Implementation State

September 20, 2026 · View on GitHub

Continuation entrypoint for a new AI: read NEXT-AI-HANDOFF.md before changing code. It contains the current continuation point, explicit next actions, no-rebuild boundaries, browser/relay production state, v4 live evidence, and M19 no-fabrication rules.

spec_version: 0.4-master-implementation-research-expanded
spec_digest_sha256: 9b8bc907fa0da89d4b7ea2e0be886919deffdb35e398ea7897ea77d305380f5b
last_completed_gate: M20-v1-stable-core-conformance
active_milestone: M19-natural-conversation-research
stable_packages:
  - core-types
  - provenance
candidate_packages: []
experimental_packages:
  - open-world-values
  - ontology
  - semantic-graph
  - semantic-validator
  - decision-runtime
  - decision-packs
  - verifier-core
  - discourse-ir
  - formal-ir
  - action-ir
  - universal-expression
  - trace-replay
  - extension-core
partial_vertical_slices:
  - controlled-English-requirement-roundtrip
  - recorded-Jev-reference-choice
  - controlled-Vietnamese-shared-JSG
  - PIR-typed-hole-to-TypeScript
  - formal-action-universal-contract
  - m16-controlled-multitarget
  - m17-verification-hardening
  - m18-open-world-lexicon-code-switch
  - m4-controlled-parser-foundation
  - m5-controlled-realizer-roundtrip
  - m6-discourse-naturalness-foundation
  - m7-dialogue-semantics
  - m8-vietnamese-language-pack
  - m10-program-ir
  - m11-synthesis-core
  - m12-typescript-backend
  - m13-python-backend
  - m14-compiler-test-repair-loop
  - m16-universal-expression-api
  - t268-t275-extension-framework
  - t301-t310-graph-topology
  - t311-t320-scope-quantification
  - t331-t340-deixis-attitudes-evidence
  - t341-t350-presupposition-pragmatics
  - t351-t360-comparison-quantity-space
  - t361-t370-questions-dialogue-acts
  - t371-t380-reference-ellipsis
  - t381-t390-discourse-information-structure
  - t391-t400-lexicon-open-vocabulary
  - t401-t410-language-typology
  - t411-t420-translation-parser-architecture
  - t421-t430-incremental-realization
known_failures: []
blocked_items: []
next_tasks:
  - absorb frozen conversation-ranker boundary v4 evidence into NC-11 calibration analysis without post-hoc relabeling
  - execute and freeze real M19 A/B stimuli through the no-fabrication observed-runner path for NC-13
  - collect the preregistered real blinded human ratings using pseudonymous evaluator worksheets for NC-14
  - freeze observed M19 ratings/failures/report with canonical evidence digests and report the result even if negative or mixed
  - drive NC-15 only from observed failure clusters and regression evidence
  - keep non-stable packages explicitly below stable until their own v1 evidence exists
last_verified_main_commit: ca2a22c12e6a948fe2c688bbe0c62d772bd1c87c
last_merged_playground_transport_commit: 368be40a34dda3c2fcfaae33514cdc30edaebec1
last_authenticated_playground_relay_smoke_commit: 553a373bbf99aa2ae7068a9d0e9ecb0536492271
last_verified_pr_head: 6a07cd47a6ac5b00cbccde94646a5bbf68a71c6b

