Roadmap

Note: This roadmap describes intended directions, not commitments. Priorities will shift based on user feedback and real-world testing.

For the underlying readiness framework (6 dimensions, gate criteria, branching forms), see docs/PLATFORM_READINESS.md.

v0.1.0 — Foundation (released, foundational)

Status: Released (2026, initial alpha)

Done

Hybrid Search (Vector + BM25 + keyword)
Graph-RAG with ontology (12 relation types)
3-stage security model (RBAC + ABAC + Instruction Isolation)
JWT auth + rate limiting + audit log
Self-evolution scaffolding (Patch Pipeline 4-Gate)
Multi-LLM routing (Ollama-based)
Multimodal hooks (image / video / audio)
Web search (Tavily + DuckDuckGo)
11-trait personality system
Knowledge tracker (8 abilities + 6 domains)
Internationalization (English + Korean UI)

Known Gaps (to address in v0.2)

Real-data validation pending
Multimodal pipeline integration limited
Self-evolution untested at scale

v0.2.0 — Foundation Hardening (released 2026-05-08, closed 2026-05-13)

Theme: Make the v0.1 capabilities trustworthy enough to recommend to a second user. Six axes, all of them P0/P1. All six axes complete. Axis 6 (second-user gate) closed on 2026-05-13 with the user secured — v0.2 → v0.3 gate clear, project formally entering v0.3 Platform Skeleton.

Axis 1 — Architecture Separation (P0) ✅

Goal: no single file owns more than one responsibility.

Split core/reasoning_engine.py into core/reasoning/ package: engine.py (orchestration), pipeline.py (loop), modes.py (mode dispatch). PRs #37 / #38 / #39.
Consolidate memory_* into core/memory/ package with documented public API. PR #35.
Public typed interfaces for: Retriever, GraphEngine, PolicyEngine, Reasoner, OutputFilter. PR #50.
tools/self/ sub-process boundary — deferred to v0.3 (no observed need at v0.2 scale; in-proc access is contained).

Done when: ✅ import graph is acyclic and each module has < 20 KB.

Axis 2 — Evaluation Harness (P0) ✅

Goal: every change is measured against the same yardsticks.

STEP 7 13-query suite locked as committed regression baseline (eval/regression/step7_*.json, scripts/bench.py). PR #52.
RAGAS integrated for retrieval / faithfulness / answer relevance. Live /query/ integration. PRs #51 / #64 / #66.
scripts/bench.py with --check / --update-baseline. PR #52.
PR-contract: core/{retrieval,graph,reasoning} PRs paste bench numbers. CLAUDE.md rule 2 + CONTRIBUTING. PR #43.
LegalBench subset — intentionally deferred (domain-coupled; contradicts the "no parallel domains" mother-platform rule until v1.0).

Done when: ✅ a PR cannot land without bench numbers attached.

Axis 3 — Observability / Tracing (P1) ✅

Goal: any answer can be debugged without re-running it.

trace_id ContextVar end-to-end + structured stage logs (auth → retrieve → rerank → graph → tool → answer → complete). PR #67.
JAMES_TRACE_STDOUT console mirror (default ON for the single-user operator workflow). PRs #71 / #75.
GET /admin/trace/{trace_id} full pipeline replay. PR #82.
Per-stage p50 / p90 / p99 / max latency histograms in GET /admin/metrics?window_hours=24. PR #83.
7-day auto-prune (JAMES_TRACE_RETENTION_DAYS env, default 7). PR #84.

Done when: ✅ a hallucination report can be diagnosed by trace_id alone.

Axis 4 — Security Boundary (P1) ✅

Goal: policy is a layer, not a sprinkle.

core/policy_engine.py — single point of role/sensitivity decisions. Wired into retrieval / graph / output / tools. 10 PRs (#50, #53, #54, #56, #57, #58, #59, #60, #61, #63).
Capability tokens for tool access (no direct fs path strings). PRs #57 / #58 / #59.
Multimodal inputs flagged + quarantined before joining LLM context. PRs #60 / #61 / #63.
Risky-coding hard-refuse policy at pre_check. PR #70.
External red-team pass on prompt injection — deferred to v0.4 (per ROADMAP — needs an external partner, not internal work).

Done when: ✅ removing the policy engine breaks 6+ modules.

Axis 5 — Controlled Evolution (P1) ✅

Goal: self-evolution cannot deploy without a human.

Opt-in env flag JAMES_ENABLE_EVOLUTION=0 (default off) + JAMES_AUTO_APPROVE safety check (refuses to start without JAMES_DEV_MODE). PR #69.
feedback → candidate → eval → human approval → deploy → rollback pipeline.
Eval gate via scripts/bench.py --check subprocess on every /admin/patch/approve deploy. PR #77.
Audit log records approver_username / approver_role / approved_at / approval_method / before_metrics / after_metrics. Lifecycle JSONL james_patch_log.jsonl. PRs #69 / #77.
Auto-rollback on bench regression + lifecycle log records the ROLLED_BACK event. Tested for byte-identical recovery under simulated mid-write crash. PR #78.
GET /admin/patch/audit?since=&approver=&outcome=&limit= operator-facing query endpoint. PR #79.

Done when: ✅ any deployed patch has an approver_username field in the audit DB, and deploy without it is rejected.

Axis 6 — Real-Data Validation (carries forward from v0.1) ✅

Goal: numbers from real data, not just synthetic.

Wiki corpus to 161 entities (concept 62 / org 57 / person 11 / document 31, hard-deduped via PR #28).
STEP 7 13-query suite includes negative / dedup / lang-mix / security / meta categories.
Multimodal pipeline integration (image / video / audio, OCR-poison quarantine). PRs #60 / #61 / #63.
Edge case discovery: #5 / #6 / #7 / #8 / #11 / #14 / #20 all closed via real-data feedback loops.
Second-user end-to-end bench run: secured (2026-05-13). v0.2 → v0.3 gate clear.

Known cuts from earlier v0.2 plan

The following moved to v0.3 to keep v0.2 focused:

Self-evolution end-to-end demonstration → folded into Axis 5
Performance profiling → after Axis 1 (premature otherwise)
Tutorial documentation → after Axis 1 stabilizes

Deferred follow-ups (recheck before v0.3)

Change Request primitive — generalise the approver_username pattern. ✅ v0.2.x landed (CR-A~D); CR-E moved to v0.3. v0.1 hard-coded approver tracking for self-evolution alone (Axis 5). Every other write — wiki edits, workspace job runs, ontology patches, config saves — landed with no proposal, no review, no diff, no rollback. v0.2.x shipped core/change_request.py
- core/change_request_apply.py + 6 endpoints + workspace UI + two target types (wiki_entity PR #243, run_jobs PR #240). Spans Axis 4 (security boundary generalisation) + Axis 5 (controlled write authorisation generalised beyond self-evolution).
- Trust zone documented in docs/ARCHITECTURE.md §5.6.
- Cycle plan + frozen schema + invariants in docs/handovers/v0.2.x-cr-track.md.
- Deliberately closed enum for target_type — the registration API surface is the v0.3 plugin contract.
- Multi-approver, team / project / department, external notification routing → all deferred to v0.3.
- Self-evolution gate wrap (CR-E) descoped from v0.2.x (4 locked JSONL-shape test files + eval-gate + rollback chain too large for the cycle without regression risk). Moved to v0.3 deliverables — see v0.3 "Change Request — finish the primitive".
- Done-when (full): every audit row carrying an approver_username today goes through core/change_request.py — achieved for wiki_entity / run_jobs; self-evolution approvers still bypass until CR-E.
Audit Phase 4b-2 — remove 16 JSONL writer sites. The JSONL → SQLite migration completed every reader path (PRs #206 / #207 / #208 / #210 / #211): the legacy /admin/audit endpoint, /admin/dashboard, /code/surface/ all query audit_log directly, and the new /admin/audit/list categories (tools / attack / system) surface the mirrored rows in the admin UI. JSONL writers remain in place as a belt-and-suspenders safety net because core/audit_bridge.mirror_to_audit_db is best-effort (try/except), so a silent SQLite write failure would otherwise be unrecoverable.
- Re-entry criteria: 2–4 weeks of production use with no SQLite mirror gaps observed (compare audit_log row count vs JSONL line count weekly), then drop the writers.
- Alternative if mirror reliability is uncertain: gate JSONL writers behind an env var (JAMES_AUDIT_JSONL=1, default off) before deletion.
- Touch points: core/security_layer.log_attack + log_system_event; 9 module-local _log_system copies (llm/router, gemma_client, graph_engine, query_expander, orchestrator, retrieval_engine, memory/loom, memory/trust, reasoning/engine); 5 tool writers (tools/router, tools/code/{sandbox,code_reader,code_analyzer, code_editor}).

v0.3.0 — Platform Skeleton (entered 2026-05-13, closed 2026-05-25 at v0.3.3 stabilization)

Closure summary (2026-05-25): Platform Skeleton theme completed across three Zenodo-archived releases:

v0.3.0 — V3' Protocol v1 methodology spec (docs/research/v3prime-protocol-v1.md)
v0.3.1 (DOI 10.5281/zenodo.20363998) — D1 Adaptive Budgeting closure with 7-tier natural-stop gradient
v0.3.2 (DOI 10.5281/zenodo.20372649) — D5 Auto-routing on Provider Contract (10-PR sequence #474–#484)
v0.3.3 (DOI 10.5281/zenodo.20374227) — D6 retry-wiring follow-up cycle (PR #486 wiring + #487 audit + #488 native Ollama done_reason)

ROADMAP §Plugin contract / Change Request / Knowledge cascade / Governance / Carryover / v0.3-only follow-ups / Measurement framework Direction 1+4+5+6(J) — all checked. Done-when: 3/4 fully satisfied + 1/4 partial (first external pack author trial awaits Ali mid-June Gemini PR). v0.4 entry triggered.

