Synthesis agent
N mediocre inputs.
One better artifact.
synth scores N candidate drafts with deviate-style metrics, selects the top 3 by robust score, then pulls the highest-marker-density sentences from that pool into one synthesized candidate — provably scoring above the single best input on both anchor and rupture. Not an average. A strict improvement.
[higgsfield: three overlapping thin gold circles converging on one bright core point, black field, engraved line art]
synth · coral-reef synthesis · 6-candidate corpus
Why teams install
What synth earns back
Beats the best input, not just the average
The synthesized candidate scores 56.67 anchor / 42.0 rupture against a max-input ceiling of 29.31 / 31.5 — a real margin, not a tie at a shared cap.
‹74f6174a›Sentence-level density, not verbatim concatenation
Every sentence in the top-3 pool is ranked by marker density per word — a short dense sentence beats a long diluted one — before selection.
‹13421965›Naive concatenation fails, on purpose
The planted failure disables elite selection and falls back to concatenating full top-3 texts — filler dilutes the density below the max input, and the gate goes red.
‹9dcba94b›How it works
Three steps. No faith required.
01
Score
Score every candidate with the deviate-style engine family (graviton / anchor / rupture / echo).
02
Select
Rank by robust (median-of-engines) score and take the top 3, or fewer if fewer than 3 candidates exist.
03
Synthesize
Alternately pull the next-best anchor and rupture sentence from the top-3 pool until the mesh clears the strongest single input on both engines.
Proof band
One packet. Every claim accounted for.
packet rig/synth · corpus 6-candidate claim "synthesized candidate beats max(inputs) on anchor AND rupture, deterministically" top3 candidates [5, 1, 0] by robust score mesh "A 2021 benchmark (n=210) measured a 22% lift in output. Actually, the conventional wisdom here is wrong." anchor 56.6667 vs max input 29.3103 · margin +27.3564 rupture 42.0 vs max input 31.5 · margin +10.5 gate PASS — both engines cleared themis SEALED — afe51b5f
Every synth run emits a packet like this — hash-chained, verifier-signed, and queryable from your own dashboard. How ProofPackets work →
Objections
Answered before you ask
Isn't this just picking the best draft?
Picking the best draft caps you at your best input. synth pulls the elite sentences across your top 3 into one artifact that provably outscores every single input on evidence density and rupture — a strict improvement, not a selection.
Won't concatenating sentences from 3 drafts read like a patchwork?
That's the planted-failure mode, and it's exactly what synth avoids — naive concatenation dilutes density with filler and fails the gate. The real mechanism ranks by marker density per word, not raw length.
Why trust the "beats every input" claim?
The margin is computed against the max of ALL N inputs, not just the top 3 — and re-verified in a fresh scoring pass, sealed into the ProofPacket. Fewer than 3 candidates, or the mesh failing to beat the max, and the run doesn't seal.
Buyers
Who runs synth
Head of Content Ops
publishing at volume
Creative Director
agency, brief-to-draft pipeline
Product Marketing Manager
launch copy at scale
Pricing
Start free. Pay when it pays.
Generalist ships today. Domain Expert adds vertical memory + priors. Custom is built to your stack.
Free
Priced after a short conversation
10 syntheses a month, 3-candidate max.
Team
Priced after a short conversation
Unlimited syntheses, up to 20 candidates, private corpus.
Pro
Priced after a short conversation
Dedicated verifier, custom scoring lexicons, SSO.
Team pricing is set after a short conversation. No countdown — the price re-prices once the cohort fills.
Runs well with
