V9 — Gap Analysis (Metis: A Curious Ape's Guide to Reality)

Status: what is real in V1–V8 today and what is missing to ship V9, with file evidence. Date: 2026-06-13. Method: source-reads of the composed domains (services/metis/, libs/{nyx,kalika,sophia,nous,mnemosyne,isis,yemaya,hathor, iris,meditation}/, DOMAINS/{psyche,euterpe}), cross-checked against the 2026-06-12 platform agentic-content report and V_SERIES.md. Per the repo's adversarial-verification standard, "real" here means a source read confirmed a domain-specific implementation, not a checkbox or a file's existence.

Headline. V9 is unusually de-risked: the education backend, the science kernels, and the learning-science library are already real and production-grade, and the generative factory is real-but-fragmented (the exact finding of the platform's own 2026-06-12 report). What is missing is almost entirely connective tissue and consumer experience — the Atlas knowledge graph, the Hephaestus explorable runtime, the Theia wonder-director, the persona/memory glue, and a consumer-first app — none of which require inventing new science or new models.

1. What is REAL today (the foundation V9 stands on)#

1.1 Metis — a production-grade learning backend (not a prototype)#

services/metis/ is a real FastAPI/SQLAlchemy/Celery service: 16 domain routers, ~174 operations (DOMAINS/metis/specifications.md). Confirmed real:

• Course authoring — Course→Module→Lesson with lifecycle (services/metis/src/metis/models/course.py, api/courses.py).
• Source-grounded generation — full ingest→extract→review→approve→outline→ generate pipeline, with evidence bundles tracing generated content to source passages (services/metis/src/metis/api/source_ingestion.py, services/source_generation_service.py, services/evidence_bundle_helpers.py). This is the seed of Aletheia's grounding gate.
• AI tutoring incl. LIVE VOICE — text + live-voice tutoring with VAD, barge-in, turn-taking, delivery fallback (avatar→voice→text), and integrity modes (teach/hint/practice/do-not-complete-for-me) (services/metis/src/metis/services/tutoring_service.py, schemas/tutoring.py). This is the seed of Chiron.
• Concept graph — directed graph with typed nodes/edges + validation/repair (services/metis/src/metis/services/source_concept_graph_helpers.py). Seed of Atlas.
• Assessment + psychometrics — 5 question types, auto-grading, IRT calibration + DIF fairness from live data (services/assessment_calibration_helpers.py, assessment_fairness_helpers.py).
• Curriculum safety policy + the core-six taxonomy — per-discipline policies for philosophy/religion/psychology/neuroscience/anthropology/astronomy (services/curriculum_safety_policy_helpers.py, learning_subject_taxonomy_helpers.py). Seed of G6.
• Verifiable credentials — Open Badges 3.0 + CLR with cryptographic proof (services/credential_service.py).
• Agent runtime + governance — orchestration DAGs, approval gates, kill switches, champion-challenger rollouts (services/agent_runtime_service.py).

1.2 The science kernels are real#

• Nyx — libs/nyx/ephemeris (Meeus/VSOP87/ELP), libs/nyx/renderer-* (WebGL/Three.js, 10M+ stars at 60fps), star/deep-sky catalogs, events, "time-travel," exoplanets (DOMAINS/nyx/features.md). Web-native — runs in the browser today.
• Kalika — libs/kalika/cas-engine is a real Rust symbolic kernel with both a napi-rs native binding and a wasm-bindgen WASM build, plus reactive notebooks and physics kernels (DOMAINS/kalika/features.md, specifications.md). WASM build = browser-native computed physics.

1.3 The learning science is real#

• Mnemosyne — libs/mnemosyne/core implements SM-2, FSRS, IRT-based CAT, BKT/DKT knowledge tracing, rubric assessment, plus prerequisite-reasoning knowledge graphs and gamification (DOMAINS/mnemosyne/features.md per the domain survey). V9's mastery loop is mostly reuse, not build.

1.4 The grounding & model planes are real#

• Sophia — hybrid retrieval (dense+BM25), 4 vector backends, knowledge graph (15 entity / 27 relationship types), citation-credibility, retrieval evaluation (libs/sophia/{indexing,vectordb,knowledge-graph,citation-analysis, evaluation}).
• Nous — ONNX inference across CPU/GPU/Metal/WebGPU, embeddings, LLM serving, vision (libs/nous/{core,llm,vision}).

1.5 The generative factory is real (but fragmented — see §2.4)#

• Isis (ComfyUI/RunPod, 45 certified workflows, image/video/3D/audio), Yemaya (30+ agents, 76 quality validators, HTN planner, HITL types), Hathor (live LLM-NPC dialogue, worldbuilding), Psyche (real-time avatar/voice/Conductor), Euterpe (music), Iris (agent runtime + model orchestration). All confirmed real by the platform report (AGENTIC_CONTENT_GENERATION_SOTA_REPORT_2026-06-12.md §2).

