Vault Glossary

Technical terms used across the hinata-v2 vault. Maintained by Hinata (the General).

Vault Architecture

• hinata-v2: the vault — primary knowledge base at /Users/nnamdi/hinata-v2/. Git-committed, locally stored.
• supreme-court/: strategy, laws, governance, preferences, and lifecycle rules that govern the vault.
• the-government/: diataxis-classified outputs from all federation work (reference, how-to, tutorial, explanation).
• federation/: flat directory containing all agent context (_context.md) and agent definitions (_agent.md). No subdirectories except demoted/.
• inbox/: handover files, ingested content, pending classification. Entry point for all external input.
• Sandpit: /Users/nnamdi/Sandpit/hinata/ — scripts, applications, logs, audio. Everything that cannot live in the vault.
• diataxis: the 4-type documentation classification framework (reference, how-to, tutorial, explanation) governing all output to the-government/.
• standing rule: a permanently active instruction in CLAUDE.md that fires every turn without explicit invocation.
• incremental verbatim transcript: the standing rule that appends every exchange (user, agent dispatch, agent result, Hinata response) to the CLI's transcript file per-turn for crash resilience.
• Z2: the Mac Mini M2 running Proxmox — the always-on data plane for pipelines, pollers, and infrastructure. Accessed via Tailscale.
• Z2 container: a Proxmox LXC/VM on Z2 running a specific service (e.g. jimmy-neutron_brain-ops container).

Pillars and Command Structure

• FORGE: the pillar for career, learning, coding, music, reading, productivity, writing, and mastery. Colonel: Levi Ackerman.
• FOUNDATION: the pillar for work, health, fitness, finances, home, sleep, legal, security, and philosophy. Colonel: Saitama.
• FLOW: the pillar for transport, networking, side-hustle, creative strategy, sacrifice/opportunity-cost, end-state, culture, and vows. Colonel: Makima.
• FIRE: the pillar for friendship, adventure, entertainment, nutrition, fashion, dreams, legacy, and partnership. Colonel: Goku.
• SYNTHESIS: the pillar for art, NLP, marketplaces, audio, research, investment, and surveillance. Colonel: Sung Jinwoo.
• Colonel: pillar-level synthesis owner who coordinates all commanders within a pillar. The "Spirit Bomb" — deploys accumulated knowledge of all its commanders.
• Captain: cross-cutting operational rank above Commanders. Captains own intake, routing, infrastructure, and evolution — not a single life domain.
• Commander: domain specialist — the primary execution unit. ~40 across all pillars.
• Lieutenant: subordinate operational agent under a Captain. Gates and enriches Commander intake.
• signal routing: the process of directing detected signals or extracted output to the appropriate commander or system.
• stage-gate: a decision point between pipeline stages where output is validated before moving forward.