Theme: define and freeze the extension contract that all future domain packs will be built against.

Required for: any domain pack work (forbidden until this gate passes).

Plugin contract — the v0.3 core

→ Plugin API status: 8/8 PRs landed (100%) as of 2026-05-24 post-PR-C10/C6.b. The eight-PR sequence (PR-C2 / C3 / C5a / C5b / C6 / C6.b / C7 / C8 / C9 / C10) closed in the 2026-05-22~24 window. Full breakdown: docs/handovers/v0.3.x-audit-2026-05-23.md.

→ All other v0.3 gate items closed in the 2026-05-24 marathon session: CR-E shadow rows via Stage B 4-PR sequence (#433/#434/ #435/#436), module size violations via Stage C 5-PR sequence (#427/#428/#429/#431/#432), Audit Phase 4b-2 JSONL-writer removal via Stage D.1 (#430).

Change Request — finish the primitive

CR-A~D completed v0.2.x — scoping (CR-A), state machine + apply dispatcher + wiki_entity (CR-B, PRs #237/#243), workspace UI panel (CR-C, PR #239), run_jobs apply path (CR-D, PR #240).
CR-E: route self-evolution approvals (/admin/patch/approve, /admin/proposals/{id}/approve|reject) through core/change_request.py as a shadow row so the unified audit shape becomes part of the platform contract. ✅ Stage B 4-PR sequence complete 2026-05-24: - CR-E.1 (#433) — target_type enum + no-op apply dispatcher (self_evo_patch / self_evo_proposal) - CR-E.2 (#434) — /admin/patch/approve shadow write + close on bench-pass / apply_failed / regression - CR-E.3 (#435) — /admin/proposals/{id}/approve|reject shadow write + close - CR-E.4 (#436) — /admin/patch/audit unifies legacy JSONL + CR-shadow read via include_shadow=True (default ON in the endpoint, default OFF in the library function for back-compat) Dual-write permanent. Legacy james_patch_log.jsonl + james_evo_log.jsonl remain authoritative for the deploy timeline; CR row is additive shadow with a no-op apply handler. The 4 locked JSONL-shape tests (test_self_evolution_gate, test_evolution_bench_gate, test_evolution_rollback, test_evolution_audit_query) all stay green.
(Stretch) Open the target_type registration API to plugins — today it's a closed enum on purpose; this is the surface every plugin pack will hook through.

Knowledge cascade — relation provenance

Replaced the v0.2 single-confidence field with sources: [{doc_id, weight, role, ts}]. 5-phase plan completed: Phase A schema (#266), Phase B ingestion (#269), Phase C delete (#270), Phase D modify (#274), Phase E graph editor (#271, #273). Two hotfixes for cross-doc aggregation + noisy-OR derivation (#349, #350). 12 invariants locked in tests/test_relations_schema.py + tests/test_phase_b_ingestion_sources.py. Postmortem: docs/postmortems/2026-05-20-knowledge-cascade-defects.md.

Governance — license / CLA / monitoring

License decision for v0.3: MIT held. Trigger conditions (T1–T5), conversion procedure, and pre-built infrastructure (CLA §4-bis relicensing grant, plugin license: field, trademark + patent tracks) committed to docs/LICENSE_PLAN.md (2026-05-11).
CLA Assistant install + docs/legal/CLA.md + .github/workflows/cla.yml — operator window closed 2026-05-20 (B-3/B-4/B-5/B-6 all verified end-to-end with a dry-run from a separate GitHub account). External contributors can sign before opening their first PR with the relicensing grant in place.
THIRD_PARTY_LICENSES.md (dependency inventory; license-strength independent) — ✅ PR #414 (2026-05-23, idna 3.13 → 3.15 security pin refresh; full pip-licenses snapshot at repo root). Re-run only on dependency churn.
Quarterly trigger monitoring — first measurement recorded at the v0.3 release in docs/LICENSE_PLAN.md §8 — ✅ PR #415 (third quarterly measurement, 2026-05-23, stars 5→13, triggers 0/5).
Trademark + patent tracks opened (lawyer consult scheduled, progress logged in docs/LICENSE_PLAN.md §6 / §7) — operator track

Carryover follow-ups (from v0.2.x)

core/memory/store.py split — was 21 KB at v0.2 close; currently ~17 KB on origin/main (9a756b7). Reached compliance without an explicit split (incidental shrinkage from later refactors). No action needed.
Audit Phase 4b-2 — remove 16 JSONL writer sites — ✅ Stage D.1 (#430, 2026-05-24). 14 files touched, ~158 lines removed / 47 added (entry dicts + mirror_* calls retained; only the try: open(JSONL_PATH).write(json.dumps(entry)) blocks + orphaned path constants + import json removed). SQLite audit_log table via core/audit_bridge.py is now the sole sink. tests/test_router_capability.py was migrated to the mirror_to_audit_db capture-via-mock pattern (9/9 pass).

v0.3-only follow-ups discovered post-design

These items were identified during the v0.3 cycle but are not in the original design memos:

i18n contract invariant — backend label_key pattern across 7 modules (~316 entries). Convention test in 5 test_*_i18n.py files. Established by the 2026-05-22 sweep series (PRs #393/#394/#395/#396/#397/#398/#400/#404).
Reasoning cap defaults — DEFAULT_MAX_TOKENS for query_rewriter / planner / reflect / verify bumped from 200/400/400/400 to 4096 each, above the ~500-token reasoning floor measured on gemma4:e4b (PR #399 + V3'.a/.b/.c/.d 4-stage validation set, PR #407).
CASCADE vs EVENT separation axiom — established by PR #401 (architecture memo). v0.4 first ship bundle rescoped to T1+T7+T2 (PR #403).
Module size gate violations discovered by 2026-05-23 audit → ✅ Stage C 5-PR sequence complete 2026-05-24. All five files now split into mixin/composition packages with every sub-module ≤17 KB: - C.1 core/wiki_generator.py 51 KB → 6-file mixin package (max 17 KB) via PR #427 - C.2 core/cascade.py 24 KB → 4-file package (max 12 KB, Phase C delete + Phase D modify + shared helpers) via PR #428 - C.3 core/character_profile.py 24 KB → 4-file mixin package (max 11 KB) via PR #429 - C.4 core/security_layer.py 23 KB → 5-file package (max 7.2 KB; landed after Stage D.1's JSONL removal shrunk the file ~50 lines first) via PR #431 - C.5 core/graph_editor.py 20 KB → 3-file package (max 11 KB, edge mutation writes / helpers / facade) via PR #432 Pure refactor — no signatures, return shapes, or behaviour changed. PLATFORM_READINESS Gate v0.3 dimension A fully clear.
admin.html → 3-page split — inventory completed (PR #385); actual split (Option-A Phase 3) deferred. Operator decision whether to land before v0.4 entry or run as v0.3.x parallel.

Measurement framework track (V3' series, 2026-05-22~ open)

The V3'.a/.b/.c/.d (4-stage cap-budget validation, PR #407) + V3'.e (substitution/synthesis mode split, PR #440) sweeps together established a reproducible 3-axis measurement framework for LLM cost behavior. Robin Converse's 2026-05-23 same-day cross-stack replication (issue #448, repo triavalabs/gemma4-26b-mode-split) adopted the JAMES JSON schema as the cross-stack analysis template — graduating the framework from "data we publish" to "schema another lab analyses against".

Three confirmed axes:

Mode split (substitution = deterministic / synthesis = variable). Robin original + JAMES e4b + Robin 26b — confirmed on both stacks.
Workload gradient (heavy / light / no synthesis cost scales with task weight). JAMES V3'.e + V3'.a~d — confirmed on e4b.
Model-scale efficiency (synthesis cost ∝ 1/param-count; substitution invariant). Robin 26b 2026-05-23 — new axis, single point so far.

Six follow-up directions queued for v0.3.x cycle (none are domain features; all are platform-skeleton strengthening or research):