Current state

M0–M9 are verified at their milestone gates. M3 includes the required one-request live Jev acceptance run; M4–M9 use deterministic/recorded language-engine evidence. M9 passed CI #136 with 36/36 test files and 280/280 tests. M10 Program IR and M11 Synthesis Core are verified. M12 TypeScript Backend passed CI #155 with 39/39 test files and 312/312 tests. M13 Python Backend passed CI #167 with 40/40 test files and 319/319 tests. M14 Compiler/Test Repair Loop passed CI #182 with 41/41 test files and 325/325 tests. M15 Formal/Data/Action IR passed CI #191 with 42/42 test files and 334/334 tests. M16 Universal Expression API is verified. M17 Verification Hardening passed CI #202 and is verified. M18.1 is verified, and M18.2 grammar/parser expansion passed CI #218 with 45/45 test files and 368/368 tests before merge commit d6d7c19; M18 Open-world Language Expansion remains the active milestone. The cross-cutting T268-T275 extension foundation passed CI #235 with 47/47 test files and 386/386 tests; package boundaries and strict TypeScript typecheck also passed, with zero live Jev requests. T276-T288 evaluation foundation passed CI #246 with 48/48 test files and 404/404 tests after a real T286 repair-loop benchmark exposed and fixed an inline-return patch-boundary bug; merged main then passed CI #247. T289-T296 hardening passed CI with 51/51 test files and 413/413 tests and merged to main. T297-T300 performance/cache/cancellation/crash/release hardening then passed CI with 52/52 test files and 417/417 tests; merged main commit 153986e passed the same deterministic gate. Both waves used zero live Jev requests. T301-T310 graph topology passed CI with 54/54 test files and 427/427 tests, with boundaries/typecheck green and zero live Jev requests. T311-T320 scope/quantification and T321-T330 event/time/modality are verified on main. T331-T340 deixis/attitudes/evidence passed deterministic CI with 57/57 test files and 453/453 tests and merged to main commit 852fdaff; merged main also passed. T341-T350 presupposition/pragmatics passed deterministic CI with 58/58 test files and 461/461 tests and merged as 1472e87; merged main also passed its deterministic gate. T351-T360 comparison/quantity/space passed deterministic CI with 59/59 test files and 471/471 tests and merged. T361-T370 questions/dialogue acts passed with 60/60 files and 477/477 tests and merged. T371-T380 reference/ellipsis passed with 61/61 files and 484/484 tests and merged. T381-T390 discourse/information structure passed with 62/62 files and 493/493 tests and merged. T391-T400 lexicon/open vocabulary passed with 63/63 files and 503/503 tests, merged as cdebadc0, and merged main passed. T401-T410 language typology passed with 64/64 files and 513/513 tests and merged as 89c6b894; its main push gate is pending while the next wave is built. T411-T420 translation/parser architecture passed deterministic CI with 65/65 test files and 522/522 tests and merged as f05a5d3b; merged-main verification is pending while the next wave is built. T421-T430 incremental parsing/realization passed deterministic CI with 66/66 test files and 532/532 tests, merged as a4edee2c, and merged-main CI passed. T431-T440 language-pack conformance passed deterministic CI with 67/67 test files and 542/542 tests and is merged. T441-T450 synthesis grammar passed deterministic CI with 68/68 test files and 552/552 tests and is merged.

T441-T450 synthesis grammar foundation passed deterministic CI with 68/68 test files and 552/552 tests and is verified on main.

T451-T460 advanced PIR passed deterministic CI with 69/69 test files and 562/562 tests and is merged. T461-T470 CEGIS passed deterministic CI with 70/70 test files and 573/573 tests and is verified on main; its acceptance case preserves the failed first candidate's counterexample and verifies the corrected next iteration.

T471-T480 solver/proof passed deterministic CI with 71/71 test files and 583/583 tests and is merged. T481-T490 source-preserving transformation passed deterministic CI with 72/72 test files and 593/593 tests and is merged. T491-T500 formal/action/evidence closure passed deterministic CI with 73/73 test files and 603/603 tests and is merged. Extended M11-M20 evidence gates passed deterministic CI with 74/74 test files and 608/608 tests; merged main commit 7110ae9c passed the same full gate. The §§963-980 release-gate closure passed deterministic CI #331 with 75/75 test files and 611/611 tests, with package boundaries and strict TypeScript typecheck green. M19 natural-conversation protocol passed rebased deterministic CI #336 with 76/76 test files and 617/617 tests, with package boundaries and strict TypeScript typecheck green. The protocol is verified, but the original M19 research milestone remains pending-human-data until actual blinded evaluator ratings are imported. The M19 human-study execution kit now provides evaluator worksheets, strict pseudonymous rating import, and deterministic evidence freezing; this improves collection integrity but does not substitute for observed ratings. It includes executable evidence aggregation, machine-readable phenomenon coverage, an unsupported-case ledger, candidate-recall accounting, search/solver honesty, source-preservation/replay/extension checks, and anti-template/anti-hidden-generator/anti-special-case audits. Original M19 natural-conversation human evaluation and M20 v1.0 conformance remain open and are not conflated with the extended M19/M20 evidence gates.

