DRAFT prior · builder · concrete-first · multiplayer fanboy

01Bootstrap

02Plan

03Draft

04Review

05Publish

06Reader

architect: composing

layout · wide-editor + evidence + voice-mirror · density meter muted (you prefer exploration mid-draft)

Meta-harnesses

A series, target: AI engineers and builders.

3. Building meta-harnesses

If the harness is the controller, the meta-harness is the controller that authors controllers. It evolves the harness as the task and the user evolve.

The search space isn't free; every harness atom is typed and every mutation is a delta against a manifest the Architect has to emit. ▸ akhil · 1

Staying within the user's prior is the harder half. The harness must remember

stuck on framing? the architect noticed you've been still 4s on a paragraph that's mid-claim. try restating §3 as a one-line wedge first? esc · dismiss

Evidence pinned · vishal

raw.works · vishal

RLMs as new reasoning models

3 quotes pinned

↩ akhil left review note here

x.com · paraschopra

Recursive feedback: when models grade themselves

akhil · "lean on this for §3.2 close"

x.com · trq212

Multiplayer harness thread

2 quotes pinned

youtube.com

Meta-harnesses talk

transcript indexed

voice mirror · akhil's last 2 paragraphs · so you can match section voice

akhil · 4 min ago "voice fidelity isn't really a feature here, it's the constraint the whole thing has to bend around — every other knob is downstream of whether the draft sounds like you would've said it."

akhil · 12 min ago "I feel like it's the intermediary to world models for businesses. Eval the loop, not the token."

akhil left a comment on §3.2 · jump

2 authors live phase: draft yjs · 412 ops architect · composing

sync · same chain vishal · typed +24 chars · §3 akhil · typed +18 chars · §1 akhil · §1 chunk ready · review? vishal · comment on §1 opener · routes to akhil akhil · review · comment on vishal's §3.2 architect · stall-detector fired (vishal) architect · re-authored vishal's right-rail→ yjs · presence · trace

DRAFT prior · claim-first · evals advocate · dense paragraphs

01Bootstrap

02Plan

03Draft

04Review

05Publish

06Reader

architect: idle

drafting · §1 RLMs framing · claim-first paragraphs §1 claim-density 2.8/ 3

density meter · §1 (drafting)

this paragraph target ≥ 3

currently 2.8 claims/para · trending up

tension tracker · §1

Open · §1 framing

"RLMs as a new class, or a sub-genus of reasoning models?" — Akhil drafting the claim-first version now.

Akhil · drafting

Open · evals

Golden eval as a single rubric, or recursive at every layer? Akhil to address in §1 opener.

Akhil · in draft

drafting §1

Meta-harnesses

A series, target: AI engineers and builders.

1. RLMs, framed claim-first

Reasoning Language Models aren't a new species; they're what happens when the harness around an LLM gets reflective enough to grade its own steps.

The eval is the loop, not the token. Once you accept that, every architectural choice downstream — memory, tool-use, multi-pass refine — collapses into one question: what are you scoring at each step?

A claim-first opener earns the reader's patience. So: the meta-harness is an architect that

review queue · from vishal

akhil · reviewing

vishal's draft · §3.2 · "Staying within the user's prior…"

isn't this just the eval problem dressed up? if you can't measure prior-fit you can't tune for it. lead with the metric.

post review draft

both writers · cross-routing is review traffic — never handoffs of research

2 authors live phase: draft yjs · 412 ops architect · idle