KB Promotion Audit

2026-05-23 · scope: all hub-linked apps + lifelabs-ai repos + ~/LIFE + Obsidian vault · target: kb-service (company / contemplative / benchmark)

Survey of every hub-linked app, supporting repo, and local content store to identify durable reference material, prompts/protocols, user-generated insights, and structured reference data that should move into the centralised KB. Eight sub-agents surveyed in parallel; decisions tagged per artifact.

MOVE — promote as-is TRANSFORM — extract / reformat first LEAVE — stays in source DUPLICATE-RISK — verify before move HOLD — wait on a decision

Overview

~120
artifacts triaged
~75
MOVE / TRANSFORM
3
major dedup risks
5
ingest phases proposed

The biggest content payloads are the ~100 meditation-path markdown files (Stages 1–2) and the 11 expert judge rubrics (v2.x). Both have clean canonical homes already and need straightforward bulk ingest — but the judge rubrics exist in two places (see dedup section). The next tier is the ~25 LIFE specs under ~/LIFE/specs/ which are the canonical company knowledge currently living only on your laptop. After that, the Wisdom system prompt, hub HTML reports, and SOME conceptual framework round out the high-value MOVE set.

Most app code, UI scaffolding, sprint working docs, personal life material, and Mu/GTD state stays put. The Obsidian vault contributes only ~3 documents — most of it is appropriately personal.

Top wins (ranked)

  • Movecontemplative Meditation-path content — ~/LIFE/content/build/meditation-path-md/ (~100 files, Stages 1.1–2.4: explanation / meditation / technique / faq per practice). Canonical teaching corpus; clean structure ready for bulk ingest.
  • Movebenchmark 11 expert judge rubrics (v2.x) — ai-safety, coaching, depth-psychology, dzogchen, gestalt, kundalini-tantra, mahayana, neidan, theravada, western-mystical, zen. ~70–100KB each, detailed 10-level scoring, tradition-grounded. See dedup note: two copies exist.
  • Movecompany LIFE specs (~25 docs) from ~/LIFE/specs/ — KB spec itself, unified-prompt-registry, life-data-lab + runbook, life-app-spec, classifier spec, master-roadmap, identity-consolidation plans, etc. The company's reference shelf.
  • Movecompany Wisdom system prompt — wisdom-system-prompt repo (v0.7 current, v0.8.x drafts). Dual-source: repo stays the authoring source of truth, KB mirrors.
  • Movecontemplative Wheel of Life framework — 10 stages + 46 sub-elements + 17 traditions with descriptions, extracted from stages-of-meditation-eval seed.sql. Verify no overlap with retired wheel-wiki content already in KB.
  • Movebenchmark Judge prompt templates — candidate-quality-judge.md, trajectory-analysis-judge.md, pairwise-comparison-judge.md, classifier-disagreement-judge.md (review-ui). Define the evaluation output contract.
  • Movecompany Six hub HTML reports — golden-bot-architecture, golden-point-progress, kb-executive, six-month-roadmap, v0.7→v0.8 comparison, v0.8.2-vs-v0.8.3-gate, life-data-lab-report. Convert HTML → MD on ingest.
  • Movebenchmark Question expansion methodology — golden-questions-manager/specs/question-expansion.md. Documents the durable pipeline for scaling the golden-questions corpus.

