Inside the enso lab we run controlled experiments on every platform that ranks, recommends, or distributes attention — LinkedIn, X, TikTok, Reddit, Substack, inboxes, search. Each dossier documents an observed seam in the wall, the signal measured leaking through it, and the operational protocol before the algorithm closes it.
- CLR · ω-04FILE 001 // LINKEDIN · distributionEYES ONLY
LIVE · DISTRIBUTION60 Minutes Inside the LinkedIn Algorithm: An Agent in 11 Cross-Promo Pods Lifted Median Reach 143× Across n=24 Posts.
LinkedIn's ranking system makes its viral/non-viral decision inside the first 60 minutes after publish. We pointed an agent at 11 long-running cross-promotion communities — places where members already trade engagement — and let it harvest reactions inside that window. The reach curve breaks at minute 18 and never returns to baseline.
live · n=24 postsDecrypt - CLR · ω-03FILE 002 // INSTAGRAM · dubbingEYES ONLY
FIELD STUDY · AMBASSADORA 48-Hour Field Study: A Three-Model AI-Dubbing Pipeline + 9-Second Retention Edit Recovers 230,185 Views Cross-Lingually on Instagram.
Composed ElevenLabs v2 → Sync 3 → CapCut pipeline clears the AI-smell threshold across all six perceptual attributes. v1 returned 1,703 views. A targeted 9-second retention edit on diagnosed cliffs (t=2s, t=8s) produced 230,185 views in 48h — 135× v1.
field study · n=4 variantsDecrypt - CLR · ω-02FILE 003 // REDDIT · karmaEYES ONLY
LIVE · THOUGHT LEADERSHIPHermes Agent, 1 Account, 11 Subreddits: 270 Karma in 1 Week, Zero Promotional Comments.
A single Hermes agent retains the full visible state of a subreddit in long context and routes through a high-bandwidth proxy mesh. 14-day comment survival: 87%. Downstream effect: 28 inbound calls.
pilot · n=11 subredditsDecrypt - CLR · ω-01FILE 004 // AUTHORITY · wikiEYES ONLY
LIVE · AUTHORITYA Glowing Seam in Wikipedia: 14 of 62 Citations Live at Day 30.
6-week pilot across n=9 client domains. No additions pitched — only documented maintenance gaps filled. 30-day edit survival: 23%. Downstream: LLM answer systems began quoting the citations back.
pilot · n=9 domainsDecrypt - CLR · γ-04FILE 005 // SDR · forecastEYES ONLY
DRAFT · AGENTICThe 90-Minute Window: Predicting LinkedIn Thread Saturation from Early Reply Velocity
A per-account forecaster that observes the first 90 minutes of reply arrivals on a LinkedIn thread and emits a stop/hold/push signal once audience saturation is statistically detectable.
training · v0.3Decrypt - CLR · β-11FILE 006 // SDR · promptEYES ONLY
DRAFT · PROMPTCastle 2.0: Anchoring the Second Touch to First-Touch Surface Yields a +38% Reply-Length Effect
A second-touch protocol that names the empirically observed first-touch surface. Reply volume invariant; mean reply length +38% with no change in send volume.
pilot · n=11 accountsDecrypt - CLR · α-02FILE 007 // INBOX · opsEYES ONLY
DRAFT · AI AGENTInbox Triage Classifier: 87% Top-1 Agreement With Human Triager on n=1,200 Replies
An LLM-based classifier that partitions inbound replies into warm, needs-human, and ignore, with a calibrated confidence score on every call.
internal use onlyDecrypt - CLR · δ-17FILE 008 // FIELD · captureEYES ONLY
SCRATCH · IDEAVoice-Note-to-Brief: Sub-30s Latency from ≤90s Voice Note to Structured Campaign Brief
Founder records a ≤90-second voice note; the agent returns a structured brief and three candidate opening lines before locomotion ends.
design dossierDecrypt
