Snapshot of the three tracked domains. Full detail on the Nodes tab; what changed and when on Version History.
Cowork exposes no per-session or per-model token telemetry, so this is a transparent turn-weighted estimator calibrated against one real Settings > Usage reading — provisional, not a settled constant, recalibrated whenever a fresh reading lands. Full methodology, weighting rationale, and worked examples: CL07 claude-session-load.md; raw weekly data points: outputs/usage-capacity-ledger.md.
Weekly Usage % ≈ k × Σ(turns × weight) — weight 3 for browser-automation/screenshot-heavy turns, 2 for large-context/audit/research turns, 1 for routine text-only turns. Example: a 200-turn browser session ≈ 200×3×0.0118 ≈ 7.1% of the week's quota by itself. This is a fallback for sessions you can't inspect directly — always prefer a direct read_transcript count when a session is inspectable.Pulled via mcp__session_info__list_sessions during this dashboard update — a real, current count of inspectable sessions, not a modeled figure. This number is refreshed whenever Claude is actively working in this KB (system-document regen, dashboard updates, etc.); it is not part of the automated daily launchd refresh, since that runs as a standalone script with no access to session-introspection tools. Treat the "as of" date as the actual freshness of this number.
| Era | Sessions | Heavy ratio | Dominant register |
|---|---|---|---|
| ~05-26 to 06-12 (genesis) | 55 | ~36% | Opus Clip/HSTV production + KB scaffolding |
| ~06-13 to 06-22 | 61 | ~49% | DOAV/Upper Hand production + KB library build |
| ~06-23 to 06-30 | 77 | ~26% | Gravitas character-portrait generation + audits |
| ~07-01 to 07-09 | 36 | ~17% | n8n v7 debugging + Gravitas UI/voice work |
| 07-10 (real anchor) | 58 | 3.4% by count / ~48% of usage | 2 outlier sessions dominate — count ≠ usage share |
claude-session-load.md §Case Study + §Forward capacity planning · outputs/usage-capacity-ledger.md · outputs/usage-capacity-audit-2026-06-12-to-2026-07-10.md. Cross-referenced on the Economics tab's "Where the credits go" section (this tab is the dedicated, current-state view; Economics keeps the cost-modeling angle).Every tracked node, grouped by domain. version_added marks which dashboard version first surfaced it.
One entry per check that produced a real change. INIT first build · PATCH a node updated · MINOR a node added or structural change · NOOP checks (nothing changed) are logged to the refresh log only, not here, to keep this list signal-only.
A complete file-based agentic operating system, from a Karpathy-pattern seed to an externally-benchmarked, multi-machine fleet. Click a phase to see its milestones.
Plain markdown on local disk, governed by a two-level router. Everything reads from a three-file session start, routes down to content, and syncs off-machine.
claude-system-architecture.md (SYS02, Jun 13) still describes the deprecated checkpoint system and a 9-task fleet. This dashboard reflects the current state; SYS02 should be refreshed.Copying files reproduces the library. Adopting these fifteen findings reproduces the system. Ordered by leverage; gold = highest.
Two paths. Clone Dean's exact KB for a same-operator machine, or fresh-build the architecture for a separate persona with its own data.
2026-07-17 refresh (living process — see below), 8 rated entries across 14 sources this pass, superseding the 2026-06-29 static snapshot. Ahead on governance/concurrency-safety, at parity on core memory-consolidation practice, one honest open gap (multi-agent orchestration), one trade re-validated against new field data.
What the system saved, what it costs to run, and what a comparably-skilled engineer would spend to build it from scratch. No raw billing telemetry exists — credit figures are a transparent model (frequency × model tier × token cost), labeled measured / modeled / estimated, now anchored to periodic real Settings > Usage readings starting 2026-07-10.
~440 man-hours (280–660) = 2.2 FTE = ~$23.5K wages avoided in 5 weeks. Top roles: Software Dev + Knowledge Manager — not admin. (OP13)
The recurring operation displaces ~$150K–$250K/yr in labor once one-time build spikes are stripped.
Same output ordered from agencies: $70K–$236K (mid ~$150K). The Gravitas platform drives the high end.
Comparable-skill solo engineer to reproduce the apparatus: ~460 hrs (300–630) × $100–175/h = ~$60K ($30K–$110K), 3–6 months. Actual: ~150 operator-hrs + ~$300 tooling → ~3× compression, >95% cash saving.