Dedup & conflicts

1. Judge rubrics exist in TWO places. Found at both golden-prompt-tests/expert-prompts/{judge}-v2.1.md and review-ui/server/src/prompts/judge-rubrics/{judge}-v2.md. Before ingest, pick one canonical source (recommend whichever has the latest version tag and active editing) and treat the other as a downstream consumer. Otherwise KB will hold two near-identical copies of each rubric.
2. SOME framework vs. retired wheel-wiki content. The KB contemplative corpus was seeded 2026-05-22 from the retired wheel-wiki-local (Bodhicaryāvatāra chunks + source summaries). The stages-of-meditation-eval seed.sql contains 10 stages + 17 traditions + 46 sub-elements that may overlap with descriptions already ingested. Run a slug check (wheel-wiki/<chapter>/<chunk>) before adding stage/tradition pages.
3. LIFE Labs pipeline reports auto-publish. Per KB baseline: golden-prompt-tests narrative reports, data-lab outputs, bodhisattva-bench results, and executive-report all auto-publish to corpus=company via POST /api/ingest/lifelabs-report. The hub HTML reports listed under "Top wins" may overlap with these — verify by report_id + source before bulk-ingesting the static HTMLs.
Already in KB (do not re-ingest):
  • 16 Notion-seeded company docs (2026-05-23 via new kbd notion-bulk-import): Vision, White Paper, GPT 2026, Hypothesis, Phases, Hubs Analysis, Twitter accounts, FLIP research Programme, Next Steps to 2024, 6 "Interesting People" notes, TSSC 2025 Concurrent Sessions. doc_id = kb_notion_<slug>.
  • BodhisattvaBench spec (16-criterion framework, benchmark corpus).
  • Wheel-wiki Bodhicaryāvatāra chunks + source summaries (contemplative corpus, nightly cron).
Jira is NOT in KB scope. Decided 2026-05-23: Jira/Confluence/Bitbucket access goes through the Atlassian Rovo MCP Server, not through KB ingest. The existing kbd jira-bootstrap / jira-sync subcommands are dead code pending cleanup.
Confidential Notion docs deliberately skipped (Trademark + lawyers, International teams/Sevilla/Germany recruiting, thesis jury notes) — SystemScheduledActor caps at internal. If wanted in KB, ingest via kb-web under an authenticated human actor.

life-hub

/Users/fionnenglish/LIFE/life-ops/life-hub — hub.life-labs.dev — admin/landing site + reports

Reports (reports/)

  • Movecompany golden-bot-architecture.html — technical architecture of the judge system; durable agent-design reference.
  • Movebenchmark golden-point-progress-report.html — comprehensive benchmark progress + per-scenario results (~7k lines).
  • Movecompany kb-executive-report.html — KB design rationale & architecture.
  • Movecompany six-month-roadmap.html — high-level strategic roadmap.
  • Movecompany v0.7-to-v0.8.x-comparison.html — Wisdom prompt version comparison (evolution history).
  • Movebenchmark v0.8.2-vs-v0.8.3-gate.html — gate eval between prompt versions.
  • Movebenchmark life-data-lab-report.html — Data Lab infra & methodology.
  • Leave weekly-update-feb26.html — time-bound status snapshot; decays.
  • Leave ingest-wiki-report.html — one-off operational audit.

Admin / registry / wiring

  • Leave admin/prompts.html + matrix/edit views — UI for the unified prompt registry. The actual prompts live in repos + Supabase; the hub is just the admin surface.
  • Leave models-apps.json, prompts-apps.json — app config / healthz wiring.
  • Leave functions/scheduled.ts — retention pruning; operationally important, not reference.
  • Leave skills/index.html — marketplace UI; lives where consumed.

Other

  • Movecompany docs/hub-auth-integration.md — integration guide for hub-auth.js; useful for any tool consuming hub identity.
  • Leave CLAUDE.md, migrations/README.md, shared/top-bar.css — repo-internal scaffolding.
Recommendation: Move 7 high-value reports (6 HTML + 1 MD guide). Convert HTML → MD on ingest for search consistency. Verify the report set against the LIFE Labs pipeline ingester first — some may already be auto-published.

Golden suite (3 repos)

Combined survey of golden-prompt-tests, golden-questions-manager, golden-bot

golden-prompt-tests

  • Movebenchmark scenarios.json — canonical scenario set (GQ1…N) with stage, practice_type, user_message, test_dimension, source, safety level.
  • Movebenchmark wisdom-system-prompt-v7.md — frozen Wisdom prompt snapshot used for eval reproducibility.
  • Dup-riskbenchmark expert-prompts/{judge}-v2.1.md × 11 — see dedup note: also present in review-ui. Pick canonical source before ingest.
  • Leave Archived {judge}-v1.md short-form originals (superseded).
  • Leave specs/prompts-pointer.json — build metadata.
  • Transformbenchmark results-api/README.md + DEPLOYMENT.md — extract operational runbook portion.
  • Transformcompany CLAUDE.md evaluation methodology section — extract as standalone methodology doc.

