https://w3id.org/rl2# (namespace already swapped in-repo).
  • Regenerate/re-check docs/RL2_Vocabulary.md after each ontology change (standing sub-agent task-spec line: every batch touching rl2.ttl updates the Vocabulary doc).

5. Implementation track

  • Go reference evaluator + ODRL importer (reconciled estimate: 6–10 person-months to a credible reference implementation with executable conformance).
  • Lean 4 model of the kernel; differential random testing Go ↔ Lean (Cedar precedent, arXiv:2407.01688). Proof order: compile-soundness theorem (RL2_Compilation.md §8) → evaluator determinism/totality → verified RDF compiler last (largest trusted boundary). Mechanized proofs: +6–12 person-months, do not gate conformance.
  • Benchmarks before any scale claim: policy count, clauses/policy, condition depth, hierarchy size, snapshot facts, evidence, conflicting atoms; compare OVAL/OPA/Cedar on the shared authorization subset.

6. Deferred, with reopen triggers

Item Trigger to reopen
Fact-vs-fact comparison (right operand is authored; dynamic-right needs widened RuntimeReference or stays assembler doctrine) Commercial profile metering scenarios
Recurrence (bounded period+count, compile-time expansion to finite windows) A real recurring-obligation contract
Maintenance boundary algorithm (lift the P8 fact/clock-only invariant restriction via recursive boundary collection) An invariant genuinely needing status operands
Evidence-existence operand (event-triggered activation, Cimmino P4d) Migration demand
rl2:enforcementMode metadata (regimentation vs sanction, eval-ignored) PEP integration work
Policy containment / conflict detection tooling (SMT over the 3-valued condition language; authoring-time, no solver in Eval) Negotiation or change-impact requirements
Assembly contract → normative A deployment exercises assemble(), at which point the informative SelectedPolicyUniverse manifest (docs/RL2_ExternalData.md §6) also becomes normative
Port churn under a live agreement (product terms extending to later-created ports; amendment/lifecycle) A deployed data product creates ports under an existing Agreement
RDFC-1.0 source-RDF digests Cross-party identity over raw RDF re-ingestion
Violation-evidence vocabulary for the Remedies profile Cross-party remedy interop
Declarative operator DSL with a normative interpreter (machine-checkable profile-operator semantics, replacing prose denotations) A profile requires machine-verified operator semantics (e.g. the Lean-verified evaluator track)
Full per-operand data contract (value schema, cardinality, owner, provenance fields beyond Attribution, freshness semantics beyond maxAge, trust-policy digest, completeness scope) plus a normative, versioned RequiredInputs companion (F-13) Cross-party snapshot exchange or a second independent assembler implementation

7. Do not relitigate

Settled during the 2026-08-01 review cycle (rationale in commit messages and the spec):

