stale · Day 01/14/14 · Ideate · Last heartbeat: agent · 21h ago
Background Research

How it works

A three-layer multi-agent system inspired by OrgAgent. Deterministic orchestration, bounded rounds, committee-gated phase transitions. This page shows the structure, the lifecycle and the ship gate.

01

Team and reporting lines

Three layers, nine roles. Governance sets direction and arbitrates. Execution produces work. Compliance reviews and convenes the author-persona committee.

Layer A — Governance

CEO
Project chair, risk mandate, arbitration, daily email.
Moderator
Per-committee meeting chair.
Conductor
Deterministic state machine; orchestrates every heartbeat.

Layer B — Execution

Drafter
Phase-aware author: ideate / expansion / consolidation / polish.
Narrator
Public posts — Bluesky, LinkedIn, Substack.
Librarian
Corpus gap-closer with WebSearch + WebFetch.
Moltbook Agent
Heartbeat + seed only; 4x/day → 2x/day.

Layer C — Compliance

Reviewer
Three sub-rubrics: style · fact · structure. Triggers LIBRARIAN-ACTION.
Committee — phase gates
Adam Grant
Evidence + assumptions
Ethan Mollick
Practice + calibration
Malcolm Gladwell
Narrative + insight
Moderator
Meeting protocol
Scribe
Verdict writer
orchestration + direction advisory verdicts CEO arbitrates across all layers; Conductor spawns all roles
02

Phase lifecycle

The essay moves through five phases. Each gate is committee verdict + CEO decision. Within each phase an inner Drafter ↔ Reviewer loop runs until the rubric clears or the rounds-budget ceiling is hit.

01
Ideate
Drafter produces outline candidates. Committee evaluates. CEO selects one and commits outline.md.
gate: ideate-gate
02
Expansion
Drafter over-writes to 130–150% of target. Reviewer critiques on three rubrics.
budget: 4 rounds
03
Consolidation
Drafter cuts back to 100–110% of target. Reviewer critiques again.
budget: 6 rounds
04
Polish
Drafter tightens to 95–100%. Final reviewer pass before ship gate.
budget: 4 rounds
05
Ship
Committee ship-gate, CEO arbitration, operator approval, publish to Substack + henrykernot.com.
gate: ship-gate
Expansion inner loop
1. Drafter over-writes 2. Reviewer 3-rubric critique 3. revise within budget
Consolidation inner loop
1. Drafter cuts 2. Reviewer 3-rubric critique 3. revise within budget
Polish inner loop
1. Drafter tightens 2. Reviewer 3-rubric critique 3. revise within budget
RESTRUCTURE → if any phase hits its rounds-budget ceiling without the committee gate clearing, the Conductor returns the essay to ideate. The escape valve against non-converging loops.
03

Review criteria

What the Reviewer and committee actually check. Two layers: objective signals sampled every round, and subjective judge scores on the three author personas' lenses.

Signals

Mechanical measurements. Logged every round. A red value doesn't halt — it flags for the Reviewer to weigh in.

Word count Target 2,500. Expansion rounds may exceed by 20–30%; Consolidation cuts it back.
Readability (Flesch-Kincaid) Grade 10–12. Outside this range, the register has drifted — too dense or too thin.
References Floor 12. Every empirical claim needs a corpus-backed citation or a flagged assumption.
Rounds remaining Budget per phase. Running out triggers CEO arbitration on whether to ship, revise, or restructure.

Judge scores

Subjective ratings from the three author personas. Each is scored 0–10 every round. Below 7 on any lens blocks the ship gate.

Style — Gladwell Narrative clarity, register, pacing. Does the piece move? Does it hold register?
Fact — Grant Evidence density, citation quality, assumption labelling. Are the claims load-bearing?
Structure — Mollick Argument architecture, practical utility, calibration. Does the reader leave with something usable?
04

Ship gate

The three-tier gate that lets an essay out into the world. No single agent — not even the operator — can ship alone.

01

Polish phase reached

CEO triggers a ship-gate committee convene. Three personas are spawned independently.

02

Independent reviews

Grant (evidence + assumptions), Mollick (practice + calibration), Gladwell (narrative + insight). No cross-talk in this round.

03

Reactive round

Personas respond to each other. Moderator checks for convergence. Scribe writes the verdict.

04

Verdict → proceed / revise / restructure

Revise: CEO directs targeted rework. Restructure: back to ideate. Proceed: continue to arbitration.

05

CEO arbitration

CEO reads the verdict. If CEO disagrees the override reason is logged in decisions/. Otherwise ship recommendation written to pending-human.md.

06

Operator approval

Human reviews pending-human.md. Reject → feedback.md loops back to drafter. Approve → publish.

07

Final essay rendered and published

Publish to Substack and henrykernot.com. Ship complete. Essay archived with full verdict trail.

05

Thesis abstract

Coming soon — full thesis on Substack.