Direction 4 — Substitution bypass verification on e4b (1-day, JAMES standalone). Patch v3prime_e_mode_split.py to record unique_response_count per cell; verify Robin's Finding 1 ("mode bypasses sampling layer") on our e4b data. If unique=1 confirmed → publishable mechanism axis 1 strengthened.
Direction 1 — Adaptive Budgeting in Cognitive Middleware (closed 2026-05-25). PR #461 (TaskBudget module + experiment driver + cognitive-stages extension), PR #463 (v2 heuristic CAP_LIGHT 800→1200 + closure result docs + 7-tier monotonic natural-stop gradient, 62→1681 tokens, 27× dynamic range, cross-sweep stable <5% per tier), PR #469 (entity_extract cap 1500→4096 alignment). 4 cognitive stages all zero truncation + zero quality regression. Ships as safe / latency-positive / memory-positive / defensive-bound. Sub-finding: verify is a high-clustering cognitive stage (~12.5% unique across 40 baseline calls) — adds task-type axis to Mechanism 2.
Direction 2 — Task-weight metric formalization — absorbed into Direction 5 as a dependent fragment, not run as an independent 1-2 week cycle. The 7-tier natural-stop gradient from Direction 1 closure provides the ground-truth dataset. Per Build-don't-broadcast principle: research-only directions are run on collaborator cadence; D2 in isolation is collaborator-coupled but the piece D5 needs is product work and lands inside D5.
Direction 6(J) — Methodology spec standalone (closed 2026-05-24, PR #457). docs/research/v3prime-protocol-v1.md shipped: 441 lines, 12 sections + KO, JSON schema frozen v1, 12-month grace policy, BibTeX (author=Hashevolution), §11 external adopter table (JAMES + Triava Labs). Other labs can cite the protocol without depending on JAMES code.
Direction 5 — Auto-routing on Provider Contract (closed 2026-05-25, 10-PR sequence PR #474-#483). New core/reasoning/router.py (D5.A skeleton + D5.B capability tags + D5.C.1 policy decision tree) consumes Direction 1's budget to select backend (small/medium/large tier × local/sovereign/cloud provider). 5-stage wiring complete (query_rewriter / planner / reflect / verify / synth via trace_helpers) — every production LLM call path consults the router, gated by JAMES_AUTO_ROUTER flag (default OFF → byte-identical to pre-D5). Audit row reason:route per resolve call. Verify stage is grounding-critical → policy escalates to large tier when registered (e.g. JAMES_ENABLE_CLAUDE_BACKEND=1). Provider Contract surface unchanged (router sits above L1). Cross-lingual RAG option 3 (core/entity_alias_pack.py + graph_engine.build_entity_map_snapshot augmentation, PR #483) bundled — KO↔EN surface forms now resolve to the same wiki entity_id without per-install frontmatter edits. Per Build-don't-broadcast principle (memory: feedback_build_dont_broadcast) this is a product cycle — no public broadcast, no Robin coupling. Operator-run STEP 7 sweep numbers (D1 7-tier ground truth) integrate into the closure result doc when operator runs the measurement; the wiring is bench-neutral at flag-OFF default.
Direction 3 — Cross-family generalization (~2-3 weeks, research). V3'.e on Llama 3.1 / Qwen 2.5 / DeepSeek v2 via Ollama. Output: mode-split universality vs Gemma-only judgment. Result first-shared with Robin (axis-1 owner) per collaborator informational-notice pattern.
Direction 6(I) — Joint paper consolidation (Track 5 essential). Three-author byline (Ali / Robin / Hashevolution) after Directions 1-5 produce input. Track 4 scope-lock note applied.

Collaborator interaction pattern: Track 1 success pattern (land first + "no action needed" DM) applies to all directions except 6(I) joint paper. Direction 3 result first-shared to Robin; Direction 5 design preview shared to Ali; otherwise informational notice on PR landing.

v0.4 / v0.5 alignment: Direction 1 (adaptive budgeting) + Direction 5 (auto-routing) integrate naturally with v0.4 Layer 4 GOVERNANCE/EVENT tracks (budget decisions = audit trail; routing decisions = events). Direction 5 also strengthens v0.5 Domain Pilot value proposition (domain pack-aware model routing). Detailed handover: docs/handovers/v0.3.x-measurement-framework-track.md.

Done when

Criterion	Status
A new contributor can build a no-op pack from `docs/PLUGIN_AUTHORING.md` alone in < 1 day, load it, and observe its effect	⚠️ — PLUGIN_AUTHORING.md exists (#412); end-to-end author run not yet validated. Awaits first external pack author.
The dogfood test passes: `packs/general/` produces byte-identical STEP 7 results to v0.2 main; deleting the pack breaks the server cleanly	✅ — `packs/general/` registers at startup via PR-C5b (#418); `scripts/dogfood_check.py` (PR-C10 #420) locks the four runtime invariants (default loads general; empty `JAMES_PACKS` refused; missing pack refused; path-traversal refused) on every PR via `.github/workflows/packs-eval.yml`. Byte-identity is preserved because the overlay is empty.
Every self-evolution approval row has a paired Change Request row (CR-E acceptance)	✅ — Stage B 4-PR sequence complete. PR #433 (target enum + no-op apply), PR #434 (`/admin/patch/approve` shadow write), PR #435 (`/admin/proposals/{id}/approve
CLA Assistant blocks any unsigned external PR at the workflow gate	✅ — verified end-to-end 2026-05-20

→ 3/4 fully satisfied + 1/4 partially satisfied as of 2026-05-24 (post-Stage-B/C/D marathon session). Only remaining partial: first external author trial against PLUGIN_AUTHORING.md (infrastructure complete; awaits Ali's mid-June Gemini backend PR as the first external trial). See docs/handovers/v0.3.x-audit-2026-05-23.md for the staged plan and the post-marathon reconciliation table.

Out of scope (deferred to v0.4)

Any domain-specific pack (legal, food, retail)
External plugin marketplace
Plugin signing infrastructure beyond manifest hash
Multi-approver workflows / team-project-department scoping on CR (still v0.3 plugin contract surface; ships only if a real second user needs it)

v0.4.0 — Layer 4 Lifecycle Semantics (entered 2026-05-25, T1+T7+T2 first bundle shipped 2026-05-27 at Sprint 5)

Status: T1 (Temporal Validity) + T7 (Supersede Chain) + T2 (Contradiction Arbitration) first bundle landed via Sprint 5 (8 PRs, #523~#543). v0.4.0 is release-ready — the CASCADE / EVENT separation invariant is provable end-to-end via the release-gating tests in tests/test_t7_release_gating_invariants.py.

Entry handover: docs/handovers/v0.4.0-entry-track.md — 6-sprint plan covering data correctness, UI consistency, plumbing, retrieval quality, Layer 4 main theme (T1+T2+T7), and long-term backlog. Sprint 0 (this entry) + Sprint 1 (graph entity-event relation diagnostic + language detection&matching) ran first.

Sprint 5 first-bundle entry: docs/handovers/v0.4.0-sprint5-layer4-first-bundle-entry.md — the locked-decision 7-PR sequence (PR-0 schema validators → PR-T1.A migration → PR-T1.B expiration cascade → PR-T7.A supersede chain → PR-T7.B release-gating invariants → PR-T2.A classifier → PR-T2.B A-path routing → PR-T2.C B-path routing → PR-T7.C closure) that shipped here.

Original Layer 4 scope (preserved below for the eventual T1/T2/T7 sprint)

Theme: deepen the memory lifecycle beyond the v0.3 Layer 3 cascade (Memory OS) into Layer 4 — two complementary tracks EVENT/TEMPORAL (semantic evolution) and GOVERNANCE (write-time control), plus a Layer 3 extension for causality. Seven areas (T1–T7) chosen from the 2026-05-21 user critique series.

Core architectural axiom — CASCADE vs EVENT separation (docs/architecture/memory-lifecycle-architecture.md §1.5):

Class	Trigger	Handling	Example
Invalidated → CASCADE (Layer 3)	Explicit invalidation signal (doc delete, ingestion error, source revoke)	Source removal + propagation + orphan sweep	Mis-uploaded doc / withdrawn report / typo correction
Superseded → EVENT/TEMPORAL (Layer 4-A)	New doc / time progression / policy change / state transition	Validity window close + supersede chain (past data preserved)	CEO change / policy revision / contract termination

Many naive RAG/KG systems collapse both into one mechanism (overwrite/append) → contradiction explosion, stale retrieval, hallucination increase. v0.4 enforces the separation as a test-level invariant (T7 invariants: supersede_does_not_trigger_cascade / cascade_preserves_supersede_chain).

Why v0.4 was retargeted from "Domain Pilot": the v0.3 cascade (Phase A–E + 2 hotfixes) revealed that "Layer 3 alone" doesn't cover real operational governance — stale facts accumulate, contradicting sources reject silently (Gate 5), manual entries have no reviewer hierarchy, and event sourcing for arbitrary-time graph replay is absent. The Domain Pilot moves to v0.5 (it requires Layer 4 governance to be production-credible anyway).

Required for: any v0.5 domain pilot (forbidden until Layer 4 T1+T7+T2 minimum lands and the CASCADE/EVENT separation invariant is provable).

Scope — 7 areas across 3 tracks (T1–T7)

Stage	Area	Track	Goal
T1	Temporal Validity & Expiration	EVENT	Fact valid window definition + auto-expiration (mark, not delete)
T2	Deterministic Contradiction Arbitration	GOVERNANCE → routing	LLM-free deterministic resolution + A/B classification (CASCADE vs EVENT)
T3	Evidence Aging & Trust Decay	EVENT	Per-source weight decay by domain function
T4	Reviewer Authority Hierarchy	GOVERNANCE	Multi-level manual source governance (analyst / manager / admin)
T5	Replayable Audit Graph	EVENT	Event-sourced reconstruction, arbitrary-time graph replay
T6	Causality Chain Tracking	CASCADE extension	Derived fact base-fact tracking + auto-invalidation propagation
T7 ⭐	Supersede Chain	EVENT	Edge-level `status: {active, superseded_by}`. World change → past preserved + new edge created + chain walk

Full design: docs/design/v0.4-lifecycle-semantics-roadmap.md. Reference architecture: docs/architecture/memory-lifecycle-architecture.md.

Phase plan

v0.4.0 (2026 Q4 ~ 2027 Q1): T1 + T7 + T2 (EVENT core + routing). T1 (validity window) and T7 (supersede chain) ship together — separation is meaningless without both. T2's A/B routing enforces the separation at write time (misclassification becomes immediately testable).
v0.4.1 (2027 Q1 ~ Q2): T6 — Causality Chain (CASCADE extension)
v0.4.2 (2027 Q2): T3 — Evidence Aging (EVENT)
v0.4.3 / v0.5 prep (2027 Q3 ~ Q4): T4 + T5 — Reviewer Authority + Replayable Audit

Invariants (extending v0.3's 12 invariants)

Existing 12 locked in tests/test_relations_schema.py + tests/test_phase_b_ingestion_sources.py (Layer 3 cascade).
New T1–T7 invariants: ~27 (cumulative 39). Per-stage contract test files following the 2026-05-22 i18n sweep pattern (tests/test_*_i18n.py × 5 already pin label_key conventions — same _contract-style files for T1–T7).
Critical separation invariants (T7):
- test_supersede_does_not_trigger_cascade — EVENT mutation must not call Layer 3 cascade_remove
- test_cascade_preserves_supersede_chain — CASCADE invalidation must not break status.superseded_by links (active history preserved)
- test_historical_replay_via_chain — reconstruct_view_at(t) accurately reconstructs the graph state at any past time

Cross-cutting contracts (preserved from v0.3)

i18n label_key — every UI-exposed label from new Layer 4 modules (reviewer_rank, approval_state, contradiction reason, aging policy names) must follow the backend label_key contract established by the 2026-05-22 sweep across 7 modules. Convention test in the corresponding test_*_i18n.py blocks silent regression.
Bench gate (CLAUDE.md rule 2) — T1 expiration cascade, T6 invalidation cascade, and any other change touching core/retrieval / core/graph / core/reasoning must paste STEP 7 bench numbers in PR body.
Module size gate (CLAUDE.md rule 5) — new core/lifecycle/*.py files each < 20 KB. Split first if a change would push over.