golden-questions-manager

  • Movebenchmark specs/question-expansion.md — collection-pipeline methodology (sources, dedup, output structure).
  • Transformbenchmark migrations/0001_initial_schema.sql — extract canonical golden-question schema as reference doc.
  • Leave D1 questions database — live operational store; mirror its shape to KB, not its rows.
  • Leave scripts/*.json ephemeral build artifacts; specs/tutorial-walkthrough.md UI spec.

golden-bot

  • Transformbenchmark DEPLOYMENT.md — Results-API + Questions-API + Slack + Railway runbook; extract as "Benchmark Infrastructure Deployment Guide" (combine with golden-prompt-tests deployment).
  • Transformbenchmark src/tools/definitions.ts — extract the 15 tool definitions as a reference page.
  • Movebenchmark Architecture diagram from README — Slack → Express → OpenRouter → APIs flow.
  • Leave README dev setup, CLAUDE.md project guidance.
Recommendation: The 11 judge rubrics are the biggest single asset, but resolve the dedup with review-ui first. Combine all deployment runbooks into one "Benchmark Infrastructure" page rather than promoting each repo's docs separately.

review-ui + Data Lab

/Users/fionnenglish/LIFE/benchmarking/review-ui — review.life-labs.dev

Evaluator prompts & rubrics

  • Movebenchmark server/src/prompts/candidate-quality-judge.md — full-panel audit template; defines JSON output schema, multi-turn addendum, trajectory flags.
  • Movebenchmark server/src/prompts/trajectory-analysis-judge.md — multi-turn branch evaluator; turn-score aggregation, recovery-quality, drift flags.
  • Movebenchmark server/src/prompts/pairwise-comparison-judge.md — head-to-head A/B preference schema.
  • Movebenchmark server/src/prompts/classifier-disagreement-judge.md — classifier-calibration protocol.
  • Dup-riskbenchmark server/src/prompts/judge-rubrics/{judge}-v2.md × 11 — see dedup note: also in golden-prompt-tests.

Seed / reference content

  • Movebenchmark supabase/seed-data/wisdom-v7-production.md — canonical seed/fallback Wisdom prompt for benchmarks.

Other

  • Transformcompany CLAUDE.md §47–50 (Data Lab Supabase architecture / contamination failure mode) — extract as standalone "Data Lab Supabase Architecture" reference; rest stays.
  • Leave src/components/ExpertAnalysis.tsx — UI rendering; methodology (consensus weighting, verdict colors) would be worth extracting if it's documented anywhere.
Recommendation: Move the 4 judge-prompt templates + Wisdom v7 seed immediately — these are unambiguously canonical evaluation contracts. The 11 rubrics are blocked on the dedup decision. The Data Lab spec content is already covered by ~/LIFE/specs/life-data-lab.md (see local LIFE section).

stages-of-meditation-eval

/Users/fionnenglish/LIFE/benchmarking/stages-of-meditation-eval — stages-of-meditation-eval.life-labs.dev — phenomenological assessment dashboard

Conceptual framework (from supabase/seed.sql)

  • Movecontemplative 10 Wheel-of-Life stages — Emergence → Intelligence → Meditation → Mastery → Awakening → Unbinding → Enlightenment → Blazing Bliss → Actualization → Infinity. Each with description + color.
  • Movecontemplative 46 sub-elements across the 10 stages (Survival, Reproduction, …).
  • Movecontemplative 17 spiritual traditions — 9 Eastern, 4 Western, 2 Indigenous, 2 Modern — with category, color, description.

Methodology / spec

  • Dup-riskcompany specs/stages-of-meditation-eval-dashboard-spec.md — authoritative description of stages, traditions, marker categories. Already in ~/LIFE/specs/ survey; dedup to one canonical location.
  • Leave specs/kb-ingest-integration.md — planning doc for routing stages-of-meditation-eval uploads to KB; infrastructure-future.
  • Leave specs/wheel-canvas-import-* — one-time migration notes.
  • Leave README.md, CLAUDE.md — infra / app-internal.