Settings>Usage read 26% weekly / Fable 0%. Turn-weighted attribution: ONE browser-automation session (character factory, 350 turns) ≈ 48% of that week's usage alone — 9× past its 40-call checkpoint threshold, none fired. Formula k≈0.0118%/weighted-turn calibrated from this reading — provisional, not a cardinal rule. Full detail: CL07 claude-session-load.md; trend: outputs/usage-capacity-ledger.md.
The census above first shipped historical turn counts marked "not measured, too costly" — untested, and wrong; 3 real read_transcript pulls came back cheap. Found: at least one session_id spans multiple day-resets (session_id ≠ work-unit). Fix rolled forward into planning: weekly-review-friday Step 4C computes a weekly capacity budget/headroom; daily-briefing's Monday run surfaces it (no re-computation). Full diff for Jim: outputs/jim-update-audit-2026-07-10.md.
~25 long Sonnet runs/week (curiosity, kaisen, youtube, claude-opt). #1 recurring sink.
browser_batch embeds base64 images — ~10× a text-tool flow. #1 per-session burn.
~7× token multiplier. Powerful fan-out, priciest per unit of work.
~22K tokens across 4 core files × ~20 daily runs = a fixed daily floor. (Halved by CLAUDE 512→265.)
Highest per token, low frequency (Gravitas strategy, decisions, contradiction resolution).
Haiku mechanical tasks, disabled one-time reminders (zero cost), routed pointer-memory, short briefings.
Text-tools over screenshots (~10×) · JIT routing (~4.7K tok/session) · CLAUDE 512→265 (~½ read cost) · spent-output pruning · Advisor (11%) · suppress-explanations (30–50%) · no vector DB · task condensation · lane system (kills re-read waste).
Done 2026-07-01: nightly research consolidated (4→1) · frequency-tuned (YouTube + Claude-opt → M/W/F) · JIT loading rule. Remaining: Haiku-first (interactive / API only — no model param on tasks) · n8n/API migration (tabled) · skill-chains (already CL03 standard).
The live five-panel map: Panel A the operating architecture as a labeled concept map · Panel B the FULL KB content graph — every tracked .md file in the KB, not just registered wiki articles — grouped by topic, Canvas-rendered for scale · Panel C the citation network — typed edges (Internal Findings / Wiki Cross-link / Raw Source / Task Output) traced from the KB's own [EXT]/[INT]/[WIKI]/[RAW]/[TASK] citation markers, colored by type · Panel D knowledge clusters — the clean, registered-wiki-only view (242 articles) grouped by domain · Panel E File References — directed, arrowed edges showing exactly which file references which, typed/colored by mechanism (article-code mention / markdown link / the KB's own "→ file.md" arrow convention), styled after Jim's autelligence-command.pages.dev KB Graph tab; click any node for an ID/domain/degree/hub/isolated info card. Panel E also formalizes the relationship between the content graph and the citation network: of 103 formal citations, 81 (78.6%) are backed by a discoverable structural link, 22 are narrative-only citations with no matching link, and the graph carries 4,851 structural links that were never formally cited — most links in the KB simply aren't citation-marked prose, which only appears inside wiki articles. Panels B and D also carry a ⇄ Direction toggle (top-left, below the search box): off by default (undirected, single-color, easiest to scan by topic); switch it on for directed arrows colored by each edge's source node's cluster, so you can see at a glance where a region's outbound connections are headed. Panels C/D/E added 2026-07-17 per Dean's directives; the direction toggle added same day. If the frame is blank, open system-concept-map.html directly (it sits beside this file in outputs/).
Panel B — content graph (FULL KB): 1,528 files · 4,932 links · 155 hubs · 509 isolated (no detected cross-refs — mostly changelog-lane archives and raw source dumps) · 33 topic clusters. Panel C — citation network: 77 nodes · 118 typed edges (INT 56 / WIKI 37 / RAW 22 / TASK 3) · 286 external citations counted (not graphed). Panel D — knowledge clusters: 242 registered wiki articles · 1,094 links · 64 hubs · 5 isolated. Panel E — file references: same 1,528 files/4,932 edges as Panel B, directed + typed by mechanism (code 3,911 / mdlink 556 / arrow 465). Panel A — architecture map: 31 nodes / 7 subsystems. Source builders: outputs/build_full_kb_graph.py, outputs/build_citation_graph.py, outputs/rebuild_kb_adjacency.py, outputs/build_edge_reconciliation.py + outputs/build_system_concept_map.py.