Done when — ✅ All items satisfied at v0.4.0 (2026-05-27)

✅ T1 + T7 + T2 (minimum) shipped, with new invariants green (tests/test_lifecycle_*.py + tests/test_t7_*.py + tests/test_t2_*.py, 123 tests).
✅ Separation invariant provable — test_t7_release_gating_invariants.py pins three release-gating invariants against the actual wiki fixture (not mocks): test_supersede_does_not_trigger_cascade, test_cascade_preserves_supersede_chain, test_historical_replay_via_chain. The B-class CEO-change scenario (Joby CEO Alice → Bob) is end-to-end exercised in test_supersede_chain_replayable_after_dispatch.
✅ Operator scenario A (CASCADE): cascade_remove_doc_from_sources unchanged from v0.3 + now reachable via contradiction_router.route_a_invalidate with audit row carrying mutation_type=invalidated.
✅ Operator scenario B (EVENT): expiration_cascade (T1) + supersede_edge (T7) + route_b_supersede (T2.C) land. Edges are marked status.active=False not deleted; historical replay works via reconstruct_view_at.
✅ Reference architecture memo + 7-area design memo published (docs/architecture/memory-lifecycle-architecture.md + docs/design/v0.4-lifecycle-semantics-roadmap.md).
⏭️ Carried into v0.4.1: ingestion-path caller that invokes dispatch_contradiction at a real contradiction-detection point. Sprint 5 ships the primitive surface + the routing wire; the call site is the next operator integration. End-to-end STEP 7 supersede bench (Alice → Bob historical replay) deferred to the ingestion-wire PR.

Out of scope (deferred to v0.5)

Any domain pack (legal / food / retail) — moved to v0.5
External customer onboarding playbook
Public eval results in eval/RESULTS.md
6-month production track record

v0.4.1 — T6 Causality Chain (CASCADE extension, shipped 2026-05-28)

Status: ✅ CLOSED. T6 Causality Chain landed as the v0.4.1 cycle's main theme (5-PR sequence T6.A → T6.B → T6.C → T6.C.b refinement → T6.D + the v0.4.0 carry-over T2.D-1/2/2.b/3 + QVT α track full closure). 21 PRs in a single multi-day session (#548 ~ #566).

Entry handover: docs/handovers/v0.4.1-t6-causality-chain-entry.md — 4-LOCK decisions (eager trigger / operator-tagged + LLM flag / strict cycle reject / C.b foundational-vs-corroborative semantics — Decision 4 refined in T6.C.b vs the original memo text).

Done — T6 5-PR sequence + 4 release-gating invariants

PR-T6.A (#562) — derived_from schema field + validate_edge_t6_derived_from cycle validator (Decision 3 LOCK) + apply_t6_edge_defaults idempotent helper. scripts/migrate_v041_lifecycle.py (--dry-run / --apply / --verify / pre-write snapshot). 23 contract tests.
PR-T6.B (#563) — core/lifecycle/derivation.py extract_derivation_chain (operator-tagged path + JAMES_T6_LLM_DERIVATION flag-gated LLM-inferred path). 14 contract tests.
PR-T6.C (#564) — core/lifecycle/causality.py invalidate_derived_facts(base_fact_id, entity_root, *, additional_empty_bases, audit_emit). Soft-invalidate (status.active=False + mutation_type=invalidated, sources preserved for T7 replay). Atomic per-file writes. 19 contract tests.
PR-T6.C.b (#565) — foundational-vs-corroborative refinement. transitive / inferred are hard deps (any base empty → invalidate); operator is corroborative (only invalidates when no hard deps AND all operator bases empty). 22 tests (19 original + 3 new C.b cases).
PR-T6.D (#566) — cascade integration. cascade_remove_doc_from_sources now calls invalidate_derived_facts for every relation whose sources became empty (single walk batched via additional_empty_bases). tests/test_t6_release_gating_invariants.py (5 tests): test_derived_invalidated_when_base_removed, test_partial_base_loss_preserves_derived (T6.C.b), test_self_reference_rejected_at_write + test_two_hop_cycle_rejected_at_write (Decision 3), test_cascade_invalidate_emits_audit_row. Run against tmpdir wiki fixtures + real cascade_remove_doc_from_sources — no mocks of the cascade itself.

Done — v0.4.0 carry-over `dispatch_contradiction` ingestion wiring (T2.D, 4 PRs)

PR-T2.D-1 (#558) — core/lifecycle/contradiction_ingest_detector.py (P1 different_tail + P2 divergent_validity patterns). 19 tests.
PR-T2.D-2 (#559) — core/lifecycle/ingest_contradiction.py + _merge.py pre-merge hook. Flag-gated by JAMES_T2D_INGEST_DISPATCH (default OFF). 10 tests.
PR-T2.D-2.b (#561) — A_invalidate cascade race fix via PendingCascade deferred-execution. 15 tests.
PR-T2.D-3 (#560) — step7 v6 q17 "Anthropic의 CEO는 누구야?" + acceptance integration test. 6 tests.

Done — QVT α track full closure (6 PRs)

PR-α-1 (#550) — docs/design/v0.4-qvt-alpha-non-saturating-oracle.md. 3-axis non-saturating oracle + per-PR Quality Delta Card pattern + 5 exemption labels.
PR-α-2 (#551) — step7 fixture v4 → v5 (gold_signals + abstention_truth).
PR-α-3 (#552) — eval/qvt/oracle.py + scripts/qvt_capture_baseline.py (operator wrapper).
PR-α-4 (#553) — .github/PULL_REQUEST_TEMPLATE.md + CLAUDE.md rule 2 extension + docs/ARCHITECTURE.md §5.7.10 QVT subsystem.
PR-α-3 baseline capture (#555) — eval/qvt/baseline_2a31b20.json (N=3 paired reruns, canonical reference).
PR-α-3 oracle calibration (#556) — Korean security-block phrase additions + blocked=True short-circuit. abstention_f1 0.29 → 0.67 median.

Done — also in v0.4.1

Replayable RAG positioning (#548) — README + ARCHITECTURE category framing.
F9 cycle full closure (#549) — q15 zero-recall closure result doc + audit-trail bench JSON.
v0.4.0 post-mint DOI (#554) — 10.5281/zenodo.20411354.
v0.4.1 entry memo (#557).

Default-off invariant preserved across every opt-in

Flag	Default	Verification
`JAMES_T2D_INGEST_DISPATCH` (T2.D-2)	OFF	`_merge.py` pre-merge hook only fires when `=1`
`JAMES_T6_LLM_DERIVATION` (T6.B)	OFF	`extract_derivation_chain` LLM-inferred path needs provider AND flag
T6.D cascade integration	ON (default-on per cycle scope)	byte-identical retrieval because migration adds `derived_from: []`; no actual derivations yet

Done when — ✅ All items satisfied at v0.4.1 (2026-05-28)

✅ T6 5-PR sequence merged with all 83+ contract tests green.
✅ T6.D 4 release-gating invariants run against real wiki fixtures (tests/test_t6_release_gating_invariants.py).
✅ T2.D ingestion wiring (carry-over from v0.4.0) shipped with race-free pending_cascades pattern.
✅ QVT α track complete: oracle module + canonical baseline JSON + PR-gate template + CLAUDE.md rule 2 extension + ARCHITECTURE.md §5.7.10.
✅ Migration script (scripts/migrate_v041_lifecycle.py) is idempotent + byte-stable; --verify mode confirms.
✅ 69 pre-T6 cascade tests still green (no regression).
✅ Closure docs published: docs/release_notes_v0.4.1.md + CHANGELOG [0.4.1] + .zenodo.json v0.4.1.

Out of scope (deferred to v0.4.2+)

T3 Evidence Aging — confidence decay over time (EVENT track)
T4 Reviewer Authority Hierarchy — multi-level governance (GOVERNANCE track)
T5 Replayable Audit Graph — full event-sourced reconstruction (partial in v0.4.0 via reconstruct_view_at)
~~v0.4-end QVT ablation matrix capture (18 cells × N=3 reruns, ~20 hours operator-run)~~ — re-shaped 2026-05-30~31 into the α-5 cycle below.
JAMES_T2D_INGEST_DISPATCH default flip to ON (waits for fixture coverage)
Production seeding of derived_from (waits for operator workflow or v0.4.2+ LLM path)

v0.4.x — α-5 Ablation Matrix Cycle (2026-05-30~31, T0 closed 2026-05-31 15:52 KST)

Theme: the v0.4-end ablation matrix that v0.4.1's "out of scope" line deferred — re-shaped end-to-end during execution. The single bullet "18 cells × N=3, ~20 h" turned out to be too thin a frame for what the matrix actually needs to deliver. Recording the corrected sizing, the external-benchmark adoption, and the dual purpose so the next operator entering this cycle reads the right shape on the way in.