M20 v1 stable-core conformance for the non-vacuous stable set core-types + provenance is merged on main commit 3dcc2269 and passed main CI #344 with 77/77 test files and 624/624 tests; package boundaries and strict typecheck were green. Public API docs, explicit ABI/schema versions, conformance suite, migration policy, correctness/replay baseline, machine-readable limitations, executable replay demo, and zero-generative audit are present. The package manifest promotes only these dependency-safe packages to stable. M19 human-data remains independently pending and is not satisfied by M20 artifacts.

The experimental GitHub Pages Playground is being built as a public BYOK test surface. It is explicitly constrained to current Jev-shaped interactions (typed yes/no, bounded choices, and coverage checks), keeps keys in tab memory only, and does not upgrade the engine's free-form generation claim. Pointer-reactive NUI visuals, reduced-motion/high-contrast alternatives, browser-runtime security checks, and deterministic conformance tests are part of the same change.

The GitHub Pages Playground is deployed at https://nolane-x.github.io/JEV-language/. The NUI inspection-light interaction, TypeSafe wire-response validation, pointer alignment fix, literal-markup regression check, and accurate network/CORS diagnostics are merged. A browser-equivalent preflight probe on 2026-09-20 proved that direct TypeSafe calls from https://nolane-x.github.io currently lack Access-Control-Allow-Origin and are therefore blocked by provider CORS. The locked-down self-hosted relay transport was merged by PR #70, then deployed through the secret-backed Cloudflare workflow. The deployment path is therefore closed as an engineering blocker; only authenticated BYOK use remains a live credential-bound check. Deployment evidence jev-language-relay-deployment/v2 records verified: true, source SHA 7a0e9d8491534f0f817074c712d82c1ecd178c6d, relay URL https://jev-language-typesafe-relay.nolane-file.workers.dev, successful deploy, successful health verification, and successful GitHub Pages CORS preflight. The production-closure revision makes this verified relay the browser default while retaining Direct TypeSafe as a fallback. The authenticated relay path is now live-verified. A one-request smoke on 2026-09-20 used the existing TYPESAFE_API_KEY GitHub secret without exposing or persisting it, traversed https://jev-language-typesafe-relay.nolane-file.workers.dev, required X-JEV-Relay: 1, verified the GitHub Pages CORS origin, and returned a valid jev-1.13.0 Noul response. Sanitized evidence is frozen in docs/evidence/PLAYGROUND-RELAY-LIVE-SMOKE.json at commit 553a373: exactly 1 request, 372 input tokens, 20 output tokens, credential_persisted=false.\n\n## Implemented foundation

  • Strict TypeScript / Node 20+ baseline.

  • Machine-readable package maturity/dependency DAG plus source-import boundary enforcement.

  • Structured Result/Error, semantic IDs, versions, trust/sensitivity labels, canonical JSON and SHA-256 digest utilities.

  • Provenance records/store.

  • Open-world source spans, normalization maps, stale-span detection, numeric/quantity/temporal/path/URL/email/filename literals, opaque integrity validation, sensitivity/redaction, and secret-safe state projection.

  • Ontology namespace/store, expanded reusable core primitive inventory, provisional concepts, compositional concepts, domain/range traversal, transactional merge, and deprecation/alias handling.

  • JSG typed node/value algebra, atomic transactions, revisions, canonical serialization/deserialization, restore, semantic diff/query, and the verified V0–V8 staged semantic-validation pipeline with ontology domain/range/cardinality, scope/binding, invariant, provenance/trust, stable-diagnostic and profile-validation hooks.

  • Recorded and TypeSafe JDR adapters, shared provider-response runtime validation, normalized typed answers, bounded transport retry policy, request/token budgets, cache, calibration hooks, structured errors and trace accounting.

  • Decision-pack registry/lifecycle contract, candidate-source provenance and recall evidence, deterministic decision-DAG batching, calibration profile registry, abstention policy, and evidence-based candidate/production maturity gates.

  • JDR request/token budgets, deterministic trace events, cache accounting, and calibration hooks.

  • Ontology transactional batch merge, parent-cycle rejection, ancestry queries, deprecation/replacement resolution, and replacement-cycle rejection.

