v1.9.0 is here — /autoresearch:reason just dropped and it changes EVERYTHING about subjective work

[uditgoenka]

autoresearch can now THINK for you. Not just code. THINK.

I just shipped v1.9.0 with the 10th subcommand: /autoresearch:reason — and honestly? This might be the most important addition since the original loop.

---

The backstory

autoresearch has always had one limitation that bugged me: it only works when you have a number to measure. Test coverage? Easy. Bundle size? Done. API latency? Chef's kiss.

But what about:

• "Should we use event sourcing or CQRS?"
• "Is this pitch deck actually compelling?"
• "Which authentication architecture is right for us?"
• "Is this blog post good enough to publish?"

No metric. No val_bpb. No mechanical verification. Just... vibes? Gut feel? Ask Claude to "make it better" and pray?

Not anymore.

---

Introducing /autoresearch:reason

It 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.

Here's what happens when you run it:

Generate version A
  ↓
A fresh critic ATTACKS it (minimum 3 weaknesses)
  ↓
A separate author who only sees A + the critique produces version B
  ↓  
A synthesizer who has NO history with either produces AB
  ↓
A blind judge panel with RANDOMIZED LABELS picks the winner
  ↓
Winner becomes the new A → repeat until judges can't improve it

Every. Single. Agent. Is. A. Cold. Start.

No shared session. No sycophancy. No "this is great but here's one tiny thing." The critic is FORCED to find 3+ weaknesses. The judges see X/Y/Z labels, not A/B/AB — they literally don't know which version is the "original."

This is peer review for AI outputs. And it WORKS.

---

Show me the money

# Debate an architecture decision (converges in ~5-8 rounds)
/autoresearch:reason
Task: Should we use event sourcing for our order management system?
Domain: software
Iterations: 8

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

# The killer combo: converge on a design, then validate it with 5 experts
/autoresearch:reason --chain predict
Task: Design the caching strategy for our high-traffic API
Domain: software

---

The --chain flag is where it gets WILD

You know how predict has --chain debug? Reason has it too. For EVERYTHING:

Chain	What happens
reason → predict	Converge on best design → 5 expert personas stress-test it
reason → plan,fix	Debate approach → plan it → implement it
reason → scenario	Converge on design → explore 25+ edge cases
reason → debug	Converge on proposal → validate empirically
reason → ship	Refine content → publish it
reason → learn	Iteration lineage → automatic ADR documentation

That last one is sneaky powerful: the iteration lineage automatically becomes an Architecture Decision Record. You debated event sourcing vs CQRS through 8 rounds? The lineage.md file IS your ADR. For free. No extra work.

---

3 modes for different situations

Mode	When to use
convergent (default)	You want THE answer. Stops when judges consistently pick the same winner.
creative	You want ALTERNATIVES. Generates diverse candidates without forcing consensus.
debate	Pure A vs B. No synthesis. For when you want a clear head-to-head winner.

---

What I shipped

• 618-line protocol defining the full 8-phase workflow
• 341-line command guide with examples, chain patterns, tips, anti-patterns
• 307-line scenario walkthrough showing a real 8-round event sourcing debate
• Updated README, COMPARISON, guide index, chains guide, scenario index
• Full --chain support for all 8 downstream autoresearch commands
• Oscillation detection — if the winner keeps flip-flopping, the system detects it and stops
• Convergence detection — N consecutive majority wins (default: 3) triggers auto-stop

15 files changed, +2,215 lines. Zero breaking changes.

---

Why this matters

Karpathy's autoresearch question #7 was:

"Could autoresearch work for non-differentiable systems? What domains satisfy the 'scalar metric + fast eval' requirement?"

With /autoresearch:reason, the answer is: all of them. The blind judge panel IS the scalar metric for subjective work. It's val_bpb for decisions, content, strategy, and design.

autoresearch now covers:

• Objective domains → /autoresearch (mechanical metric)
• Subjective domains → /autoresearch:reason (blind judge panel)
• Everything in between → chain them together

The loop is universal. The metric is optional. The improvement is inevitable.

---

Get it now

/plugin update autoresearch

Or if you're new:

/plugin marketplace add uditgoenka/autoresearch
/plugin install autoresearch@autoresearch

Then try:

/autoresearch:reason
Task: [your hardest subjective decision]
Domain: software
Iterations: 5

And watch it converge on something better than any single prompt could produce.

---

Full release notes | Command guide | Scenario walkthrough

Let me know what you reason about first. I'm genuinely curious.