tonydzi/second-brain-starter-kit/alpha-judge

🅰️⚖️ /alpha-judge — the reusable LLM-judge for every alpha miner

🧒 When reporting to a non-technical operator: end with a child-simple "In plain words" recap in their language (their standing request; reports TO the operator only).

The alpha-extraction engine = cheap deterministic detector (0 tokens) → LLM judge → weekly ≤10 + yes/no calibration, repeated for each of the 10 miners. The DETECTORS are per-miner scripts in $IMPORTS_ROOT/alpha/. The JUDGE is the SAME logic every time — so it lives HERE, once, and is reused. Canon: 02-Decisions\decision-alpha-extraction-engine-variant-a; memory [[alpha-extraction-engine]].

Model policy (per [[model-routing-sonnet-grunt]] + the operator's 2026-06-14 session override)

• In-session, now: judge on Opus (the operator steers; he'll say when to switch to Sonnet as Opus nears its limit).
• Nightly unattended: Sonnet (idle free bucket) via judge_nightly.cmd (headless claude -p --model sonnet). LIVE + verified 2026-06-15 — Windows task "Alpha Judge Nightly" Mon 05:00, chained after the 04:15 scan. Now judges ALL 10 miners (was bets-only; expanded 2026-06-15) — one headless session reads THIS spec, loops the registry, applies the dedup-vs-home gate, writes each <miner>-judged-latest.md. Vault writes are forbidden in nightly mode (judge-only; home-note additions wait for the operator's in-session approval). (Auth is in the Windows keychain; check with claude auth status, re-login via claude auth login. Keep .cmd pure-ASCII.)
• The judge is grunt-classification → Sonnet is the natural default once the session-override lifts.

Miner registry (detector → home → verdict scheme)

Flow (what YOU do)

1. Pick the miner from the arg (default bets if none; if unclear, ask). Read its candidates file (or query its db for the top unjudged candidates — cap ~25; never the whole corpus).
2. RECALL the HOME note (e.g. insight-prediction-ledger / insight-contradictions) so you DON'T re-propose what's already catalogued. Token-cheap: read only that one note + the candidate digest.
3. JUDGE each candidate against the miner's core question — "is this ANTON'S OWN <X-of-the-right-type>, not noise / not someone else's / not a restatement?" Classify per the verdict scheme + a ONE-line reason. The detector optimises RECALL; the judge supplies PRECISION.

• 🛡 DEDUP-VS-HOME GATE (mandatory — added 2026-06-15): before you KEEP a candidate, check it against the HOME note you recalled in step 2. If the candidate merely restates something already catalogued there (same belief / value / bet / question / shift / bridge — even if worded differently), DROP it with the verdict 🗑 already in the home note (restatement). Only candidates that add something the home note does NOT already contain survive. This is a semantic check (you, the judge, do it — NOT a blind word-overlap filter, which would false-drop real nuance). Why: on 2026-06-15 the identity (11 raw keeps → 0 truly new after dedup) and stance miners re-proposed items already in _Self-Bible-MOC / insight-contradictions; this gate makes that dedup a required step, not a manual afterthought.

1. Write the clean signal → _imports\alpha\candidates\<miner>-judged-latest.md (utf-8, \n): only the ✅/🟡 keepers, each with quote(s), source [[wikilink]], date if known, verdict, reason. If zero signal, say so plainly (that's a valid, honest result — don't invent).
2. Propose HOME additions (the keepers not already in the home note): show the operator BEFORE→AFTER ([[show-before-after]]); on their OK → vault_backup.py first ([[vault-backup-rule]]) → append to the home note (draft status, they edit) → recalc if the home has an engine (bets → ledger_calibration.py). Never auto-write a curated vault note.

Guardrails

• Token-economy ([[vault-data-architecture]]): judge reads ONLY the candidate digest + the one home note. NEVER load the corpus. The detector already did the 0-token narrowing.
• Honesty over yield: "0 real" is a legitimate verdict (contradictions hit it 2026-06-14 — the operator is consistent + already has the curated map). Don't manufacture signal to look productive.
• No new ledgers: every miner FEEDS a pre-existing curated home (bets→prediction-ledger, contradiction→insight-contradictions). Surface NEW candidates; don't duplicate the home.
• Vault writes = backup → preview → approve → reindex. Concurrent-git aware: if the vault index.lock is held by live git, defer the write and say so.

Not this skill

Building/refining a DETECTOR script = direct work in _imports\alpha\ (not here). Running the deterministic scan = the miner's own .py / weekly task. This skill is ONLY the judge + home-feed stage.

<!--kit-footer-->

Like this skill? It is part of second-brain-starter-kit: the second brain we built for ourselves and run every day at Palo Alto AI Research Lab. Install the whole set with npx skills add tonydzi/second-brain-starter-kit. Everything is open source and free, so take what you need.

Flagships worth a look on their own: secondop-panel (a second opinion from a panel of external models), claude-memory-tidy (stop your agent's memory from rotting), telegram-mcp-kit (your own Telegram over MCP in about 15 minutes).

Author: Anton Dziatkovskii, Palo Alto AI Research Lab. Telegram @tonydzi - WhatsApp +1 341 222 9178 - X @Tony_Stef_

Engineers: want to test-drive this setup? Message me. I hand out free starter seeds to engineers who test and report back, and custom skill requests are welcome.