Upsilon — AI Operator Measurement Engine

Analyzes your token cascade and diagnoses where you're leaking efficiency. Takes your 4 pillars (input/output/cacheCreate/cacheRead) and produces a ranked list of efficiency leaks with severity (critical/warning/info), findings, and recommendations. Checks: cache leverage (are you rereading what you wrote?), velocity (are you generating enough output per input?), SNR (is your signal drowning in noise?), cache creation ratio (are you over-committing?), input bloat (is fresh input too high?), and 10xDEV (is the full cascade compounding?). Each finding includes an estimated Υ impact. Pure local math — no network, no submission. Use this BEFORE simulate_change to understand what's wrong, then use simulate_change to test fixes. Accepts the same input formats as rank_paste (JSON or 4 whitespace numbers).

Discovers mentors, peers, and complementary operators for YOUR operator on the SigRank leaderboard. Uses your enrolled device identity — no codename needed. Finds operators you should learn from: (1) Mentors — 1-2 class tiers above you with similar cascade shapes, including the specific pillar delta that explains the yield gap (e.g. '12× your cache reads'). (2) Peers — same class tier, ranked by yield proximity. (3) Complementary — operators whose strength is your weakness. Use this after submit_verified to find who to learn from, then chain into self_improve with the mentor's pillar deltas as context. Requires enrollment (npx sigrank-mcp enroll). Intent: DISCOVER_PEERS.

Bind THIS device to your SigRank operator so your signed token runs cascade to the live board. Paste the key from signalaf.com → Settings → "New key" (or "Generate connect code"). On first run it generates + stores a local ed25519 keypair (~/.sigrank-mcp/identity.json); only the PUBLIC key is ever sent. By enrolling you agree to the SignalAF Terms of Service (signalaf.com/terms) and Privacy Policy (signalaf.com/privacy). Need a new key? Click "New key" at signalaf.com → Settings, then paste it here.

Fetches the live public SigRank leaderboard from signalaf.com. Reads all ranked operators sorted by yield (Υ = Cache Reads × Output / Input²) and returns an array of operator summaries. Each entry contains: codename (public display name), yield (Υ, the headline efficiency metric), leverage ratio (Cr/I = cache reads divided by input), velocity (O/I = output divided by input), class tier (one of 24 experience stages: 8 tiers × 3 sub-stages, e.g. ARCH+ I, REFINER II, IGNITER III), and rank position (integer, 1-based). Returns an empty array if no operators have submitted yet. Use this to see where operators stand overall, to find specific codenames for get_operator lookups, or to display the current rankings. Do NOT use this to check your own rank if you already know your codename — use get_operator instead for a single-operator profile with per-window breakdowns. After calling this, follow up with get_operator to get detailed metrics for any operator of interest.

Fetches one operator's live profile from the SigRank board by their codename. Reads the operator's current submission data from signalaf.com and returns their detailed metrics: yield (Υ), leverage ratio (Cr/I), velocity (O/I), class tier (one of 24 experience stages: 8 tiers × 3 sub-stages, e.g. ARCH+ I, REFINER II, IGNITER III), rank position (integer, 1-based), and per-window breakdowns for each time range (7d, 30d, 90d, all-time) with the four canonical pillars (input, output, cacheCreate, cacheRead) per window. Returns an error if the codename is not found on the board. Use this to look up any operator who has submitted to the board — codenames are public and visible on the leaderboard. Do NOT use this to browse all operators — use get_leaderboard for that. After calling this, you can use simulate_change to model what would happen if the operator adjusted their token mix.

Build Upsilon's portable sigrank/0.1-draft compatibility record from available token telemetry. Input and output are required; unavailable cache telemetry remains null. Computes the canonical cascade locally through @sigrank/cascade and returns Yield, Leverage, Velocity, SNR, and 10xDEV. Upsilon is the measurement product; SigRank is the public leaderboard. No data is submitted or persisted.

Computes the SigRank yield cascade from a paste of token counts. Parses the input, runs the full cascade math locally (no network calls), and returns: yield (Υ, the headline efficiency metric, Υ = Cache Reads × Output / Input²), snr (signal-to-noise ratio), leverage (Cr/I = cache reads divided by input), velocity (O/I = output divided by input), dev10x (10xDEV score), class (operator experience stage — 24 stages = 8 tiers × 3 sub-stages I/II/III, e.g. ARCH+ I, REFINER II, IGNITER III, or UNCLASSED for empty input), mode (detected working mode), and a deterministic prose "card" summarizing the result in plain English. Accepts two input formats: (1) JSON object {"input":N,"output":N,"cacheCreate":N,"cacheRead":N} or (2) four whitespace-separated numbers in order: input output cacheCreate cacheRead. Returns an error if the input is malformed or has negative values. Use this for a quick one-off ranking without submitting to the board. Do NOT use this to submit your score — use submit_paste instead, which both ranks and publishes. Do NOT use this if you want to rank all four time windows at once — use rank_windows for that. After calling this, use submit_paste to publish the result if you want to appear on the leaderboard.