  • Generic VerificationObligation/Verifier ABI, evidence grading, deterministic-verifier precedence, and critical semantic-preservation checks including role bindings and temporal values.

  • TypeScript 7 CLI compatibility with the official TypeScript 6 programmatic compiler API bridge for embedded compile checks.

  • First four narrow bootstrap vertical slices: controlled English, recorded reference choice, controlled Vietnamese, and PIR typed-hole → TypeScript.

  • M4 controlled grounding/parser foundation with reversible normalization, packed syntax forests, bounded recorded-JDR ambiguity choice, JSG commit, and corpus coverage for event/negation/quantity/time/condition/cause/requirement/permission/prohibition/comparison/question.

  • M5 constrained English realization with discourse/clause plans, lexical/morphology planning primitives, semantic source maps, attribution-safe realization, fallback policy, and a 100% semantic round-trip target on the current 11-fixture controlled corpus.

  • M6 verified bidirectional English expansion: semantic discourse-relation planning, safety-gated aggregation, explicit paraphrase lattices, repetition/style/audience planning, collocation scoring, bounded pragmatic Decision Packs, human-eval export, verified-by-round-trip synonyms/active-passive/temporal/condition/cause/reported-speech variants, relative clauses, pronoun-linked multi-sentence discourse, exact unknown-name preservation, and a passing 14-sample held-out template-leakage benchmark.\n- M7 verified dialogue semantics: revisioned transactional state, topic stack, salience/reference candidates, questions, requests, commitments, correction/retraction history, ellipsis/follow-up reconstruction, bounded reference Decision Pack, semantics-preserving compaction, and a passing 22-turn long-reference acceptance fixture.

  • M8 verified Vietnamese language pack: direct Vietnamese↔JSG parsing/realization, bilingual semantic-equivalence corpus, language-specific classifier/aspect/address behavior, and the shared Section-251 HumanLanguagePack ABI for English/Vietnamese.

  • M9 verified multilingual semantic gate: independent English/Vietnamese G→surface→G round trips, resolved-reference and instruction-as-content JSG structures, and dimension-level diagnostics for predicate/roles/polarity/modality/quantity/time/condition/causality/attribution/reference/instruction content.

  • M10 verified Program IR: backend-neutral graph/model with modules, symbols, expanded types/expressions/statements, contracts/effects, typed holes, source bindings, atomic transactions, deterministic validation/serialization, derived CFG/def-use analysis and all eight required corpus fixtures.

  • M11 verified synthesis core: typed expression/statement holes, six generator families, deterministic hard pruning, best-first/beam frontiers, state deduplication, bounded search budgets, bounded recorded-Jev ranking, acceptance verifiers, structured partial failures and traceable multi-step composition.

  • M12 verified TypeScript backend: backend ABI/manifest, compiler-API parser, AST↔PIR subset, source bindings, AST/source printer, minimal patching, normalized diagnostics, strict typecheck adapter and conformance over the M12 source subset plus the verified M10 PIR corpus.\n- M13 verified Python backend: stdlib AST parse/unparse, optional annotations with dynamic unknowns, AST↔PIR subset, mapping-safe record lowering, source bindings/patches, normalized compile diagnostics, all-eight M10 lower/compile proof and TypeScript/Python cross-backend PIR semantic fixtures.

  • M14 verified compiler/test repair loop: stable repair diagnostics, implicated-node location, bounded guard/import/argument/return repair generators, bounded candidate ranking, compile/test/regression verification, rollback, hard budgets and executable broken-code acceptance.

  • M15 verified formal/data/action IR: stable Data/Schema/Query/Math/Logic/Command/Action schemas, deterministic validators, canonical renderers, read-vs-mutation query safety, declared-capability Action IR validation and seven-family conformance.

  • Graph-structured Discourse IR foundation with deterministic prerequisite-aware ordering.

  • Formal IR family foundation (Data/Schema/Query/Math/Logic/Command), capability-validated Action IR, and harness-neutral Universal Expression contract.

  • Registry-driven Universal Expression runtime plus trace DAG/config-digest/replay-manifest foundation.

