Product.ai / Join / Projects / Verification Ladder + Back-Pressure Hooks — close the deterministic gate on one Cortex or AIOS agent workflow
Project Open to Alpha Team

Verification Ladder + Back-Pressure Hooks — close the deterministic gate on one Cortex or AIOS agent workflow

Audit one agentic workflow inside Cortex or AIOS — a cron-triggered skill, an autonomous pipeline, a sub-agent fan-out — and ship the deterministic verification ladder on it. Tier 1 (mandatory human review where applicable), Tier 2 (back-pressure hooks: typecheck, lint, tests, schema validation on every agent stop), Tier 3 (lightweight 20-50 task eval substrate built from real failures), and a named external truth anchor between steps. The deliverable is a workflow that physically cannot reach production state in a broken configuration.
Project Overview
Discipline
AI Systems — AI Engineer · ai-systems-engineer
Duration
2 weeks
Compensation
Your stated freelance rate
Surface
Engineering · Product.ai · Cortex
Kernels
engineering · productai · cortex
Outcomes
team-visible · dev-integrate · chat-expert
Tier
Applied
Alpha Team
Open to alpha members who want to take this on
Tooling
Claude Code or Co-work

Why we want this done

The Stripe Theorem is FORGED: agent reliability is bounded above by the reliability of its surrounding deterministic harness, not by model capability. Compound failure math does the rest — 95% per-step accuracy across 20 steps yields 35.8% end-to-end success. Cortex and AIOS run dozens of agentic cron paths today. Many have probabilistic discipline (rules, advisory hooks) but lack deterministic gates. The Multi-Agent Physics axioms (CLAUDE.md §5.2 rules 7-10) are explicit on this: probabilistic discipline has a structural 41-87% failure ceiling at multi-agent scale. Whichever workflow ships verification gates first becomes the template for the rest. The candidate also gets to demonstrate the most load-bearing AI engineering instinct in the rubric — verification-first under pressure — on real Product.ai code, not a contrived exercise.

Scope

  1. Pick one workflow (the candidate proposes; we pressure-test the choice). Strong candidates: an existing cron-triggered skill in ops/cron/ or a sub-agent dispatch inside an AIOS skill
  2. Audit the current verification ladder — what tiers exist, what's missing, what's advisory vs. deterministic
  3. Implement Tier 2 back-pressure hooks — typecheck, lint, tests, schema validation, hard iteration cap
  4. Add lightweight Tier 3 eval substrate (20-50 task cases from real past failures, not synthetic)
  5. Identify and instrument a named external truth anchor — a verification step the agents cannot influence
  6. Document the 9-class race-class taxonomy mapping for any near-miss the workflow has produced (per CLAUDE.md §5.2 §archaeology ritual)
  7. Ship a one-page enforcement-companion table — for every advisory rule in the workflow, the matching deterministic gate

What success looks like

  • The workflow is observably more reliable in the trial window — measurable, not theoretical
  • Hook failures produce non-zero exit codes and clean error messages
  • An engineer reading the new gates can reproduce them on a different workflow without re-asking
  • The race-class taxonomy entry lands in aios/DEPENDENCY-PAIRS.md or equivalent
  • The external truth anchor is nameable in one sentence ("Y is the verification step the agents cannot influence")
  • A failing PR in the workflow's CI is blocked, not warned

References

references.md
Cortex CLAUDE.md §5.2 rules 7-10 (Multi-Agent Physics, deterministic gates over probabilistic discipline)
axioms/Engineering/Multi-Agent Physics (2026-04-18).md
AI Engineering Phase 3 briefing axioms A4 (Stripe Theorem), A5 (External Truth Anchors), D4 (Verification Substrate)
ops/cron/ for existing cron architecture; aios/.claude/skills/ for skill examples
aios/dashboard/scripts/validate-html-js-contract.py and existing CI gates as reference
Anthropic Claude Code April 23 2026 postmortem (the closed-loop-verification failure case)

Constraints

  • Claude Code as primary substrate
  • Hooks must be deterministic (PreToolUse-style guards, exit-code enforcement) — advisory rules with reminders do not satisfy the brief
  • IP separation: no aios-methods access required for application-layer cron paths; if the chosen workflow touches methodology, swap to one that doesn't
  • Race-class taxonomy entry follows the existing nine-class scheme (AHCR / HMR / CDR / BSR / MCAR / AIR / FBOR / WAW / PDE)
  • The 24-hour archaeology ritual applies to any incident the workflow produces during the trial
Apply
01

Read the Codex (10 min)

The operating principles we work by. If they resonate, the rest of this will land. Open the Codex →

02

12-minute video screen

Hireflix, async. Questions are calibrated to this project specifically.

03

Chemistry call (30-60 min)

Direct call with the CEO. Strategic alignment and mutual fit. No problem-solving exercise.

04

Project begins within 2-3 weeks

1099 contractor agreement, NDA, paid at your stated rate. Day 1 in Santa Monica.

Alpha Team members can take this project without the screen-and-call sequence. Reach out via the Alpha Team channel.