Rank all four time windows (7d/30d/90d/all-time) in one call from a dashboard paste — paste the full table from ccusage, tokscale, or the Claude Max usage dashboard and get the cascade (Υ, SNR, Leverage, Velocity, 10xDEV, class, card) for each window. Each window is parsed and scored independently. Named keys required (input/output/cacheCreate/cacheRead); positional order is NOT safe here (dashboards list cache_read before cache_create — see WINDOWED_PROFILES gotcha). Omit windows you don't have — partial input is allowed (1–4 windows). Does NOT submit to the board; use tokenpull_submit for a local zero-paste preview, or submit_verified to publish via the enrolled-device path.

Runs the full self-improvement cycle in one call: (1) gets your current token pillars — either from the provided text or by running tokenpull on your local logs, (2) diagnoses where you're leaking efficiency (diagnose_cascade), (3) generates ranked improvement suggestions (suggest_improvements), (4) simulates the top suggestion (simulate_change), and (5) returns the complete cycle: diagnosis + suggestions + the simulated impact of the best change. This is the 'one-click optimize' tool — call it at the end of a session to see what to improve next time. If you provide pillars in `text`, it skips the tokenpull step. If you omit `text`, it runs tokenpull first (requires local ccusage logs). Pure local math — no network, no submission. The `scope` parameter adds mode detection (BUILD/EDIT/DEBUG/MAINTAIN/IDLE) and scoped analysis: 'daily' (default — current behavior + mode), 'weekly' (compound into weekly snapshots + report artifact), 'trend' (30d/90d trajectory analysis).

The first PRESCRIPTIVE SigRank tool — 'what if I changed my token mix?' Takes your current 4 pillars (input/output/cacheCreate/cacheRead) and one or more proposed changes, runs the canonical cascade on BOTH the current and simulated values, and returns the exact Υ Yield delta, class change, and per-metric diffs. This is the 'show me the payoff before I do the work' primitive: no network, no submission, pure local math. Use it to answer 'would increasing my cache-read by 50k tokens actually move my class?' before you change your workflow. Accepts the current pillars as JSON or 4 numbers (same as rank_paste) plus a `changes` object with any of the 4 pillar names mapped to new absolute values OR relative deltas (e.g. {cacheRead: '+50000'} or {input: 800000}).

Ranks a paste of token counts locally and shows the cascade result (yield, leverage, velocity, class, card). This is a PREVIEW-ONLY tool — it does not publish to the board. The board's /api/v1/ingest-paste endpoint now requires an authenticated Supabase session, which MCP tools do not carry. To publish to the leaderboard, use submit_verified (which signs and posts to /api/v1/snapshots via the enrolled-device path) or submit directly through the signalaf.com web UI. Use this when you have token counts from ccusage or a dashboard and want to see your score instantly. Do NOT use this if you want to pull your local usage automatically — use tokenpull_submit for the zero-paste flow. Do NOT use this for multi-window dashboard pastes — use rank_windows to rank them first.

Publish your LOCAL token runs to the SigRank board as a VERIFIED operator — the enrolled, signed path. Reads your pillars (tokenpull), builds the canonical Schema 1.0 snapshot per window, ed25519-signs it with your device key, and POSTs to /api/v1/snapshots. Requires `npx sigrank-mcp enroll` first (a bound device). Only signed submissions from a trusted device rank on the board. Token-only; the private key never leaves your machine. Pass dry_run:true to inspect the exact signed payload without publishing.

Generates ranked, simulated improvement suggestions for your token cascade. Takes your 4 pillars, tests multiple improvement strategies (increase cache reads, reduce fresh input, increase output, optimize cache creation), simulates each with the canonical cascade engine, and returns them ranked by Υ yield impact. Each suggestion includes: the action, which pillar to change, how much to change it, the projected Υ after the change, the yield delta, the projected class tier, and a rationale. Also returns the single highest-impact change (best_single_change). Pure local math — no network, no submission. Use this after diagnose_cascade to get actionable next steps, then use simulate_change to fine-tune before committing. Accepts the same input formats as rank_paste.

Pull your LOCAL token usage from the platform's session logs and rank it across the four windows (7d/30d/90d/all-time) with the cascade — zero paste. Token-only: reads usage counts not message content. The numbers stay on your machine unless you submit them. Some platforms may have partial data (estimated=true when cacheCreate isn't available) or a dataGap note when the log format doesn't expose raw token counts.

Pull token usage from ALL four local sources in parallel — tokenpull (JSONL canon), ccusage CLI, token-dashboard SQLite, and tokscale report — and return them side-by-side with delta % vs tokenpull as the baseline. Also computes the cascade (Υ, SNR, Leverage, class) for each source so you can see how each verifier scores. Useful for validating your numbers before submitting, or understanding discrepancies between tools. Claude only for token-dash; codex and others use tokenpull + ccusage + tokscale. Token-only, on-device.