  • rl2:priority never applied to Promises/Duties — its domain is Privilege ∪ Prohibition by design; crystallized Duties have nothing to inherit.
  • Set/Offer/Agreement stay distinct classes; operativity is structural (Promise ∉ Norm), not a mutable status property. Promise ≠ Duty deontically: directive-in-force vs commissive-contingent-on-acceptance, binding different parties.
  • Atomic vs Logical constraint classes stay separate (structural arity checking).
  • neq/isNoneOf are authoring preference over not(eq)/not(isAnyOf), not a projection rewrite: truth values agree, but causes differ (ComparisonSite embeds the operator).
  • Independent-Duty obligate atoms are action-local (derive only for requests matching the Duty's own action); an unmet prerequisite yields NotApplicable (no atom), not implicit Deny.
  • Never widen an rdfs:range (to owl:unionOf or otherwise); domains widen only via the in-file owl:unionOf precedent; structure sharing across classes goes through SHACL.
  • Full Boolean canonicalization rejected (exponential; Salas 2026); the narrowed one-encoding- per-primitive-relation claim plus enumerated local rules is the standing position.
  • The three-test rule (demand / determinism / runtime-independence) governs every future vocabulary question; odrl:implies stays rejected; includedIn stays core.
  • F-01 (sentinel Duty subject/object with no instance binding): resolved via new rl2:consequentDuty (Privilege → Duty, non-gating, the post-use/companion counterpart to rl2:prerequisiteDuty). The regression review (fix-codex.md R-01..R-03) correctly called this cycle's original "resolved" claim premature: a single bind(d, a, o) substitution at all four request-context sites left template Duties reachable by dutyStatus unbound (R-01), gave a bound Duty no per-grant occurrence identity (R-02), and gave Duty conditions incompatible request/status evaluation scopes (R-03). The completed model instead splits the four sites into two mechanisms: prerequisite gating and attached-duty reporting keep unchanged, pure bind(d, a, o) substitution, with a request.* path in these request-free contexts yielding an attributed Missing; consequentDuty emission and independent-clause atom emission use occurrenceOf(d, Env) instead, which additionally consumes the Duty's condition (evaluated once, in the request environment), resolves window to Absolute or None via resolveWindow at grant time (never an unresolved Relative window — a resolveWindow failure yields an indeterminate atom, not a broken occurrence), and assigns an occurrenceId — (sourceIdentity(d), Request.id) when the request carries one (PER-GRANT), else (sourceIdentity(d), agent, asset) (PER-MEMBER fallback, coalescing repeated grants). dutyStatuses (RL2_Model.md §6) is restricted to concrete(d) Duties under their own identity — a template contributes no entry there — plus a lazily-computed entry per bound occurrence, keyed by occurrenceId (from occurrenceOf) or by BoundIdentity (from a bound prerequisite at gating); the two key spaces are disjoint by construction (RL2_Semantics.md's boundOccurrenceStatuses). Evidence.dischargeOf (RL2_Model.md §4.3) correlates evidence to a specific PER-GRANT occurrence; PER-MEMBER occurrences and authored concrete Duties are unaffected. Envelope atoms and dutyStatus queries are always concrete and sentinel-free; a bound occurrence recorded in an EvaluationResult is concrete and re-evaluable, and re-identifiable by its recorded key. Sentinels remain legal in Duty templates but are SentinelMisuse on a Promise (a Promise crystallizes at materialization with no binding source). ODRL odrl:duty from a Permission is clarified, not normalized: a TranslationConfiguration must declare per profile/policy whether it means prerequisiteDuty or consequentDuty; an undeclared interpretation is MissingTranslationInterpretation.
  • F-03 (plug-in rl2:ProfileOperator semantics unconstrained): operators stay (D8), quarantined — the declaring profile must ship a total-function denotation (same register as core apply) plus positive/negative conformance vectors, or the profile is non-conforming; STRICT core conformance (spec/RL2_Compilation.md §8, §10) excludes profile operators entirely; a request-independent profile judgment should be precomputed as a WorldSnapshot fact rather than declared as an operator.
  • F-10 (validation regime pinned, spec/RL2_Compilation.md §10.1): single merged default graph as phase (1)/projection input, named graphs ignored, owl:imports not followed, no network/file fetching beyond supplied inputs; base IRI is an explicit caller-supplied compile input, a remaining relative IRI rejects; NO RDFS/OWL entailment before or during validation; SHACL Core + SHACL-SPARQL only, Advanced Features (rules, functions) inactive; sh:Violation rejects, sh:Warning never rejects but must be reported; canonical report projection is (source shape IRI, focus node, result path, severity), deduplicated and sorted, messages excluded.
  • F-06/F-07 (Offer bundling / output-port boundary): no materializeBundle and no overloading rl2:Offer/requiresProfile to carry a bundle of tagged Set policies — materialize(Offer, Acceptance) stays single-Offer. Concrete-at-acceptance data products (Level
    1. are expressed with one Promise per named output port plus a product-wide prerequisite attestation Duty (see conformance/usecases/data-product-offer.md). Per-request policy-universe selection (e.g. tag-based, per port/asset) is the assembler's job, kept replayable via the informative SelectedPolicyUniverse manifest (docs/RL2_ExternalData.md §6), not a kernel concept.
  • F-13 (external-data admissibility language unpinned): trust verification (signatures, provenance chains, trust anchors, connector authentication) is pre-Eval, performed by the trusted assembler before WorldSnapshot exists; Eval never does cryptographic or chain-of-trust verification. EvaluationConfiguration carries admissibility as a closed, finite three-kind record (allowedSources, maxAge per left-operand resolution path; evidenceSigners per Duty-evidence scope) — not a predicate language, not profile-extensible — with a canonical JSON form covered by the configuration echo/digest in EvaluationResult. A fact or evidence item failing the filter is treated exactly as absent, yielding the ordinary attributed Missing (never a distinct error kind, never a silent skip). The full per-operand data contract and a normative RequiredInputs companion stay deferred (§6).
  • R-04/R-05 (admissibility completion and closure complexity, fix-codex.md): evidenceSigners is keyed by the owning Duty's own stable identity (sourceIdentity(d)), not by EvidenceSelector shape, so a bound occurrence inherits its template's entry unchanged regardless of who requested it. A configured constraint (allowedSources, maxAge, or evidenceSigners) whose candidate lacks the attribution field it reads fails that filter, deterministically and conservatively in both directions. The "visible in causes" claim for admissibility filtering is qualified, not retracted: a filtered fact surfaces as an ordinary attributed Missing; filtered-to-empty evidence is a definite status (typically Pending) with no distinct cause of its own; an over-restrictive evidenceSigners entry is visible only by comparing two evaluations' configuration digests side by side, never by inspecting one evaluation's causes. Separately, the Complexity Analysis section's closure-construction claim (O(n_hier) space and O(1) membership for the same representation) was not simultaneously achievable and is corrected: n_closure (reachable pairs, up to V²) names the materialized closure's actual size, and the section now states the materialized-vs-compact space/lookup-cost trade-off explicitly instead of asserting both sides of it at once.