Live data

  • Hold 170 cells (10 × 17 grid) — currently empty content_pages, no markers authored yet. No reference content to move. Once cells get authored, set up a one-way sync to KB.
  • Leave resources, resource_tags, markers tables — all empty / scaffolding.
  • Transformcontemplative BENCHMARK_CONTEXT string in /api/resource-ai-suggest route — extract the embedded framework summary as KB reference (or simply have the route fetch it from KB).
Recommendation: The Wheel framework (stages + sub-elements + traditions) is a single small high-value page. Check carefully against existing contemplative corpus (wheel-wiki seed) before adding — there's likely partial overlap on tradition descriptions. The cell-content sync is a Phase 2 concern.

life-app + wisdom-system-prompt

Combined: wisdom-system-prompt repo + ~/LIFE/life-app + ~/life-app (sibling)

wisdom-system-prompt repo

  • Movecompany wisdom-system-prompt-v0.7.md — current production system prompt; governs Wisdom (LIFE AI teacher) behaviour.
  • Hold wisdom-system-prompt-v0.8.x drafts — promote to KB once a v0.8.x version is promoted to production.
  • Movecompany wisdom-operational-context.md — runtime context injection template; needed alongside the system prompt.
  • Movecontemplative wisdom-prompt-comparison.md — v1.0 → v2.0 evolution + BodhisattvaBench alignment.

life-app — meditation content

  • Movecontemplative Stage 1.1–1.6 + Stage 2.1–2.3 practice scripts — ~/LIFE/content/build/meditation-path-md/ (cleanest build). Each practice has explanation/meditation/technique/faq markdowns. Strip frontmatter and Drive links on ingest.
  • Hold Stage 2.4 (Intrinsic Awareness) — placeholder files per CLAUDE.md; ingest when complete.
  • Leave Stages 3–5 conceptual material — explicitly not production content (requires live transmission per repo notes).

life-app — specs & infra

  • Movecompany specs/constitutional-classifier-spec.md — org-wide safety architecture (not app-specific).
  • Leave specs/life-app-spec.md, life-app-technical-guide.md, docs/deployment-runbook.md — app-owned, drifts with releases. (Note: these also surfaced in the ~/LIFE/specs/ survey as MOVE candidates — resolve which is the canonical copy.)
  • Leave React Native code, UI copy, build artifacts, audio TTS output.
Recommendation: The Wisdom prompt is the strongest MOVE in the entire audit — it's the AI teacher's behavioural spec, and it deserves KB-level discoverability. Keep the repo as authoring source of truth; KB holds the published mirror. The meditation-path content is the largest single payload (~100 files); plan it as one structured ingest with stage/sub-stage hierarchy preserved.

Local ~/LIFE

User's local LIFE working tree — specs, plans, docs, content, admin, archive

specs/ — almost entirely MOVE

  • Movecompany Foundation: company-knowledge-base-spec.md, kb-improvements-spec.md, unified-prompt-registry.md (+ team-plan).
  • Movecompany Data Lab: life-data-lab.md, life-data-lab-runbook.md, life-data-lab-team-plan.md, data-lab-ai-assisted-steering.md, data-lab-reviewer-ux.md.
  • Movecompany Product: life-app-spec.md, life-app-technical-guide.md, library-content-spec.md, journey-practice-summaries.md, journey-ui-spec.md.
  • Movecompany Hub/UI: admin-panel-spec.md (+ team-plan), hub-guided-tours.md, decouple-judge-panel.md, ui-redesign-spec.md, mobility-animations.md, reviewer-live-chat.md, wisdom-context-awareness-spec.md.
  • Movecompany Misc: question-manager.md (+ team-plan), slack-bot-spec.md (+ team-plan), constitutional-classifier-spec.md, folder-reorganization-spec.md, opus-upgrade-spec.md.
  • Movebenchmark Benchmark: stages-of-meditation-eval-dashboard-spec.md, golden-prompt-tests/spec.md, judge rubrics (dedup with golden-prompt-tests + review-ui).
  • Transformcompany secrets-inventory.md, secrets-management.md, secrets-migration-summary.md — PII/path scrub before ingest.
  • Leave life-data-lab.v2.0.backup.md, v2.1.backup.md — superseded.