Dual purpose (per 2026-05-31 user clarification)

Routing decision — flag-ON / tier-gating for each Layer 4 cognitive routing layer (AUTO_ROUTER D5, ADAPTIVE_BUDGET D1, SCOPE_ROUTING LEO). The original framing.
Reasoning-capability evidence — publishable proof that JAMES's layer stack outperforms peers on external benchmarks. Ties to eval/RESULTS.md's external benchmark line that v0.3 deferred — this cycle is the unblock.

Both purposes must be served by every cell verdict; bucket-(d) measurement artifacts must be removed before either narrative ships externally (per feedback_oracle_phrase_artifacts 4-step rule).

Actual sizing (vs the original 18×N=3 ≈ 20h estimate)

19 cells (18 standard + 1 sanity M_M/L1 think=ON) — extra cell carries A2 default-flip evidence inside the matrix.
5-axis oracle — 3 quality (path / graded / abstention) + 2 cost (token_cost, latency_cost) per cell, Pareto verdict (strong-adopt / adopt / efficiency-adopt / tier-gated / reject / zero). Cost integration was added to serve user requirement #3 (token / time efficiency).
Workspace runtime (MultiHop-RAG balanced-100, 931-entity workspace, A2 think=OFF):
- Ingest: 264 min for 183 articles (latency grew with workspace size — first 25 s, last 197 s)
- Baseline N=1: 107 min (100 queries × ~64 s)
- T0 smoke (N=1, 3 cells): ~5.3 h estimate
- Full N=3 sanity-included T0+T1: ~16 h
- Full N=3 T0+T1+T2 across all 3 tiers: ~30 h (vs original 20 h)
Adaptive tiering — --t0-smoke (~2-5 h) gates whether T1 (M_M, ~5.5 h) is worth running; T1 gates whether T2 (M_S + M_L, ~13 h) adds tier-gating signal.

External benchmark adoption

eval/RESULTS.md originally said "BEIR / MS MARCO — too large for laptop, defer to v0.3" and "KLUE-RC — no clean public option yet, watch for v0.4." This cycle finally lands an external benchmark via MultiHop-RAG (Tang & Yang, EMNLP 2024) — 2,556 multi-hop QA over 609 news articles, CC-BY-4.0, with 4 question types (comparison / inference / temporal / null) that map directly onto the routing / abstention measurement need.

Workspace isolation via the existing JAMES_WORKSPACE env (config.py:74, core/plugins/workspace.py) — zero code change to the data-dir resolver. Production wiki is untouched throughout the cycle.

PR sequence (24 PRs to date, 2026-05-30 → 2026-05-31, T0 smoke mid-flight)

PR	Layer	Note
#615	reset	MultiHop-RAG external benchmark + 5-axis + per-question-type matrix
#616	exec	Ingest wrapper + bench timeout fix + path-axis finding
#617	docs	§7.4 first per-question-type baseline signal
#618	fix (d)	source-recall — bench was dropping `response.sources`
#619	fix (d)	abstention phrases — gemma4:e4b English refusals
#620	tool	re-score tool + §7.4 76% → 36% correction
#621	docs	4-bucket diagnostic taxonomy + dual purpose
#622	fix	render_report defensive path + bucket retroactive
#623	fix (d)	session_id suite-aware + 3 narrow abstention phrases
#624	docs	ROADMAP §v0.4.x α-5 cycle entry (this section, initial form)
#625	fix (a)	matrix runner — bench subprocess respects `--suite` (was hardcoded step7)
#626	docs	T0 smoke result analysis TEMPLATE (`qvt-ablation-T0-smoke-analysis-TEMPLATE.md`)
#627	docs	Pareto verdict walk-through + CLAUDE.md rule 2 `fix` label exempt + diagnostic post-mortem
#628	feat	T0 analysis fill script — auto-populate template from cell JSONs
#629	docs	oracle.py → package split design (deferred to post-matrix)
#630	docs	α-5 cycle summary outline — reviewer-ready closure read
#631	docs	publishable narrative draft — "Don't Build a Layer for the Bug"
#632	docs	bucket-(c) LLM-judge abstention detector design memo
#633	feat	`qvt_promote_findings.py` — auto-draft memory entries from findings.md

All 19 cycle PRs above land on main between f7762a3 and ba50c47. Two landmark findings of the cycle — path_recall = 0 and null_query hallucination = 76% — were both bucket-(d) oracle artifacts, fixed in #618 / #619 / #623. Without the 4-step verification rule (memory feedback_oracle_phrase_artifacts) they would have generated wrong-bucket follow-ups (new citation layer, grounding architecture change) and the matrix verdicts would have read as "JAMES fails the benchmark" when the failure was on the measurement side.

Wrong-fix avoided count: 3 (path_recall=0 → new citation layer prevented #618; 76% hallucination → grounding rewrite prevented #619+#623; matrix near-zero verdict → AUTO_ROUTER decommission prevented #625). Cumulative system code change attributable to α-5 measurement debt: 0 lines.

Mid-cycle infra additions (#626–#633): 1 template + 1 fill script + 2 design memos + 1 closure outline + 1 publishable narrative + 1 finding-promotion script. Locked deferred-bucket-(c) follow-up + locked the methodology lesson in a publishable form before the cycle closes.

Cycle deliverables — T0 closure 2026-05-31 15:52 KST

5-axis matrix report at reports/promo-assets/v0.4-qvt-ablation-matrix-20260531T065209.md. Both L1/M_M (baseline) and L5/M_M (full stack) classified reject vs corrected baseline; sanity cell L1/M_M-thinkON also reject.
T0 analysis at reports/research-runs/qvt-ablation-T0-smoke-result-20260531-1552.md.
Rescore audit at reports/research-runs/qvt-ablation-rescore-summary.md — 3/3 cells QQ-bugged at write time, all rescored in one pass.
Findings log at reports/research-runs/qvt-ablation-findings.md — all entries bucket:-tagged (a/d).
ROADMAP / backlog sync — this section locked, status moved from "in flight" to "T0 closed". A1 in backlog §2 marked closed.
Routing-flag default-flip PRs per layer — not applicable at production tier (Branch B verdict). Tier-gated assessment requires T1 (M_S + M_L) which is post-T0 operator decision.
T1 / T2 operator decision — see §"T0 verdict + post-closure" below.

T0 verdict + post-closure (2026-05-31)

Read	Number	Compared to
L1/M_M baseline (production)	path 0.419 / graded 0.327 / abst_f1 0.591	corrected baseline `3a961a3_rescored` (path 0.404 / graded 0.343 / abst_f1 0.609)
L5/M_M full stack (all routing ON)	path 0.412 / graded 0.317 / abst_f1 0.500	abst_f1 -0.091 vs L1 — real regression
Sanity L1/M_M think=ON	path 0.404 / graded 0.370 / abst_f1 0.533 / token -109 chars / latency -1.7 s	think=ON not unambiguously winning; A2 default-flip gate (real-query QDC) unchanged

Verdict (post-closure-corrected 2026-05-31 PM): routing-layer stack as exercised in α-5 reduces to ADAPTIVE_BUDGET + SCOPE_ROUTING + (AUTO_ROUTER no-op) because the matrix env registered only the always-on ollama_local backend (core/reasoning/backends/__init__.py:329); without multi-tier backend registration the routing policy in core/reasoning/router.py collapses to legacy on every branch. The combined stack regresses abstention_f1 by 0.091 without meaningful cost benefit (token -1.5%, latency +3.6%) at the M_M production tier. Both cells classify reject under the 5-axis Pareto rule.

AUTO_ROUTER verdict is not in evidence at this cycle; proper measurement requires multi-tier backend registration (engineering preconditions documented in α-6 design memo). ADAPTIVE_BUDGET was under-instrumented — judged by quality axes when its design intent (per core/reasoning/budget.py docstring) is cost-optimization at quality-neutral. Per-layer isolation (L2 / L3 / L4 cells) defers to T1. Branch B of the publishable narrative §6 applies as "conditional Branch B" — routing-layer stack inert as measured by quality axes, with explicit caveats. Opt-in flags remain available; no default-flip PRs file at this verdict.

Cycle totals: 36 PRs (#608 → #645+), 5 measurement-side fixes (3 bucket-d + 2 bucket-a), 4 wrong-fix-averted, 0 lines of JAMES code changed against α-5 measurement debt. The publishable claim (see §6.1 of reports/research-runs/alpha-5-publishable-narrative-DRAFT.md) is the cycle's load-bearing artifact.

Operator decision — T1 / T2 sizing

T1 (M_M + more rows) and T2 (M_S + M_L) are operator-budget calls. Recommended:

T1 (~5.5 h on M_M) — measures whether individual layers (L2/L3/L4) show effects that all-on L5 smears together. Could surface ADAPTIVE_BUDGET (L3) wins inference-query graded while AUTO_ROUTER (L2) hurts abstention.
T2 (~13 h on M_S + M_L) — tests the tier-gated routing hypothesis: smaller models might benefit MORE from routing (compensating for base-capability gap); larger models might benefit LESS. T2 closes the cycle's "model is universally tested" caveat.

Both deferable. T0 closure is itself shippable as "production-tier evidence: routing layers inert" + Branch B publication. T1/T2 fold into an α-5.1 or α-6 cycle.