Commanders and Agent Roles

• Sung Jinwoo: SYNTHESIS colonel — owns intelligence pipelines, voice memo synthesis, and cross-commander signal merging.
• Levi Ackerman: FORGE colonel — owns career, learning, coding, music, and productivity domains.
• Saitama: FOUNDATION colonel — owns work, health, finances, home, and security domains.
• Makima: FLOW colonel — owns networking, transport, side-hustle, creative strategy, and vows.
• Goku: FIRE colonel — owns friendship, adventure, entertainment, and legacy domains.
• Nujabes: SYNTHESIS commander for audio sourcing, curation, and content quality.
• Heimerdinger: SYNTHESIS commander for NLP, ASR transcription, and extraction pipeline design.
• Zepile: SYNTHESIS commander for marketplace discovery, ecosystem mapping, and arbitrage signals.
• Madara: SYNTHESIS commander for market surveillance and macro signal detection.
• Hashirama: SYNTHESIS commander for capital allocation, portfolio growth, and investment theses.
• L (Lawliet): SYNTHESIS commander for research methods, novel approaches, and scientific rigour.
• Deidara: SYNTHESIS commander for visual art, creative coding, and code-art.
• Zuko: FORGE commander for career intelligence, job applications, and professional positioning.
• Shikamaru: FORGE commander for learning, quiz-based study, SRS, and skill acquisition.
• Trunks: FORGE commander for coding craft, software engineering, and technical builds.
• Squidward: FORGE commander for music production, virtuosity, and audio creative work.
• Ging: FORGE commander for reading, book synthesis, and knowledge extraction.
• Sasuke: FORGE commander for productivity, focus, and deep work.
• Jiraiya: FORGE commander for journaling, creative writing, and lyrics.
• Hisoka: FORGE commander for mastery tracking and level-up progression.
• Kakashi: FOUNDATION commander for work performance, VMO2, and professional engineering.
• Allmight: FOUNDATION commander for mental health, wellbeing, and burnout recovery.
• Zoro: FOUNDATION commander for fitness, training, and physical performance.
• Bulma: FOUNDATION commander for finances, budgeting, banking, and ISA management.
• Simba: FOUNDATION commander for home, property, and life infrastructure.
• Snorlax: FOUNDATION commander for sleep, rest, and recovery routines.
• Light Yagami: FOUNDATION commander for legal literacy, contracts, and rights.
• Itachi: FOUNDATION commander for digital security, credentials, and operational security.
• Thorfinn: FOUNDATION commander for philosophy, ethics, and values.
• Melfi: FOUNDATION commander for clinical-frame psychology, sovereignty audits, and anthropology.
• Lelouch: FLOW commander for networking, relationship management, and social signals.
• Dragonite: FLOW commander for transport, commuting, and route optimisation.
• Rock Lee: FLOW commander for side-hustle and income-opportunity signals.
• Uncle Iroh: FLOW commander for creative strategy, long-term narrative, and contingency planning.
• Erwin Smith: FLOW commander for opportunity-cost audits, project kill decisions, and sacrifice.
• Cell: FLOW commander for end-state planning, destination coordination, and convergence.
• Hange: FLOW commander for culture, intellectual curiosity, and exploration.
• Kurapika: FLOW commander for personal vows, commitments, and accountability.
• Killua: FIRE commander for friendship, social skills, and interpersonal dynamics.
• Luffy: FIRE commander for adventure, spontaneity, and experience-seeking.
• Brook: FIRE commander for entertainment, media consumption, and cultural intake.
• Sanji: FIRE commander for nutrition, cooking, and food strategy.
• Black Panther: FIRE commander for fashion, personal style, and visual identity.
• Naruto: FIRE commander for long-term dreams, vision, and aspiration.
• Mufasa: FIRE commander for legacy, estate planning, and 20-year wealth trajectory.
• Roxanne: FIRE commander for romantic partnership and relationship growth.
• Jimmy Neutron: Captain — system infrastructure, vault health, scripts, plugins, agent schemas. Owns all Bash execution.
• Canary Organiser: Captain — inbox pipeline owner. Routes incoming items: BACKLOG / ASSIMILATE / ROUTE / HOLD / DISCARD.
• Orochimaru Scout: Captain — pipeline health, daily/weekly/monthly audits, improvement loops.
• Meruem Evolution: Captain — federation evolution tracking, quality arbiter, system maturity progression.
• Minato: Canary lieutenant — morning briefing, proactive surface, scheduling.

Workflow and Pipeline Terms

• voice memo pipeline: the staged workflow that processes voice memo transcripts from ingestion to synthesis and routing.
• Audio-Career Workflow: the career-focused extension of the voice memo pipeline that produces career intelligence and networking content.
• Ingestion: the stage that reads transcript files and applies ASR post-correction.
• Routing: the stage that classifies a memo by domain and assigns primary and secondary commanders.
• YAML extraction: the stage that extracts structured entities from transcript text into YAML.
• Synthesis pass: cross-memo consolidation that produces combined career and networking intelligence.
• career intelligence routing: the process that forwards career-focused output to the Zuko workflow.
• networking signal routing: the process that forwards relationship and named-person signals to the Lelouch workflow.
• Iroh strategy pass: the final synthesis stage that produces a long-term career narrative and contingency plan.
• Auto-Extraction: deterministic extraction logic that uses regex and vocabulary lists rather than AI inference.
• Domain Vocabulary: a curated list of technical vocabulary terms used to detect domain-specific language in transcripts.
• content keyword scan: a fallback routing method based on keyword hits when path-based classification is unavailable.
• path-based category detection: routing based on known directory or filename path segments.
• tie-break: a deterministic rule used when multiple routing categories score equally.
• raw_* fields: fields like raw_person_candidates, raw_technical_terms, and raw_action_candidates that store verbatim auto-extracted candidates.
• scaffolded fields: placeholder YAML fields with PENDING or empty lists reserved for downstream commander enrichment.
• priority memo set: the initial batch of high-value transcripts selected for early pipeline processing.
• speaker sentiment arc: aggregated sentiment tracking across an audio memo or batch of memos.
• named entities: extracted people, organizations, and other proper nouns from transcripts.
• decision points: transcript moments where choices or commitments are explicitly made.
• action items: tasks or follow-up work explicitly identified in transcripts.
• language_detected: the detected language label applied during transcript processing.
• transcript_source: the origin of a transcript, e.g. apple-openai or whisperx.
• synthesis_output: the output vault path for consolidated synthesis documents.
• whisperX detect_language=True: an ASR configuration option that enables automatic language detection.
• wav2vec2 forced alignment: a technique for producing word-level and phoneme timestamps from audio.
• SKIP_EMPTY log: a pre-flight check that skips transcripts smaller than a threshold.
• ASR post-correction: rule-based correction of speech recognition errors after transcription.
• REVIEW_REQUIRED: a flag used when a correction is ambiguous and should be verified by a human.
• PENDING: a placeholder value inserted for downstream fills in YAML scaffolding.