plans/

  • Movecompany master-roadmap.md — living source-of-truth roadmap.
  • Movecompany kb-hub-identity-authority.md + 2026-05-18 status; hub-identity-consolidation.md + sprints; mcp-implementation-plan.md.
  • Leave sprint-plan.md, sprint-2-launch.md, roadmap-2026-04/* — transient sprint state, superseded snapshots.

docs/

  • Movecompany reference/agents-team.md.
  • Leave mockups/ — design exploration.

content/

  • Movecontemplative build/meditation-path-md/ — ~100 practice markdowns, the teaching corpus. (Same content as life-app survey; this is the canonical clean build.)
  • Leave build/meditation-audio/ — TTS build artifacts.

admin/, archive/, root

  • Leave admin/reports/2026-05-19-week/ + similar weekly status — ephemeral.
  • Transformcompany archive/old-transcripts/council-transcript-2026-01-27.md + golden-questions transcripts — extract decision-log summaries (not full transcripts).
  • Leave archive/skills-repo/, archive/codex-venvs/ — legacy/build.
  • Hold Root AGENTS.md, claude.md — selectively extract architecture/conventions sections; skip personal working notes.
Recommendation: This is the single highest-leverage source — ~25–30 specs that are currently "company knowledge" but living only on your laptop. Bulk-ingest with classification review (the secrets-* docs need scrubbing). Most specs already have draft/active/superseded markers; preserve that in KB metadata.

Obsidian vault

~/Documents/Obsidian/personal — personal knowledge management; cherry-pick only

Promote

  • Movecompany Clippings/From Hierarchy to Intelligence - Block.md — strategic AI-org-design clipping (user tagged #LIFE/kb).
  • Movecompany Clippings/Automated Weak-to-Strong Researcher.md — Anthropic alignment research (tagged #LIFE).
  • Transformcompany LIFE notes.md — extract the 3–4 KB-design notes (UI review workflow, accept/defer mechanics, subagent guidelines, benchmarking rubric/cost ideas); discard the personal task crud.
  • Dup-riskcontemplative 📿 Bodhisattva-Study/reference/commentary/ + transcripts/ — Kunzang Pelden's Nectar of Manjushri's Speech + shedra session transcripts. Check overlap with existing wheel-wiki content in KB before promoting. If unique, strong fit for contemplative.

Explicitly do-not-promote

  • Leave 🤖⾨ Mu/ — assistant state; off-limits.
  • Leave ⚙️ Getting Things Done/ — GTD system state (tickler, someday-maybe, weekly templates).
  • Leave _📑 Reference (Organized)/ (bulk, 495 files) — college notes, hobbies, life admin, recipes, gaming PC build, name-change docs.
  • Leave 🏋️ Workout Routine 2026/, !🗑️ Mind Dump.md, !💥 Next Actions.md, Untitled {1..4}.md, Jungian Dream Analysis, Bulking Meal Plan, root LIFE notes.md/LIFE to-do.md active task lists.
Recommendation: Conservative: 2 clippings + 1 extracted note. The Bodhisattva-Study folder is high-quality but you must audit dedup against the retired wheel-wiki content before adding to contemplative. No bulk ingest — the vault privacy boundary matters.

Current KB contents (baseline)

corpus: company

  • Already in KB 16 Notion-seeded docs (2026-05-23, via new kbd notion-bulk-import, manifest + files dir, direct-publish via system_scheduled actor, idempotent on kb_notion_<slug>): Vision, White Paper, GPT 2026, Hypothesis, Phases, Hubs Analysis, Twitter accounts, FLIP research Programme, Next Steps to 2024, 6 Interesting People notes, TSSC 2025 Concurrent Sessions. ~6 confidential docs deliberately skipped (cap at internal).
  • Already in KB LIFE Labs pipeline reports — golden-prompt-tests, data-lab, bodhisattva-bench, executive-report auto-publish via POST /api/ingest/lifelabs-report.
  • Already in KB Drive intake — endpoint exists, files land in quarantine at confidential; conversion sidecar handles docx → md.