Open methodology questions for the next cycle

bucket-(c) LLM-judge abstention detector — 6 of 9 remaining FN on baseline_f7762a3 use "not possible to" / "is not available" phrasings that broader phrase additions would FP-flood. An LLM classifier ("does the answer refuse to answer the question?") would resolve them without phrase-overlap risk. Deferred candidate.
Korean-corpus version of MultiHop-RAG — current cycle is English only (gemma4:e4b handles both per core/i18n.py, but the matrix measures English routing only). No public Korean multi-hop RAG benchmark exists; translation cost vs. value is a v0.5+ question.
Production-real-query Quality Delta Card — bridging the benchmark-vs-production gap for the A2 default-flip (JAMES_GEMMA4_E4B_THINK_OFF) decision. Plan is documented but the card itself is a separate cycle.

v0.4.x — α-6 Sector × LLM Ablation Cycle (2026-05-31~, Phase 3a in flight)

Theme: successor to α-5. α-5 toggled 5 routing flags at one model tier and found the routing stack inert at production tier (Branch B). α-6 reshapes to:

Sector-level ablation (10 sectors of JAMES infrastructure instead of 5 routing flags). Cells are sector combinations, not flag combinations.
Multi-LLM extension — gemma3 1b/4b/12b/27b + gemma4:e4b + (later) cross-family (qwen2.5 / llama3.1 / deepseek-v2).
JAMES-vs-vanilla comparison — α-5's L1 baseline already had every other JAMES sector on; α-6's C_minus cell strips them off and measures bare gemma against benchmark.

Cycle sizing (vs original α-6 design memo §3):

4 phases (Phase 0 sector flags / Phase 1 M_M / Phase 2 M_S / Phase 3a scale ladder / Phase 3b cross-family deferred).
75+ PRs through Phase 3a entry, 8 wrong-fix-averted, 0 lines of JAMES code changed against measurement debt.
Per-tier wall-clock: M_XS ~15 min / M_S ~30 min / M_M ~107 min / M_L ~65 min / M_XL ~6-8h (GPU/CPU split, see below).

Phase status (live)

Phase	Tier	Status
Phase 1	M_M (gemma4:e4b)	✅ closed 2026-06-01 AM
Phase 2	M_S (gemma3:4b)	✅ closed 2026-06-01 AM — tier-gated hypothesis REVERSED
Phase 3a step 1	M_XS (gemma3:1b)	✅ closed 2026-06-01 PM
Phase 3a step 2	M_L (gemma3:12b)	✅ closed 2026-06-01 PM — gemma4-only hypothesis REVERSED
Phase 3a step 3	M_XL (gemma3:27b)	🟡 in flight (bench timeout overrides applied, see PR-pending)
Phase 3a closure	recovery curve + closure PR	⏳ post-27b
Phase 3b	cross-family (qwen / llama / deepseek)	deferred (gated on Phase 3a verdict)

Findings (4 tier-tagged per `feedback_finding_size_honest_framing`)

Finding	Tier	Status
S4 citation tier-invariant — path Δ +0.397~+0.420 across 1b/4b/12b/e4b (4-point series, 27b pending)	⭐⭐⭐ candidate	universal-law promotion gated on 27b confirmation + cross-fixture sanity
JAMES S5 effect non-monotonic in pure-model abstention (1b 0 / 4b -0.074 / 12b +0.375 / e4b +0.033)	⭐⭐ partial	mechanism candidate = instruction-following capacity threshold
Graph layer regresses graded (-0.054 at M_M C_rag-graph step)	⭐⭐ measurement debt	resolves on α-7 graph top-K fix; logged in `qvt-ablation-findings.md`
Withdrawn framings — "JAMES amplifier", "inverted-U capability floor", "gemma3 vs gemma4 family", "REVERSES tier-gated hypothesis"	❌ self-deception	superseded by 12b / 27b data; honest framing memory applied

Cycle PR index

reports/research-runs/alpha-6-cycle-pr-index.md.

Closure deliverables (in flight on `feat/v0.4-alpha6-phase-3a-closure`)

Phase 3a 1b analysis (alpha-6-phase-3a-gemma3-1b-analysis-20260601.md) with post-12b reconciliation header
Phase 3a 12b analysis (alpha-6-phase-3a-gemma3-12b-analysis-20260601.md) with honest framing tier table
Phase 3a 27b analysis (post-bench)
Recovery curve doc (alpha-6-phase-3a-recovery-curve-20260601.md) with §5 withdrawn-claims registry
Closure PR consolidating all 3 analyses + recovery curve + 2 timeout fix patches + CLAUDE.md sync + α-7/α-8 design memo seeds

v0.4.x — α-7 Graph Top-K Cycle (post-α-6 closure)

Theme: first JAMES code change against α-5 + α-6 measurement debt. Per α-6 Phase 1 §3 finding, the graph layer (expand_dynamic in core/graph_engine.py:314) is the net regressor on graded at M_M (-0.054). 41-161 entities surface per query at the 931-entity workspace, swamping the LLM context window with low-signal nodes.

Status: design memo landed (docs/design/v0.4-alpha-7-graph-topk.md); implementation gated on α-6 closure + re-baseline.

Scope (per design memo §2):

New module core/graph_topk.py (~3 KB target) — post-DFS top-K filter, default K=10
Tighten DFS_SCORE_THRESHOLD 0.05 → 0.08
Wire into expand_dynamic return
Avoids forcing core/graph_engine.py split (already 20.4 KB, marginally over the 20 KB gate — new module routes around it)
Re-baseline multihop_rag post-PR (~107 min operator action)
5-axis Quality Delta Card with per-question-type cross-tab

Acceptance band (per design memo §4): graded_answer Δ ≥ +0.030 for adopt; +0.010-0.030 for tier-gated; ≤ +0.010 → reject + sub-finding investigation.

Honest framing (per memory feedback_finding_size_honest_framing):

⭐ operational — top-K filtering is a standard GraphRAG technique (MS GraphRAG, Neo4j LLM KG Builder both do variants). PR's value is the measured Δ numbers, not the mechanism.
Don't frame as "JAMES discovered top-K filtering" — false claim.

Cycle dependencies:

Predecessor: α-6 closure (the re-baseline depends on Phase 3a closure being on main)
Successor: α-8 ontology typed-filter (A/B compares against α-7's K-bound baseline)
Parallel: T3 Evidence Aging (orthogonal axis)

α-7 cycle CLOSURE (2026-06-02) — REJECT

Status: cycle closed as REJECT. PR #680 closed without merge. α-7 graph top-K is a research artifact, not a production change.

5-tier remeasurement verdict (1000 queries, 0 timeouts, 0 errors, 10/10 cells written):

Tier	α-6 contribution	α-7 contribution	change	mode change
M_XS (1b)	0.000	0.000	0	unchanged (inert)
M_S (4b)	-0.074	-0.067	+0.007	unchanged (disrupt)
M_M (e4b)	+0.033	-0.283	-0.316	CRASH amplify → disrupt
M_L (12b)	+0.375	-0.076	-0.451	CRASH create → disrupt
M_XL (27b)	-0.181	-0.359	-0.178	worse disrupt

Mechanism (CONFIRMED): top-K=10 removes "evidence-of-absence" signal at every tier where pure-LLM had nonzero abstention capability. Universal regression. K=50/K=25 tuning would not fix — wrong-knob mechanism, not wrong-value problem.

S4 universal-law candidate (path Δ) survives the context reshape: spread ±0.013 across 5 tiers (vs α-6 ±0.022). ⭐⭐⭐ candidate strengthens — citation pipeline is graph-entity-count independent.

Wrong-fix-averted: 10th cumulative (7 α-5 + 2 α-6 + 1 α-7). α-7 cycle's value = preventing the wrong fix from shipping, plus mechanism finding informing α-8.

Carry-forward to α-8:

Bucket-(d) phrase additions (7 total: 5 α-7 PR + 2 follow-up) — eligible for separate docs-only PR for oracle improvement.
α-7 closure analysis preserved at reports/research-runs/alpha-7-closure-analysis-20260602.md.
α-7 5-tier 10 cell JSONs preserved at workspaces/hotpot_eval/reports/research-runs/qvt-ablation-cells/.

Memory entries added 2026-06-02:

feedback_alpha7_top_k_destroys_positive_contributions
feedback_s4_citation_survives_context_reshape
feedback_12b_pure_capability_misclassified_as_plateau
project_alpha_7_closure_state

v0.4.x — α-8 Ontology Typed-Filter Cycle (post-α-7 closure)

Theme: measures whether typed filtering (only surface entities whose type matches query intent) beats α-7's type-agnostic K filter on multihop_rag. A/B isolation: Δ = C_rag-ontology − C_rag-graph_post-α-7.

Status: design memo landed (docs/design/v0.4-alpha-8-ontology-typed-filter.md); implementation gated on α-7 closure.

Scope (per design memo §2 + §3):

Extend core/ontology.py (10 KB → ~14-15 KB; under gate):
- Abstract root Entity + ENTITY_TYPES registry
- 5 new horizontal types: event, date, location, quantity, project (additive; existing person / org / concept / document unchanged)
- 6 new relations: OCCURRED_AT, HAPPENED_ON, LOCATED_IN, INVOLVES, MEASURED_AS, WORKED_ON
Migration cost: 0 — additive only; old types and 931 wiki entities unchanged.
New sector flag JAMES_DISABLE_TYPED_FILTER (disable-polarity, default OFF = sector ON = production byte-identical)
New matrix cell C_rag-ontology (between C_rag-graph and C_rag-full)
Query intent → expected-type classifier (keyword heuristic; LLM classifier deferred to v0.5+)
5-axis QDC + per-question-type cross-tab against post-α-7 C_rag-graph

Acceptance band (per design memo §4):

⭐⭐ adopt — graded Δ ≥ +0.030
tier-gated — +0.010 ≤ graded Δ < +0.030 (enable per question_type only)
reject — graded Δ < +0.010 (heuristic top-K is sufficient; ontology layer stays in codebase but disabled by default)

Rule #1 boundary test (per design memo §2.3): each proposed type checked horizontally. The 5 accepted types pass; regulation, transaction, recipe flagged for deferral or rejection.

Forward compat hook (per design memo §3.4): since: field anchor allows v0.5 domain packs to register pack-specific types via the same mechanism without absorbing them into the mother schema.

