Skip to content
Back to skills

Organ Formula Foundry Bakeoff

ASecurity

Foundry Core B bakeoff harness and promotion/demotion ledger that runs synthesized formula candidates up a staged promotion ladder and records every gate verdict to an append-only ledger. Use when formula candidates need scoring, promotion, or demotion through the staged ladder, or when invoking the bakeoff to test and govern formula quality.

  • 2 stars
  • 0 votes
  • 0 copies
  • 3 views
  • Added September 10, 2026
toolsgobashnodesecurity

Works with

  • cli
  • mcp

Security analysis

A100/100

Scanned September 10, 2026

npx -y skills add nexuslinkproductions/yuri-os --skill organ-formula-foundry-bakeoff --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Organ Formula Foundry Bakeoff?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Organ Formula Foundry Bakeoff
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/nexuslinkproductions-organ-formula-foundry-bakeoff-yuri-os/badge)](https://www.skillsdirectory.com/skills/nexuslinkproductions-organ-formula-foundry-bakeoff-yuri-os)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: organ-formula-foundry-bakeoff
description: "Foundry Core B bakeoff harness and promotion/demotion ledger that runs synthesized formula candidates up a staged promotion ladder and records every gate verdict to an append-only ledger. Use when formula candidates need scoring, promotion, or demotion through the staged ladder, or when invoking the bakeoff to test and govern formula quality."
triggers:
  - "organ-formula-foundry-bakeoff"
  - "how do I use formula-foundry-bakeoff"
  - "formula-foundry-bakeoff usage"
  - "formula-foundry-bakeoff guide"
  - "Formula Foundry Bakeoff + Promotion Ledger (Core B test layer)"
generated: true
source_node: "formula-foundry-bakeoff"
source_file: "_SYSTEM/Scripts/math/formula-foundry-bakeoff.mjs"
scope: harness
invocation: ability
---

<!-- GENERATED from the canonical graph node "formula-foundry-bakeoff" (mechanism.guide) by _SYSTEM/Scripts/yuri-guide-project.mjs.
     DO NOT hand-edit — edit _SYSTEM/organ-guides.json, then run: node _SYSTEM/Scripts/yuri-guide-seed.mjs && node _SYSTEM/Scripts/yuri-guide-project.mjs -->

# Organ Guide — Formula Foundry Bakeoff + Promotion Ledger (Core B test layer)

**Module:** `_SYSTEM/Scripts/math/formula-foundry-bakeoff.mjs` · **Layer:** Energy & Math · **Invocation:** both · **CLI:** `ledger`

**Purpose.** Foundry CHUNK 3 — the bakeoff harness + the promotion/demotion ledger (the SCORER + GOVERNANCE side of Core B). Takes already-synthesized candidates as INPUT and runs them up a staged promotion ladder, recording every gate verdict to an append-only ledger. Deliberately the GENERATOR≠SCORER boundary: it never imports the synthesizer, so a synthesizer can never grade its own candidate.

## Exports
- `stableHash(obj)`
  - in: any JSON-serializable object
  - out: a deterministic stable hash string (sorted-key, no clock)
- `PROMOTION_LADDER (frozen const)`
  - in: —
  - out: the ordered, frozen list of promotion rungs
- `readLedger(ledgerPath = LEDGER_PATH)`
  - in: optional ledger path
  - out: array of gate records (malformed lines skipped)
- `appendGateRecord(record, ledgerPath = LEDGER_PATH)`
  - in: a gate record + optional path
  - out: appends one JSONL line to the ledger (mkdir-safe)
- `canPromote(formulaId, toRung, ledger = readLedger())`
  - in: formula id + target rung + optional ledger
  - out: boolean — whether the ladder + evidence permit promotion to that rung
- `promote(card, toRung, ledger = readLedger())`
  - in: card + target rung + optional ledger
  - out: the promotion result; records a gate entry
- `demote(formulaId, toRung, reason, opts = {})`
  - in: formula id + target rung + reason
  - out: the demotion result; records a gate entry with the reason
- `stage0Intake(candidate)`
  - in: a candidate
  - out: stage-0 intake verdict
- `stage1Fixtures(card)`
  - in: a card
  - out: stage-1 fixture-test verdict
- `async stage2Counterexamples(card, opts = {})`
  - in: a card + opts
  - out: stage-2 counterexample-search verdict
- `stage3RealData(candidates, opts = {})`
  - in: candidates + opts
  - out: stage-3 real-data verdict with ACTUAL corpus totals (no top-N laundering) or a refusal
- `async runBakeoff(candidates, opts = {})`
  - in: candidates + opts (incl. opts.stamp for determinism)
  - out: the full staged bakeoff result

## Security boundary
MUTATING but bounded: appends to the promotion/demotion LEDGER (_SYSTEM/state, append-only JSONL, mkdir-safe) — never deletes or rewrites. No protected-path access. DOMAIN-BLIND: nothing inspects a candidate's source domains to loosen/tighten a gate — music, frequency, magnetism, numerology, alchemy, information-theory clear the SAME rungs on the SAME evidence bar; a candidate fails only for insufficient evidence. Deterministic: no Math.random, no Date.now (timestamps via opts.stamp).

## When to use
Promoting/demoting a formula candidate through the evidence ladder, reading the gate-decision history, or running the full staged bakeoff over a candidate set. The governance/scoring counterpart to formula-foundry's generation.

## Gotchas
- GENERATOR≠SCORER: this module never imports synthesizeFormulaCandidates — keep that separation; do not wire the synthesizer in.
- Determinism requires the caller to pass opts.stamp — there is no clock inside; a missing stamp loses reproducibility.
- stage3RealData reports REAL corpus totals or REFUSES — it will not emit a top-N severity-laundered count.
- The ledger is append-only; a later record for an id supersedes earlier ones (read reconstructs current state).

## Session Notes

### 2026-06-13
- session: 40m | peak ctx: 0% | compacts: 0
- tools: Bash×824, Read×163, Edit×17, StructuredOutput×16, Write×8, TodoWrite×4, ToolSearch×2, ScheduleWakeup×2, Workflow×1, mcp×1, AskUserQuestion×1
- corrections: none
- errors: none

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.