AI, NLP, and Machine Learning Terms

• AI: artificial intelligence, especially the generative and inference tooling referenced in project docs.
• LLM: large language model, used for text generation and contextual inference.
• ASR: automatic speech recognition, the technology used to transcribe voice memos.
• NLP: natural language processing, the field of extracting meaning from text.
• NER: named entity recognition, the extraction of entities such as names, dates, and organizations.
• BGE-small: a BAAI sentence embedding model used for local semantic similarity in email intelligence.
• BERTopic: a topic modeling algorithm used to discover clusters in email corpora.
• LogReg: shorthand for logistic regression classifier used in email recommendation and classification.
• LogisticRegression: the specific sklearn classifier model used to predict email categories.
• FAISS: a vector similarity search library used as a local vector store.
• spaCy: an NLP library used for NER and text processing.
• sentence embeddings: vector representations of text used for similarity and classification.
• topic ID: a discrete cluster label produced by BERTopic.
• calendar_signal: a boolean flag set when an email contains event or date/venue signals.
• delete confidence score: a numeric estimate of how safe it is to delete an email.
• priority scorer: a ranking mechanism that assigns email priority from 1 to 5.
• confidence calibration: model calibration technique such as Platt scaling.
• uncertainty ensemble: a combined model disagreement score used to identify uncertainty.
• active learning query ordering: the process of selecting high-uncertainty examples for review.
• centroid: the geometric center of a cluster, used in deleted-email profile analysis.
• reply-chain lookup: thread graph traversal used to evaluate email reply context.
• multi-label classifier: a model that can assign multiple labels to a single example.
• local ML: machine learning that runs locally without external API token usage.
• zero LLM tokens per email: a design constraint for the email intelligence pipeline.

Email Intelligence Terms

• Hinata Email Intelligence System: the multi-layered architecture for email classification, recommendation, and sweep.
• Layer 1: structured classification using deterministic sender and header rules.
• Layer 2: unstructured classification using embeddings, topic modeling, and local ML.
• Layer 3: recommendation engine that weights signals and actions over time.
• Layer 4: API endpoints that expose feed, digest, events, sweep, and alerts.
• sender rules: deterministic patterns used for immediate email categorization.
• domain allowlist: trusted domains exempt from deletion and special handling.
• structured categories: deterministic categories such as career, finance, housing, and events.
• marketing signals: textual patterns that indicate promotional or newsletter content.
• feedback loop: the process of using user actions to improve recommendations.
• FastAPI: the Python web framework used for email service endpoints.
• Telegram bot: the messaging integration used to deliver alerts and summaries.
• Studio: the workspace interface that consumes email intelligence outputs.
• tasks.json: the task registry used for routing signals and capturing work assignments.
• michael_id: the unified identity field used for cross-account email deduplication.
• Message-ID fingerprint: a deduplication key derived from message metadata and subject.