Honest framing (per memory feedback_finding_size_honest_framing):

⭐⭐ partial if Δ exceeds noise — typed filtering is a known technique; JAMES-specific value is the empirical position, not the mechanism
⭐ operational if Δ ≤ noise — confirms heuristic top-K is sufficient; finding logged either way
❌ Don't claim "JAMES discovered typed semantics matter" — decades- old territory

Cycle dependencies:

Predecessor: α-7 closure (baseline = post-α-7 C_rag-graph)
Successor: cross-fixture sanity (deferred to v0.5 entry)
Parallel: T3 Evidence Aging (orthogonal)

v0.4.x — Sequencing rule (2026-06-01 strategy handover §4)

Graph-touching cycles serial to avoid measurement baseline drift: α-6 closure → α-7 → α-8. Confidence/time-axis cycles parallel to graph track: T3 Evidence Aging is orthogonal to entity surfacing/typing and can land in any cycle alongside the graph track.

The v0.5 first-domain candidate is enterprise internal knowledge ontology (horizontal, audit/ownership/correction moat) per 2026-06-01 strategy handover §6. This framing applies to v0.5 domain selection criteria, not to v0.4 cycle scope. Mother-platform rule #1 holds — v0.4.x cycles add only horizontal infrastructure (top-K, abstract types, forward-compat hooks), no domain-specific types.

v0.5.x — Cycle close + post-close consolidation (2026-06-12 → 2026-06-13)

v0.5 entry declared 2026-06-12 after v0.4.4 closure (LRB v0.2.3 S3 publication-scale + cycle γ 4-bench infrastructure closure; Zenodo DOI 10.5281/zenodo.20652679).

Operator decision at entry: "이제는 진짜 엔터프라이즈 온톨로지 장착으로 간다" (entry handover).

v0.5 closed 2026-06-12 PM with 21 PRs (#841 – #861) per the v0.5 close handover. 23 additional PRs (#863 – #886) landed between cycle close and 2026-06-13 implementing most of the close handover §5 work queue — see the v0.6 entry skeleton for the post-close work-queue status sweep.

Four work streams (CLAUDE.md rule #1 preserved — actual status at 2026-06-13)

Stream	Scope	Status at 2026-06-13
A — Pre-LOI materials + Hashevolution dogfooding	Customer-facing 자료 v0.4.4 sync (7 docs) + Hashevolution own-company scenario (시나리오 C: internal dogfooding + external outreach 병행) + G1.a/G1.b/G1.c tenant-id contract + G2.a/G2.b/G2.c approval-evidence contract	A.1 / A.2 ✅; A.3 / A.4 operator-pending; G1+G2 SaaS-readiness trio ✅ (#860 / #861 / #869 / #870 / #882 / #883)
B — Enterprise ontology framework (mother-level)	B.1 audit + B.2 multi-tenant + B.3 plugin API + B.5 enterprise document ontology + G8.a-c mount + SDK.a-c trio	B.5.a-d ✅; B.1 audit + G3/G4/G5/G7 ✅; B.2/B.3 design memos ✅; G8.a-c ✅; SDK.a-c trio ✅ (#875 / #876 / #881); G8.d LOI-blocked
C — Measurement infra carry-over	v0.2.3b S3 cross-model + D-alce + D-2wiki + HR full sweep + arXiv submission + graph-RAG synthesis	Graph-RAG Step 1 ✅ (#864 / #877, +0.41 path_coverage ⭐⭐⭐ n=3); Step 2 driver scaffold ✅ (#885); all others operator-attended
D — LOI-gated (blocked)	Customer-specific NDA/DPA/MSA + ingestion + pilot kickoff + vertical pack (legal) + G8.d capability grant workflow + F.2 CR.e customer theming	BLOCKED until LOI signed (Fork A of v0.6 entry contract)

Additional surfaces landed post-close (not in original v0.5 stream definitions)

Track F.1 — Time-Travel Dashboard quartet (#865 / #878 / #879 / #880): TT.a timestamp picker + TT.b corpus-state diff renderer (audit replay overlay) + TT.c reasoning trail replay at time T + TT.d now-vs-T diff view modal. Surfaces v0.4.2 T5 reconstruct_graph_at + v0.5 G3 reconstruct_corpus_view_at + replay_audit primitives as a single operator dashboard.
Track F.2 — Change Review Workspace quartet (#866 / #867 / #873 / #874): CR.a list page + CR.b detail modal with diff renderer + CR.c contradiction-arbiter visualisation + CR.d approve/reject buttons with G2.a evidence capture wire.
Track C — CSP nonce middleware (#884): core/security/csp_nonce.py per-request nonce primitive + build_security_headers composition seam + request.state.csp_nonce middleware wire-in. script-src flag safe today; style-src flag reserved for UI #6 inline-style migration.

Rule discipline (preserved + new)

Rule	Status at 2026-06-13
CLAUDE.md rule #1 (no domain features until v1.0)	PRESERVED across all 44 cycle + post-close PRs — 4-layer protection contract (code-level capability gate + doc-level "Out of scope" + naming-level domain-agnostic + trigger-level LOI tagging) held throughout
#2 (bench numbers + Quality Delta Card on core/ PRs)	Preserved (all 44 PRs touch UI / lifecycle primitives / security / SDK packaging / docs — none touch `core/retrieval` / `core/graph` traversal / `core/reasoning`)
#3 (self-evolution opt-in)	Preserved
#4 (architecture changes require `architecture` label)	Preserved
#5 (`core/` 20 KB module size)	Preserved for NEW files; five legacy modules grandfathered (largest `core/reasoning/reflect.py` at 29.2 KB — split planned, tracked in v0.6 entry skeleton §5 NEW solo-doable items)
NEW #6 Dogfooding evidence ≠ Dim F evidence	Hashevolution own-use = iteration feedback; Dim F still requires external customer pilot
NEW #7 Enterprise ontology = mother-level	Vertical-pack (legal etc.) waits for LOI or v1.0; B.5 horizontal subtypes ✅ landed
NEW #8 External-facing claims gated on measurement evidence	Zenodo DOI + result.json refs mandatory; Step 1 graph-RAG finding cites n=3 paired evidence per `feedback_n1_verdict_inflation_n3_caught`

Cycle entry → close deliverables (actual status)

v0.5 → v0.6 gate (Dim F)

Dim F gate (≥6 month external customer pilot + measured success metrics per docs/PLATFORM_READINESS.md §3) is NOT cleared. The 2-fork v0.6 entry contract from the v0.6 entry skeleton §4 governs:

Fork A — LOI signed → Track D vertical pack scoping begins. G8.d capability grant workflow + F.2 CR.e customer theming unblock.
Fork B — 6-month no-LOI → reassess strategy. v0.5 mother-platform work stays the baseline; v0.6 cycle re-scopes around the next strategic direction.

Until one resolves, mother-platform hardening continues; vertical content stays BLOCKED per CLAUDE.md rule #1.

v0.5.0 — First Domain Pilot (Fork A of v0.6 entry contract)

Theme: prove the platform contract by running ONE real domain in production for 6 months with one external customer. Moved here from v0.4 so Layer 4 governance lands first — see v0.4 retarget rationale above.

Status note (2026-06-13): this section describes Fork A of the v0.6 entry contract (LOI signed → vertical pack scoping begins). v0.5 cycle close (above) shipped the mother-platform infrastructure that this fork builds on — G8.a-c ontology pack mount mechanism, G1/G2 SaaS-readiness primitives, SDK.a-c trio for third-party pack authors, Time-Travel Dashboard surface for audit operators. The Dim F gate clock starts only when a customer LOI is signed; until then, v0.6 cycle entry waits on the operator's decision per the v0.6 entry skeleton §4.

Required for: a second domain pack (forbidden until this gate passes).

Deliverables

One domain pack chosen from informal candidates (likely packs/legal/ or packs/food/) — selection criteria: signed PoC interest + clear legal liability boundary
Customer onboarding playbook (docs/CUSTOMER_ONBOARDING.md)
External red-team pass on prompt injection (replaces pattern-only defense with ML guard + patterns)
Public eval results in eval/RESULTS.md (mother + first pack)
6-month no-core-regression production track record

Done when

One paying or formal-PoC customer has run the deployment for 6 months with no core code change attributable to their domain needs (only pack-level changes).

Out of scope

A second domain pack
Vertical Product packaging
Public marketplace

v1.0.0 — Production-Grade Mother (~6 months after v0.5)

Theme: make domain branching safe for outsiders. After this gate, external developers can publish their own packs.

Deliverables

HTTPS / SSO / SAML / LDAP — production defaults
Multi-tenancy (per-tenant data isolation, per-tenant pack selection, quota management)
SOC 2 or ISO 27001 readiness assessment
Backup / restore / rollback CLI tested under simulated failure
Prometheus + OpenTelemetry exporters
Public SDK and plugin author guide finalized
Bus factor ≥ 2 (one non-maintainer with full commit/review history)
Annual external red-team schedule established

Done when

A third party (not a customer) builds and publishes a pack against the v1.0 SDK without contacting maintainers.
The platform survives a single-maintainer 30-day absence with no customer-visible regressions.

Out of scope

Vertical Products (separate business decision per domain after v1.0)
Federation across multiple JAMES instances (Beyond v1.0 section)

Scalability — vector compression tier (trigger-gated, NOT version-scheduled)

Entry trigger (none of which is currently met): a real deployed corpus whose float32 vector index exceeds available RAM — concretely > ~1M documents at the default 384-dim, or a pilot that measures a vector-store memory bottleneck. Until a trigger fires this stays a note, not work: at 384-dim even 1M docs ≈ 1.5 GB fits the reference 32 GB machine, so adopting compression now would be optimization against an unproven premise (cf. the cloud-tier S6/S7 shelving lesson).