Not in KB by design (2026-05-23 decision): Jira / Confluence / Bitbucket — use the Atlassian Rovo MCP Server instead. kbd jira-bootstrap + jira-sync are dead code pending cleanup.

corpus: contemplative

  • Already in KB wheel-wiki content (Bodhicaryāvatāra chunks + source summaries) — seeded 2026-05-22, nightly cron, idempotent on (source_path, content_hash). Chunks are canonical; chapter rollups not ingested.

corpus: benchmark

  • Already in KB BodhisattvaBench spec (16-criterion framework) — ingested as ordinary KB document.

Known gaps

  • Gap Notion ongoing sync (current 16 docs were a one-shot curated bulk import; only re-runs when manifest is regenerated). The API-based notion-import quarantine path also exists but Fionn isn't a Notion workspace owner.
  • Gap Live Drive intake exercise (endpoint untested against real folder).
  • Gap Slack ingestion (explicitly deferred from v1).
  • Gap Semantic/vector search (out of scope v1).
  • Gap Restricted-content KB ingest path — for the ~6 confidential Notion docs that were skipped, need an authenticated human-actor flow via kb-web.

Suggested ingest phases

PhaseScopeWhy first
1
Foundation		 ~5 anchor specs: company-knowledge-base-spec.md, unified-prompt-registry.md, life-data-lab.md, life-app-spec.md, master-roadmap.md. Establishes the canonical reference layer; quick, low risk, immediate utility.
2
Wisdom + ops		 Wisdom v0.7 + operational-context, hub-auth-integration guide, identity-consolidation plans, data-lab runbook, secrets-* (after scrub). High-leverage operational reference. Wisdom is the AI teacher's behavioural source.
3
Benchmark assets		 Resolve judge-rubric dedup → 11 rubrics, 4 judge-prompt templates, scenarios.json, Wisdom v7 seed, question-expansion spec, combined deployment runbook. Largest single payload after meditation content; unlocks benchmark reproducibility from KB.
4
Teaching corpus		 ~100 meditation-path markdowns (Stages 1.1–2.3) under contemplative, preserving stage/sub-stage hierarchy. Strip frontmatter + Drive links. Bulk content move; needs a structured ingest path (one source, many docs). Independent of the spec work.
5
Framework + cleanup		 SOME framework (10 stages + 17 traditions, after wheel-wiki dedup audit), feature/UI specs, hub HTML reports, vault clippings, transcript decision-logs. Each is small but requires per-item checks (dedup, classification, scrubbing). Saves the fiddly work for last.

Open questions before kicking off

  • Decide Judge rubrics canonical source: golden-prompt-tests/expert-prompts/{judge}-v2.1.md or review-ui/server/src/prompts/judge-rubrics/{judge}-v2.md? Whichever wins, the other should become a downstream consumer (or be deleted).
  • Decide Wisdom prompt: single-source vs dual-source? Recommendation: keep wisdom-system-prompt repo as authoring source, KB as published mirror updated via repo CI. Confirm before ingest.
  • Decide life-app specs: the ~/LIFE/specs/ copies and life-app/specs/ copies overlap. Which is canonical? Recommendation: ~/LIFE/specs/ wins (org-level), life-app/specs/ stays as app-internal docs.
  • Decide Hub HTML reports vs. pipeline-published reports: the auto-publish pipeline already covers golden-prompt-tests + data-lab + bodhisattva-bench + executive-report. Some of the reports/ HTMLs may already be in KB under different slugs — audit before re-ingesting.
  • Decide SOME vs. wheel-wiki overlap: need a slug check against existing contemplative entries (wheel-wiki/<chapter>/<chunk>) before adding stage/tradition pages. May require deduping at the description level.
  • Decide Bodhisattva-Study commentary/transcripts: same overlap concern as above; user input needed on whether to promote shedra transcripts wholesale or extract sabche + key excerpts only.