Execution Modes

• runtime: execution during an active Claude Code session. Mac is on, harness is connected. Runtime services can use Mac LaunchAgents or in-session tools.
• BAU (business as usual): execution when Mac is offline (closed or sleeping). Only Z2 systemd services run. All always-on services must be BAU-classified and live on Z2.
• always-on service: a BAU service that must survive Mac sleep/shutdown. Examples: Telegram bot, chat-audit timer, bw session renew. Must be Z2 systemd, never Mac LaunchAgent.
• runtime service: a service only needed during active sessions. Examples: wrangler dev server, Studio frontend. Can be Mac LaunchAgent.

Session Lifecycle and Harnesses

• harness: a CLI tool that connects to the vault. Five harnesses exist: Claude Code, Cloud, Antigravity, Codex, Gemini.
• Claude Code: sole writing authority for the vault. Structural, strategic, high-stakes.
• Cloud: long-running/unattended ephemeral clone tasks. Never pushes to main.
• Antigravity: vision and agentic build — active building, code modification, ingestion.
• Codex: execution and mastery — high-speed scripting, DevOps. Does not write to vault.
• Gemini: diagnostics and research support harness (deprecating).
• /handover: unified session close — verifies transcript, reconciles state, writes re-attach point, git push.
• token economy (3-band law): Commander work 40-70%, state management 20-40%, Hinata main thread 10-20%.
• dont-halt: DEPRECATED (retired 2026-06-08). Was an autonomous execution mode. Now replaced by the permanent acceptEdits posture — halt only on canonical bottlenecks.
• PostCompact: context compression event that fires when conversation approaches token limits.
• crash-resilient capture: writing data to disk before operations complete, so crashes don't lose prior work.

Tools, Formats, and Infrastructure

• API: application programming interface.
• CLI: command-line interface.
• JSON: JavaScript Object Notation, a data interchange format.
• JSONL: newline-delimited JSON format.
• YAML: human-readable data serialization format used for extraction artifacts.
• CSV: comma-separated values.
• FastAPI: Python framework used to expose HTTP endpoints.
• BigQuery: Google Cloud data warehouse used in project examples.
• PostgreSQL: open-source relational database.
• Cloudflare: cloud provider — tunnels, DNS, edge hosting for michael-engineer.dev.
• Tailscale: zero-trust mesh VPN connecting Mac, Z2, and all infrastructure nodes.
• Proxmox: virtualization platform running on Z2. Hosts LXC containers and VMs.
• LXC: Linux container runtime on Proxmox.
• KVM: kernel-based virtual machine on Proxmox.
• NAS: network-attached storage.
• NFS: network file system protocol.
• SSH: secure shell protocol.
• VM: virtual machine.
• SSD: solid-state drive.
• RAM: random-access memory.
• GPU: graphics processing unit.
• Mermaid: diagramming language rendered in markdown. Encouraged for visual output in the vault.

Language, Learning, and Content Terms

• IgboMastery: the compound sprint project focused on Igbo language learning infrastructure.

• Lingopie: the external language learning stack used as a design inspiration.

• SRS: spaced repetition system used for quiz-based learning.

• tokenisation: the process of splitting text into tokens or words.

• WER: word error rate, a speech recognition accuracy metric.

• Common Voice: Mozilla speech dataset used for speech recognition benchmarking.

• word frequency corpus: a corpus used to rank tokens by frequency.

• frequency floor: a threshold below which rare tokens are skipped.

• Creative Commons: a licensing category for reusable content.

• fair use: a legal doctrine for limited reuse of copyrighted material.

• scraping ToS: terms of service constraints that govern automated content discovery.

• source discovery map: a catalog of discovered content sources and channels.

• licensing risk map: a risk assessment of sources by licensing posture.

• curated content map: a selection of audio content chosen for pedagogical use.

• listening notes: qualitative notes on audio quality, dialect, and ASR suitability.

• accent register: dialect classification and pronunciation quality notes.

• PostgreSQL schema: the database schema structure for storage and application data.

• React player: audio playback UI component in the learning stack.

• Z2 NLP container: the NLP container on Proxmox for ASR and language processing.