Priority ladder when the trigger fires — ranked by (memory saved × recall kept) ÷ replay/audit cost, not memory alone, because byte-identical replay (RAB / reconstruct_graph_at(t)) is the crown-jewel constraint:

1. Int8 scalar quantization (~4×) — replay-safe: a fixed, audit-logged scale makes it deterministic; backend-agnostic (quantize at the embedding level before ChromaDB). First and lowest-risk. Entry PR must paste a 5-axis Quality Delta Card (rule #2) — the "Recall within 1%" claim is re-measured on JAMES's own QVT/LRB queries, not imported from the literature.
2. Dimensionality reduction — requires a model swap (the default paraphrase-multilingual-MiniLM-L12-v2 is multilingual and not Matryoshka, so head-truncation is unsafe). Forces full re-ingestion + KO/EN bilingual re-validation (bilingual regression has bitten before — see feedback_d2_v2_softener_bilingual_regression). Separate measurement cycle.
3. FAISS IVFPQ (~10–100×) — strongest memory win but the trained codebook is a replay liability: historical replays must pin their era's codebook. Also a backend change (ChromaDB → FAISS = trust-boundary / docs/ARCHITECTURE.md PR, rule #4). Architecture decision precedes any build.
4. Binary + rescoring — only at very large scale; keeps float32 for rescore (so the "32×" is index-only), and rescore ordering must be made deterministic for audit reproducibility.

Rule #1 is not implicated (horizontal infra, not a domain feature) — this is a priority/trigger question, not a permission one.

Beyond v1.0 — Speculative

After v1.0, growth is by domain accumulation, not core feature addition. See docs/PLATFORM_READINESS.md §4 for the three branching forms (Pack / Distribution / Vertical Product) and selection criteria.

Long-considered, no commitment:

Optional Neo4j backend — migrate from markdown wiki to graph DB, Cypher query support, backward compatibility with markdown (was tentatively v0.3; reframed as post-v1.0 optimization)
Multi-agent system — specialist agents (researcher, coder, security), agent-to-agent communication, task decomposition (was tentatively v0.3; reframed as post-v1.0 capability)
OpenAI-compatible API for drop-in replacement
Streaming responses + Webhook support
Federation: connect multiple JAMES instances
On-device fine-tuning: LoRA adapters per user
Edge deployment: smaller models for embedded use
Plugin marketplace: community-contributed packs
Visual graph editor: web UI for ontology editing
Voice interface: ASR + TTS pipeline

How to Influence the Roadmap

GitHub Issues: feature requests, prioritized by upvotes
Discussions: longer-form proposals
Pull Requests: implement what you need

We prioritize based on:

Security-critical fixes (immediate)
Real-data feedback from users
Strategic alignment with the project's direction
Community contribution (volunteer-friendly tasks first)

Domain-specific feature requests during v0.2 — v0.4 are out of scope. See docs/handovers/v0.2.1-business-track.md §3 for the rationale and the "no parallel domains" rule.

Versioning

We follow Semantic Versioning:

MAJOR.MINOR.PATCH-PRERELEASE
0.x.y versions may contain breaking changes
1.0.0 and beyond will follow strict semver
After v1.0 ships, plugin API gets its own SemVer track with a 12-month deprecation guarantee (see docs/PLATFORM_READINESS.md Gate v0.3 criteria)

Last updated: 2026-05-22 — v0.4 retarget to Layer 4 Lifecycle Semantics + CASCADE/EVENT 분리 axiom 채택. v0.3 cycle 의 Knowledge Cascade (Layer 3 Memory OS) 가 안정화되며, 2026-05-21 사용자 비판 시리즈에서 "Layer 3 alone" 의 governance gap (stale facts / Gate 5 silent reject / manual hierarchy 없음 / event sourcing 부재) 드러남. 후속 architectural insight 로 CASCADE (invalidation propagation) 와 EVENT/TEMPORAL (semantic evolution) 의 명확 분리 axiom 채택 — 두 종류 mutation 을 같은 메커니즘으로 처리하면 contradiction 폭발. 7 영역 (T1–T7, T7 supersede chain 신규) 디자인 메모 (docs/architecture/memory-lifecycle-architecture.md §1.5 + docs/design/v0.4-lifecycle-semantics-roadmap.md §9.5) 작성. v0.4 첫 ship bundle = T1+T7+T2 (EVENT 핵심 + A/B routing). 기존 v0.4 (First Domain Pilot) → v0.5 로 shift. 2026-05-22 i18n sweep 시리즈 (7 PR, ~304 entries, 7 backend 모듈) 가 label_key 패턴을 platform invariant 로 lock-in — v0.4 신규 라벨도 동일 contract 필수. Open issues: 0.

Prior update (2026-05-13): v0.3 진입 정식. Axis 6 두 번째 사용자 게이트도 모집 완료 → 6 axes 모두 통과 → v0.2 → v0.3 gate clear. v0.2.x 가 더한 Change Request 기반 (core/change_request.py + wiki_entity + run_jobs) + dependabot 6 high-severity 마감 + 정공법 3 PR (#252/253/254: 웹 학습 LLM-triple 위임 / UNRESOLVED sweep / 노이즈 cleanup 스크립트). v0.3 Platform Skeleton 트랙 활성 — License / CLA / Plugin API / Knowledge Cascade / CR-E 가 본 사이클부터 deliverable.

FilesExpand file tree

ROADMAP.md

Latest commit

History

ROADMAP.md

File metadata and controls

Roadmap

v0.1.0 — Foundation (released, foundational)

Done

Known Gaps (to address in v0.2)

v0.2.0 — Foundation Hardening (released 2026-05-08, closed 2026-05-13)

Axis 1 — Architecture Separation (P0) ✅

Axis 2 — Evaluation Harness (P0) ✅

Axis 3 — Observability / Tracing (P1) ✅

Axis 4 — Security Boundary (P1) ✅

Axis 5 — Controlled Evolution (P1) ✅

Axis 6 — Real-Data Validation (carries forward from v0.1) ✅

Known cuts from earlier v0.2 plan

Deferred follow-ups (recheck before v0.3)

v0.3.0 — Platform Skeleton (entered 2026-05-13, closed 2026-05-25 at v0.3.3 stabilization)

Plugin contract — the v0.3 core

Change Request — finish the primitive

Knowledge cascade — relation provenance

Governance — license / CLA / monitoring

Carryover follow-ups (from v0.2.x)

v0.3-only follow-ups discovered post-design

Measurement framework track (V3' series, 2026-05-22~ open)

Done when

Out of scope (deferred to v0.4)

v0.4.0 — Layer 4 Lifecycle Semantics (entered 2026-05-25, T1+T7+T2 first bundle shipped 2026-05-27 at Sprint 5)

Original Layer 4 scope (preserved below for the eventual T1/T2/T7 sprint)

Scope — 7 areas across 3 tracks (T1–T7)

Phase plan

Invariants (extending v0.3's 12 invariants)

Cross-cutting contracts (preserved from v0.3)

Done when — ✅ All items satisfied at v0.4.0 (2026-05-27)

Out of scope (deferred to v0.5)

v0.4.1 — T6 Causality Chain (CASCADE extension, shipped 2026-05-28)

Done — T6 5-PR sequence + 4 release-gating invariants

Done — v0.4.0 carry-over dispatch_contradiction ingestion wiring (T2.D, 4 PRs)

Done — QVT α track full closure (6 PRs)

Done — also in v0.4.1

Default-off invariant preserved across every opt-in

Done when — ✅ All items satisfied at v0.4.1 (2026-05-28)

Out of scope (deferred to v0.4.2+)

v0.4.x — α-5 Ablation Matrix Cycle (2026-05-30~31, T0 closed 2026-05-31 15:52 KST)

Dual purpose (per 2026-05-31 user clarification)

Actual sizing (vs the original 18×N=3 ≈ 20h estimate)

External benchmark adoption

PR sequence (24 PRs to date, 2026-05-30 → 2026-05-31, T0 smoke mid-flight)

Cycle deliverables — T0 closure 2026-05-31 15:52 KST

T0 verdict + post-closure (2026-05-31)

Operator decision — T1 / T2 sizing

Open methodology questions for the next cycle

v0.4.x — α-6 Sector × LLM Ablation Cycle (2026-05-31~, Phase 3a in flight)

Phase status (live)

Findings (4 tier-tagged per feedback_finding_size_honest_framing)

Cycle PR index

Closure deliverables (in flight on feat/v0.4-alpha6-phase-3a-closure)

v0.4.x — α-7 Graph Top-K Cycle (post-α-6 closure)

α-7 cycle CLOSURE (2026-06-02) — REJECT

v0.4.x — α-8 Ontology Typed-Filter Cycle (post-α-7 closure)

v0.4.x — Sequencing rule (2026-06-01 strategy handover §4)

v0.5.x — Cycle close + post-close consolidation (2026-06-12 → 2026-06-13)

Four work streams (CLAUDE.md rule #1 preserved — actual status at 2026-06-13)

Additional surfaces landed post-close (not in original v0.5 stream definitions)

Rule discipline (preserved + new)

Cycle entry → close deliverables (actual status)

v0.5 → v0.6 gate (Dim F)

v0.5.0 — First Domain Pilot (Fork A of v0.6 entry contract)

Deliverables

Done when

Out of scope

v1.0.0 — Production-Grade Mother (~6 months after v0.5)

Deliverables

Done when

Out of scope

Scalability — vector compression tier (trigger-gated, NOT version-scheduled)

Beyond v1.0 — Speculative

How to Influence the Roadmap

Done — v0.4.0 carry-over `dispatch_contradiction` ingestion wiring (T2.D, 4 PRs)

Findings (4 tier-tagged per `feedback_finding_size_honest_framing`)

Closure deliverables (in flight on `feat/v0.4-alpha6-phase-3a-closure`)