  • M16 verified Universal Expression API: harness-neutral parse/realize/express/transform/verify runtime, capability discovery, same-JSG natural-language/structured-data/PIR-program/declared-Action realization, and trace/replay evidence.

  • M17 verifier foundation: semantic round-trip with parser-limit unknowns, conservative trust/provenance checks, requirement reports, grammar evidence normalization, and compile/test evidence normalization.\n- M17 verified verification hardening: registry orchestration, provenance ancestry integrity, deterministic conflict handling, evidence-floor grading, and replay bundles bound to raw/authoritative result digests.

  • M18.1 verified: Section-209 open-world lexical resolver, EN/VI code-switch evidence, provisional lexical-sense proposals, exact unknown-term preservation and explicit borrowing policy. T098-T106 passed CI #207.

  • M18.2 verified: typed grammar features/categories, GrammarRule-driven packed chart parser, expanded English controlled syntax T110-T119, lexical/morphology grammar evidence and conservative syntax-to-JSG bridging. T107-T120 passed CI #218 with 45/45 test files and 368/368 tests; live Jev requests: 0.\n- M1 validator completion verified: T026-T035 now provide V0–V8 staged validation, stable diagnostic registry, invariant registration/bundle, ontology domain/range/cardinality checks, scope/binding validation and provenance/trust closure. CI #226 passed 46/46 test files and 378/378 tests; live Jev requests: 0.

  • T268-T275 extension foundation verified: runtime-validated extension manifests, duplicate-safe dependency registry, spec-compatible version ranges for engine/semantic-schema/ontology-core/PIR, language/backend/verifier conformance runners, transactional ontology/domain-pack loading, and capability-minimal isolation admission. CI #235 passed 47/47 test files and 386/386 tests; live Jev requests: 0.\n- T276-T288 evaluation foundation verified: reproducible dataset manifests, deterministic benchmark runner, candidate-recall/calibration reporters, semantic/NLU/NLG/dialogue/multilingual/synthesis/repair benchmark reports, zero-generative audit, and replay-manifest generation. CI #246 passed 48/48 test files and 404/404 tests; the T286 executable repair benchmark also produced a regression fix for one-line return-expression patch boundaries. Main merge CI #247 passed.\n- T289-T296 hardening verified: deterministic fuzz-smoke coverage for JSG deserialization, normalization/spans, syntax forests and patch application; canonicalization/transaction properties; adversarial trust escalation and credential redaction tests. CI passed 51/51 test files and 413/413 tests; live Jev requests: 0.\n- T297-T300 bootstrap release hardening verified: deterministic performance recorder with percentile summaries, cache key/copy-isolation/stale-request correctness, cancellation and crash normalization coverage, and an evidence-bound release conformance report that blocks unverified tasks, red CI, incomplete tests, generative violations, known failures, or blocked items. CI passed 52/52 test files and 417/417 tests; merged main commit 153986e passed the same deterministic suite; live Jev requests: 0.

  • T301-T310 graph topology verified: canonical graph-view/projector APIs, reentrant shared-node analysis, explicit MentionNode/entity separation and mention indexes, provisional disconnected graph fragments with deterministic merge diagnostics, explicit cycle-permission registry/enforcement, and bounded topology property tests. CI passed 54/54 test files and 427/427 tests with package boundaries and strict typecheck green; live Jev requests: 0.

  • T311-T320 scope/quantification verified: explicit ScopeNode/ScopeConstraintNode, unresolved relative scope, generalized QuantifierNode/cardinality/distributivity semantics, explicit negation scope, staged scope validation, evidence-tier resolution pipeline, ambiguity-preserving generation filtering, and an adversarial executable benchmark. CI passed 55/55 test files and 436/436 tests with boundaries/typecheck green; merged main commit 304c2ca passed deterministic CI; live Jev requests: 0.

  • T321-T330 event/time/modality/conditionals verified: refined event/process/transition/achievement/activity ontology, explicit event-token versus event-class semantics, temporal object registry with interval-relation constraints, language-pack tense/time separation and aspect mapping ABI, dimensioned/ordinal modality with separate calibrated probability, conditional variants and counterfactual metadata, runtime boundary validation, semantic-preservation checks, and a deterministic adversarial benchmark. After correcting a round-trip test that incorrectly assumed insertion-order preservation instead of canonical node ordering, CI passed 56/56 test files and 445/445 tests with package boundaries and strict typecheck green; live Jev requests: 0.