1.6 Platform, safety, provenance are real#

• V1 governance plane (Agentic AI Studio: AgentRun envelopes, budgets, kill switches, generation ledger), V3 C2PA discipline, V7 Sekhmet safety, the meditation engine libs, and the V2–V8 game substrates all exist.

2. What is MISSING (what V9 must build or fix)#

2.1 The Atlas knowledge graph (NEW — the central missing piece)#

There are three separate graphs today — Metis's concept graph (source_concept_graph_helpers.py), Sophia's knowledge graph (libs/sophia/knowledge-graph), and Mnemosyne's prerequisite graphs (libs/mnemosyne/core) — plus rich domain data (Nyx/Kalika/Nisaba) that is not in any of them. There is no unified "map of reality" and no free-text wonder → concept resolver. This is V9's single largest net-new build. Effort: high; risk: medium (the pieces exist; unification + grounding-pin coverage is the work).

2.2 The Hephaestus explorable runtime (NEW)#

Nyx renders and Kalika computes, but nothing binds a concept node to a manipulable browser explorable, and there is no generative-widget verifier (generate a runnable interactive, then prove it hits its success state before it ships). The explorable-on-demand primitive (V9_SOTA_RESEARCH.md §4) is unbuilt. Effort: high; risk: medium.

2.3 Consumer experience layer (NEW — the biggest UX gap)#

apps/metis/{web,mobile} exist but are instructor/admin-shaped scaffolds, not a consumer product. Missing: the ask-a-wonder front door, the beautiful lesson player, the explorable canvas, the embodied-Chiron UI, the personal Atlas star-map, gamified celebration of mastery, mobile-first/offline, and the "show your sources" transparency UI. (Confirmed by the Metis consumer-gap survey: the backend is done; the delightful consumer UX is not.) Effort: high; risk: low (ordinary product build).

2.4 The generative chain is real but not connected end-to-end#

Per AGENTIC_CONTENT_GENERATION_SOTA_REPORT_2026-06-12.md, the platform's two structural deficits apply to V9 too:

• Seven agent loops, one needed. Iris is the intended shared runtime but the agentic tool-call loop is not yet the single substrate (report §2.5, F2). V9 must consume Iris v2, not add an eighth loop. Dependency on platform Phase A2.
• Yemaya HITL is post-hoc, not mid-pipeline (report §2.2, F5). V9's regenerate-with-direction needs the mid-pipeline pause/redirect/resume.
• Grounding/safety rails exist but aren't bound to generation (report F6). V9's Aletheia gate is exactly the work of binding Sophia grounding + Metis safety policy as a mandatory pass on generated lessons.

2.5 Chiron persona/memory glue (NEW)#

The parts exist separately — Psyche (avatar/voice), Hathor (persona/dialogue), Metis live-voice tutoring, Mnemosyne (memory) — but nothing binds them into one coherent teacher that remembers a learner across sessions and shifts integrity modes. This is integration glue, not new capability. Effort: medium.

2.6 Precise STEM explainer-film path (NEW external)#

No Manim / LLM-to-Manim pipeline exists in the repo; on-demand accurate math/ physics video must be built (driven from verified Kalika derivations) with the render-reliability/self-repair loop the 2026 literature says is required (V9_SOTA_RESEARCH.md §2). Effort: medium; risk: medium (render reliability).

2.7 The Bellona cook path (deferred dependency)#

Game-bridge explorables (learn-by-playing inside V2–V8) need the headless UE cook path, which is types-only today (AGENTIC_CONTENT_GENERATION_SOTA_REPORT_2026-06-12.md §2.4: "no headless RunUAT invocation exists … every generated asset dead-ends here"; corroborated by V2's single .uasset, V4's 770 JSON stand-ins). V9 P1 deliberately does not depend on this (web-first; §2.3). It is a P2/P3 motion shared with V8.

3. Gap summary table#

4. Risk-ranked verdict#

1. Lowest-risk, highest-value (do first): the consumer experience layer + the Aletheia grounding gate over the existing Metis generation — this alone is a shippable, trustworthy "ask a wonder, get a grounded lesson" product.
2. Core differentiators (the wedge): the Atlas graph and the Hephaestus computed-kernel explorables (Nyx/Kalika in the browser) — these are what no competitor has, and both build on real, web-native kernels.
3. Delight multipliers: Chiron's face/memory and the precise explainer-film path — high wow, medium effort, not P1-blocking.
4. Deferred: the game-bridge / Bellona cook path — explicitly P2/P3, shared with V8, never on the P1 critical path.

The honest bottom line matches the platform's own 2026 finding: the vision is already specified and largely implemented at the component level; the work that remains is integration, grounding, and a delightful consumer surface — not invention.