github uditgoenka/autoresearch v1.9.0
v1.9.0 — /autoresearch:reason — Adversarial Refinement for Subjective Domains

latest releases: v2.2.1, v2.2.0, v2.1.4...
3 months ago

What's New in v1.9.0

/autoresearch:reason — The 10th Subcommand

Extends autoresearch to subjective domains where no objective metric exists. Constructs a subjective fitness function through isolated multi-agent adversarial refinement with blind evaluation — the same way science uses peer review where math uses proofs.

The blind judge panel IS the val_bpb equivalent for subjective work.

How It Works

Generate-A → Critic attacks A (strawman) → Author-B sees task+A+critique →
Synthesizer sees task+A+B → produces AB → Blind judge panel (randomized labels)
picks A/B/AB → Winner becomes new A → Repeat until convergence

Key invariant: Every agent is a cold-start fresh invocation — no shared session, no history bleed. This prevents the sycophancy that makes single-model refinement fail on subjective tasks.

Quick Start

# Debate a software architecture decision
/autoresearch:reason
Task: Should we use event sourcing for our order management system?
Domain: software
Iterations: 8

# Refine a pitch with 5 blind judges
/autoresearch:reason --judges 5 --iterations 10
Task: Write a compelling Series A pitch
Domain: business

# Chain: converge → validate → implement
/autoresearch:reason --chain predict,fix
Task: Design caching strategy for high-traffic API
Domain: software

Features

  • 8-phase protocol: Setup → Generate-A → Critic → Generate-B → Synthesize-AB → Judge Panel → Convergence Check → Handoff
  • Context isolation: Each agent spawned fresh — prevents sycophancy and anchoring
  • Blind evaluation: Judges receive crypto-random labels (X/Y/Z), forced comparative evaluation
  • Convergence detection: N consecutive majority wins (default: 3 rounds)
  • Oscillation detection: Forced stop if incumbent changes 5+ times without convergence
  • 3 modes: convergent (default), creative (no auto-stop), debate (no synthesis)
  • 6 domains: software, product, business, security, research, content
  • --chain flag: Pipe converged output to debug, plan, fix, security, scenario, predict, ship, or learn

Flags

Flag Purpose Default
--iterations N Bounded mode unlimited
--judges N Judge panel size (3-7, odd) 3
--convergence N Consecutive wins to stop 3
--mode convergent, creative, debate convergent
--domain software, product, business, security, research, content (asked)
--chain <targets> Chain to downstream tools none
--judge-personas Override default judges domain-default
--no-synthesis A vs B only false

Chain Patterns

reason → predict        (converge then stress-test with 5 experts)
reason → plan,fix       (converge then implement)
reason → scenario       (converge then explore edge cases)
reason → debug          (converge then validate empirically)
predict → reason        (identify issues then debate solutions)
scenario → reason       (discover edge cases then debate how to handle)

Documentation

  • Command Guide — comprehensive 341-line guide
  • Scenario Walkthrough — 8-round event sourcing debate example
  • Chains & Combinations — reason chain patterns
  • Protocol Reference — full 8-phase spec

Files Changed

15 files changed, +2,215 lines:

  • 6 new files (reason-workflow.md, reason.md command, guide, scenario walkthrough)
  • 9 updated files (SKILL.md, README, COMPARISON, guide index, chains, scenario index, manifests)

Full Changelog

v1.8.2...v1.9.0

Don't miss a new autoresearch release

NewReleases is sending notifications on new releases.