Foundations
First automation loops, Sheets as a living desk, and the habit of measuring harvest quality rather than noise. Plenty of iteration before the formal trial clock started.
Horizon
A private Lab experiment in algorithmic spot accumulation.
Horizon is how I study steady compounding on spot coins: accumulate within gates, harvest when the plan allows, and protect free capital before chasing growth. Built for my own accounts — not a product, not for clients, and not financial advice.
Timeline
Early build work began in January 2026. The formal twelve-month trial opened on 1 August 2026 and runs through July 2027. Below is the arc — not every deploy, just the shape of the journey.
Twelve-month trial opens
A deliberate year of live learning: same personal accounts, clearer rules, and a promise to prefer resilience over spectacle.
Hard lessons from drawdowns
Volatile stretches exposed unbounded accumulation and thin free capital. That forced entry caps, capital guards, and a slower, more honest scoreboard.
Lean dual-Lambda stack
Main cycle plus a dedicated market-data worker — Python 3.13 on ARM64 — paced by EventBridge Scheduler, with DynamoDB locking so runs never overlap, and Secrets Manager for credentials.
Steady harvesting under fixed dials
Live soak with frozen risk dials: maker-first harvests, sequenced adds, and capital guard armed. Watching Balance growth quality day by day — not inventing new gates every week.
What changed
A lot of machinery came and went between January and the trial. The lasting shifts are few and deliberate:
Scoreboard honesty
Focus moved from fill notional to daily grow and capital health — whether free capital can still breathe.
Entry caps (8×)
Hard caps versus entry size quarantine bleeding names and stop infinite accumulation on the same spot coin.
Capital guard (~35% floor)
If available capital as a share of equity thins too far, new entries pause until the book can recover room to manoeuvre.
Maker-first harvesting
Unlock harvests prefer resting limit fills — typically around a ~99% maker-fill ratio among planned harvest events.
Sheets as control plane
Google Sheets stays the reactive operator UI — dials, live mirrors, proposed actions — while AWS carries the compute.
Highs, lows, and the growth showing now
High — the stack got quiet
Lean dual-Lambda pacing, a real mutex, and a Sheets desk that explains each cycle. Less drama in the infrastructure means more attention for the book.
High — harvest discipline
Staged unlocks with a min-profit floor and maker preference turned “close something” into a repeatable habit instead of a scramble.
High — growth under soak
With dials frozen, daily grow has been printing more consistently: free capital often sits comfortably above the guard floor (recently around the 80%+ available band), while harvests keep banking without inventing new risk switches.
Low — unbounded adds hurt
Early on, averaging without a hard entry cap trapped margin in names that would not recover quickly. That lesson is why the 8× cap stays locked.
Low — one bad kill-switch day
An aggressive funding-bleed eject crystallised a painful cluster of closes. The soak response was simple: leave that live kill-switch off, keep hourly funding protection on, and prefer shadow telemetry before live ejects.
8× entry cap ~35% capital floor ~99% maker harvests 80%+ available capital in recent soak
Lessons so far
Protect the floor before you feed
New names and adds only matter if free capital survives. Capital guard and name caps beat clever entries.
Fewer dials, longer soaks
Changing risk switches every bad afternoon destroys signal. Freeze a small set, measure Balance grow, then decide.
Shadow before crystallise
Telemetry on deep-red names teaches recovery behaviour. Live mass-closes teach regret.
Architecture is a risk control
Non-overlapping cycles, secrets out of the sheet, and a desk you can read under stress matter as much as the signal logic.
What I may try next
Still inside the August 2026 – July 2027 window. Nothing below is a promise — only a short list of careful experiments once the current soak has said enough:
Tighter capital discipline after feeding slows
Only once the book stops growing names — refine the auto-deposit / available-capital posture without inventing new skip gates.
Selective recycling of stuck stubs
If evidence supports it, carefully free trapped margin on post-trial opens only — never a mass close of the deep-red book.
Leaner operator surface
Keep the dozen dials that matter; merge or freeze the rest so the Sheets desk stays human-scale.
Under the Lab
Built by Phil Stone
Private Horizon work in Northern Ireland. Meet Phil · Get in touch
Core automation sits on a lean AWS serverless stack: dual Lambda (Python 3.13 on ARM64), EventBridge Scheduler, DynamoDB for hot state and locking, Secrets Manager for credentials, and Google Sheets as the reactive control plane. Strategy detail and results stay private.
Lab boundary: Horizon is personal research for Phil Stone’s own accounts. Trading spot assets involves risk, including the possible loss of capital. This page is an introduction only. It is not financial advice, not an invitation to invest, and not a product for clients.