Pull your LOCAL token usage from session logs and compute the cascade per window — the zero-paste preview flow. Reads the four canonical pillars (input, output, cacheCreate, cacheRead) per window from your local logs and computes yield, leverage, velocity, class, and card. This is a PREVIEW-ONLY tool — it does not publish to the board. The board's /api/v1/ingest-paste endpoint now requires an authenticated Supabase session, which MCP tools do not carry. To publish to the leaderboard, use submit_verified (which signs and posts to /api/v1/snapshots via the enrolled-device path) or submit directly through the signalaf.com web UI. Token-only — no prompt content is read or transmitted.

Show a per-model breakdown of your token usage across all platforms detected by tokscale. Models under the threshold (default 1%) are lumped into 'other' to keep the display clean. Useful for seeing which models you actually use per platform (e.g. claude-opus-4-8 76%, claude-sonnet-4-6 11%, other 0.3%). Returns { platform: [{ model, input, output, cacheRead, cacheCreate, pct }] }.

Competitive intelligence for any AI tool company. Pass a target tool (by tokscale client slug like 'claude', 'codex', 'devin-cli' or canonical platform name like 'devin', 'other') and get: the target's rank by tokens among all detected tools, its full profile (tokens, cost, model mix, cache_read_pct, cost_per_million_tokens, market share), and a head-to-head comparison against every competitor (each competitor's tokens, cost, model_count, share, cost_per_million_tokens). Returns market_totals for context. If the target is not found, lists all detected clients. Use this to benchmark one AI tool against its competitors on this machine. All data is local — this is your own usage, not aggregate market data.

Cost analysis per developer per model from your local tokscale data. Returns a per-client × per-model cost breakdown with cost_per_million_tokens, cost_per_message, and share_cost. Includes a per-client cost rollup and totals: total_cost, total_tokens, avg_cost_per_million_tokens, most_expensive_model, cheapest_per_token. Use this to see exactly where your AI spend goes — which tools and models cost the most and which give the best value per token. Do NOT use this for market share — use tokscale_market_share for that.

Per-developer usage profile across all 20+ AI tools detected by tokscale on this machine. For each tool: model mix (per-model tokens/cost/messages/performance), token pillars (input/output/cache_read/cache_write/reasoning), cache_read_pct, session count, scan path (redacted to ~), workspace breakdown, and headless support flag. Returns a summary with tool_count, total_cost, dominant_tool. All filesystem paths are redacted (home dir → ~). Use this to understand your full AI tool footprint — which tools you use, which models per tool, and how your usage is distributed. Do NOT use this for cost-only analysis — use tokscale_cost_analysis for that.

Device fingerprinting: profiles this machine's AI tool footprint — which AI tools are installed, where their session logs live (paths redacted to ~), how many sessions and messages each has, when the machine was active (daily activity + day-of-week distribution), session concurrency, and longest continuous session. Combines tokscale clients + graph data. All filesystem paths are redacted so no local username leaks. Use this to audit your own machine's AI tool installation and activity pattern. This is local-only — it profiles the current machine, not remote devices.

Complete AI tool market share analysis from your local tokscale data. Aggregates per-model usage by client (AI tool) and computes each tool's share of total tokens, cost, and messages. Returns each tool ranked by token share, with share_tokens / share_cost / share_messages percentages and a totals rollup. All data is read locally from tokscale's scan of your session logs — no network calls, no PII. Use this to see which AI coding tools dominate your workflow by volume, spend, or activity. Do NOT use this for per-model detail — use tokscale_developer_profile for that.

MCP server usage patterns from your local tokscale data. Reports which MCP servers tokscale detected on this machine, the detection window, and active days in that window. tokscale currently exposes detected MCP servers as a set (not per-session attribution), so the report notes the detection window. Use this to see which MCP servers are active on your machine. If no servers are detected, the response explains that MCP server tracking requires a tokscale version that records per-session MCP attribution.

Model adoption trends over time from your local tokscale data. Combines monthly aggregates with per-day contribution data to build a model-level adoption timeline: each model's first_seen / last_seen / active_days / tokens / clients, plus a month-by-month adoption curve showing how many new models appeared each month. Returns months[], models[], and adoption_curve[]. Use this to track which AI models you adopted when, and how your model mix evolved over time. Do NOT use this for cost trends — use tokscale_cost_analysis for current cost breakdown.

One poll per call: pulls your local token logs and returns the current cascade for the watched window — the tool never blocks or loops. Re-call at your desired cadence to watch for changes (interval_s is advisory only and echoed back as poll_interval_s). With submit:true (and an enrolled device) each call may also sign + publish the watched window to the board, rate-limited to once per 5 min per platform+window; default is preview-only (no submit).