Sourcing, Arbitrage, and Capital Terms

• income-arbitrage: the practice of finding income opportunities across marketplaces and labour platforms.
• resale arbitrage: identifying underpriced listings for resale profit.
• marketplace discovery: the process of discovering new earning surfaces on platforms.
• capital allocation: deciding how much money to deploy against an opportunity.
• arbitrage window: a time-bound mispricing signal that can be acted on quickly.
• capital signal: a macro market signal used to update allocation decisions.
• dual signal: a signal that is both actionable sourcing and capital relevant.
• verify: a research validation flag requiring L sign-off.
• sourcing brief: a structured recommendation for a specific marketplace move.
• value thesis: the rationale behind an investment or sourcing opportunity.
• time window: the urgency or expiry horizon for a flagged opportunity.
• micro path: Zepile’s path from sourcing brief to Hashirama deployment.
• macro path: Madara’s broader market signal route to Hashirama.
• position: the amount and target structure of a deployed capital allocation.
• target return: the expected return from a deployed position.
• exit trigger: the condition that prompts a position exit.
• stop condition: the condition that limits downside or pauses a decision.
• marketplace surface: the channel or platform where a signal is observed.
• Depop, Vinted, eBay, Discogs, Upwork, Contra, Toptal, Fiverr, StockX, GOAT, WatchCharts, Chrono24, GitHub Marketplace, RapidAPI, Musicbed, Artlist, Pond5: marketplace surfaces referenced in the pipeline.
• Rightmove, Zoopla, SpareRoom, OpenRent: property-search surfaces owned by Simba.

Project-Specific Systems and Names

• Audio-Career Workflow: the career intelligence extension of voice memo processing.

• IgboMastery: the Igbo language learning infrastructure project.

• Madara-Zepile-Hashirama Pipeline: the signal flow linking surveillance, marketplace discovery, and capital allocation.

• Deterministic Local Excel Reconciliation: a protocol for local, deterministic spreadsheet reconciliation.

• CLAUDE.md: root entry point for the vault. Contains profile, routing tables, standing rules, and governance.

• michael-engineer.dev: the public portfolio site. Must always return HTTP 200 (Severity 1 if not).

• studio.michael-engineer.dev: the authenticated workspace interface. Must be behind auth middleware.

• hinata-sandpit: the Sandpit git repo at /Users/nnamdi/Sandpit/hinata-sandpit/.

• hinata-brain: legacy name for the vault (now hinata-v2). The old iCloud path is blacklisted.

Communication and Metadata Terms

• task #: a unique task identifier, e.g. z0b488, c0y917, 47w70z.
• tasks.json: the task registry used for routing signals and capturing work assignments.
• vault: the shared workspace where generated artifacts are stored (/Users/nnamdi/hinata-v2/).
• cloud-writable: a path or artifact that can be written from Cloud sessions (branch-based, never main).
• source tier: priority level of a transcript or dataset.
• evidence chain: the research trail and validation record used for sign-off.
• confidence band: a range used to express model or ASR confidence.
• transcript: per-CLI .md file in the-government/information_reference/reference_transcripts/, verbatim, 7-day rolling.
• handover file: inbox/claude-handover-[dd-mm-yy].md — cross-session state transfer, auto-processed by Canary.

Inbox Pipeline

• Canary Inbox Pipeline: the raw capture-to-routing flow for inbox/ items. First stage of the brain's content processing.
• Canary filter question: the core decision rule: does this item bring Michael an opportunity to grow?
• BACKLOG: inbox outcome for actionable growth opportunities → task stubs.
• ASSIMILATE: inbox outcome for knowledge that improves a Commander's context → appended to context or knowledge-base.
• ROUTE: inbox outcome for reference documents with ongoing value → the-government/.
• HOLD: inbox outcome for structural blockers requiring human input → inbox/pending-review/.
• DISCARD: inbox outcome for items failing the enrichment-check → inbox/done/ with justification.
• Enrichment-check: mandatory rubric before HOLD or DISCARD — asks whether the item exposes a gap, intent, or scoping question.
• Plan-First Parallel-Execute Protocol: default Canary workflow — survey in parallel, present as decision table, execute concurrently.
• Assimilate-not-route rule: prefer merging signal into canonical Commander files rather than creating duplicates.
• Raw Telegram message: never stored directly in federation; must be assimilated or discarded after extraction.
• Nami: routing lieutenant — scores items against Commander contexts.
• Robin: intent extraction lieutenant — adds structured metadata before routing.
• Vegapunk: processing lieutenant — maintains strategies for complex content types.
• Jinbe: hold audit lieutenant — logs HELD/FLAGGED/DISCARDED items.
• Rick Sanchez: pattern memory lieutenant — logs routing decisions, surfaces automation opportunities.
• Toph: Meruem's lieutenant — grounding, quality arbiter.