Natural-conversation research wave

Latest continuation note (2026-09-20): conversational live Jev evidence now totals 70 requests: 12 multilingual baseline + 18 stress v2 + 20 mirrored v3 + 20 boundary-calibration v4. v4 produced 10/10 clear-preference agreement, 9/10 near-tie gate abstention, 0 near-tie false-certainty rate, and 8/10 mirrored order-invariant base pairs. These observations advance NC-11 but do not replace M19 human evidence. See docs/evidence/CONVERSATION-RANKER-BOUNDARY-v4.json and docs/NEXT-AI-HANDOFF.md.

The evidence-bound conversational layer now includes the bounded conversation candidate ABI, typed Response Semantic Plan, diverse verified candidate lattice, Jev conversational surface-ranker Decision Pack, bounded anti-template style memory, deterministic semantic-certification bridge, EN/VI conversational microgrammars, candidate-recall reporting, and an end-to-end bounded selection pipeline with confidence/margin abstention.

Live Jev evidence through the verified relay using jev-1.13.0 now covers 50 authenticated ranking/judgment requests across three frozen sets: multilingual baseline 12/12 expected judgments, conversational stress v2 18/18, and mirrored subtle-pair v3 20/20 agreement with preregistered linguistic hypotheses. In v3 all 10 base pairs were order-invariant after A/B reversal, option-A selection rate was exactly 0.5, mean confidence and mean probability margin were both 0.916, minimum margin was 0.4, and confidence=1 saturation fell to 0.4. These are bounded-ranker observations, not unrestricted generation or human-naturalness proof.

Applying the current selection defaults (confidence 0.62, margin 0.08) to v3 accepts 19/20 cases and abstains on ja-workplace-register:mirrored. Because v3 contains zero observed ranking errors relative to its preregistered hypotheses, threshold calibration is explicitly marked insufficient-errors; the repository does not lower production thresholds merely to recover one correct abstention.

Candidate generation now has a frozen EN/VI engineering recall baseline. Wave 1 deliberately records preferred-surface recall 0.8 with two known gaps. Wave 2 closes those two targeted gaps using bounded, opt-in Vietnamese topic-comment reshaping and epistemically gated English hedging while preserving the historical wave-1 result and negative guards. These remain engineering surfaces rather than human preference evidence.

The major remaining natural-conversation research gates are real M19 blinded study stimuli, preregistered human ratings, and continued error-driven expansion on held-out failures. M19 therefore remains partial by design.

Still partial by design

The master specification is much broader than the bootstrap wave. Broad NLU/NLG, full dialogue/pragmatics, full Vietnamese grammar, repair beyond the verified bounded M14 families, richer formal backends, broader M16 multi-target coverage beyond the controlled semantic subset and concrete program adapters, broader verifier orchestration/provenance edge cases, full conformance matrix and research-expansion tasks remain open. Package stubs say NOT_IMPLEMENTED or PARTIAL rather than pretending they exist.

Quota policy

Live Jev calls are not part of push/PR CI. M3 used exactly one explicitly isolated live smoke request: model jev-1.13.0, 356 input tokens, 26 output tokens, SDK retry 0. The temporary trigger was removed after the successful run, and ordinary deterministic development remains at zero live requests.

Completion semantics

implemented means source exists. verified means the relevant automated gate passed. Recorded Jev behavior is never reported as live success, and a narrow vertical slice is never reported as broad capability coverage.

  • T331-T340 deixis/attitudes/evidence verified: serializable Deictic/Quotation/Attitude ContextNode frames, explicit person/spatial/temporal/discourse/social deictic references, deterministic deictic resolution, nested quotation context stack, direct-vs-indirect speech exactness rules, propositional-attitude isolation from global assertions, first-class evidential source modes kept separate from node confidence and provenance, context-aware graph topology/cycle permissions, semantic-preservation checks, nested-attribution tests and a deterministic deictic-shift/quotation benchmark. CI passed 57/57 test files and 453/453 tests with package boundaries and strict typecheck green; live Jev requests: 0.