Capital OS
A low-attention investment intelligence system that watches a portfolio, speaks only when a decision matters, and remembers why — with deterministic safety, human-gated advice, and judgment improving from outcomes over time.
Capital OS
Most portfolio tools either drown you in charts or promise autonomous trading. This project asks a different question: what if the system watched quietly, explained only what matters, and got smarter from every decision you accept or reject?
Capital OS is a low-attention investment intelligence system for a non-expert investor. It syncs portfolio state, filters market noise into a few admitted candidates, and produces structured advice — observe or hold — with evidence, risk, invalidation, and confidence. An optional language model may rewrite presentation prose; it never owns sizing, permissions, or market truth. Quiet is a valid conclusion.
The problem
Retail investing tools tend to fail in the same places:
- Attention tax — platforms reward constant looking, not disciplined waiting
- Advice without memory — recommendations appear, then vanish without a trail
- Outcome amnesia — wins and losses never feed back into better judgment
- Autonomy theater — “AI trading” skips the hard part: earning trust under risk budgets
- Model overreach — language models asked to decide money when they should only explain
Those are system-design problems, not “pick a smarter model” problems. Capital OS treats them that way. We optimize for alignment, not autonomy.
Product thesis
The product encodes capital discipline as durable operating memory: portfolio health, admitted market candidates, decision journal, active theses, and outcome checkpoints.
Intelligence here is not “the model picks winners.” It is:
- watching continuously while interrupting rarely
- routing only filtered, portfolio-relevant context into advice
- keeping risk, environment, and execution gates in deterministic code
- closing the loop from recommendation → feedback → outcome → calibration
Advice comes before autonomy. Higher agency is earned through evidence — reliable data, useful recommendations, calibrated confidence, and predictable behaviour — not enabled by a flag. Chat is a bad system of record; the journal and thesis store are where judgment compounds.
What we built
An observer OS, not a trading terminal
The first vertical slice answers the questions a busy investor actually needs:
- Is my portfolio healthy?
- Has something important changed?
- Do I need to act?
- What does the system advise, and why?
- What could go wrong, and when is the thesis invalid?
The dashboard is an attention surface, not a pro desk. Market ticks pass filters and quiet suppression before anything becomes a candidate for advice.
Deterministic core, optional AI voice
Structured action, urgency, and confidence are produced by code. Language models are off by default and, when enabled, may only rewrite already-structured brief prose. Raw market ticks never enter an LLM. Fail closed: if the rewrite fails, the deterministic brief still stands.
Safety boundaries are product features: demo / read-only first, real execution prohibited at this stage, GET-only broker transport, full operational audit with secret redaction. That is the human task lane applied to capital — honest limits, not invented autonomy.
Decision memory that compounds
Accepted advice becomes durable institutional memory:
- a decision journal that records recommendation, reason, evidence, and user response
- active theses with explicit invalidate / close lifecycle
- scheduled outcome check-ins (end of day, 7d, 30d)
- calibration views that separate decision quality from outcome quality
The system is designed to learn from sparse, honest feedback — not from chasing every tick. That is living memory for investment judgment.
Advice before autonomy
Observer recommendations are constrained to observe or hold. Buy, sell, and rebalance remain reserved for later modes behind explicit policy gates. The product promise is discipline and awareness first; execution autonomy is a later privilege, not the launch story.
Principles that hold the line
- Quiet is a valid conclusion.
- Advice before autonomy.
- Explain every important recommendation — action, evidence, horizon, risks, invalidation, confidence, and the cost of doing nothing.
- Deterministic code owns money-adjacent truth; models may explain, not authorise.
- Decision quality and outcome quality are recorded separately.
- Human control over capital and risk is non-negotiable.
Honest V1 boundaries matter too: no push or email delivery of due prompts yet, no automatic confidence recalibration, no automatic thesis invalidation from market rules alone. The learning loop is real; full self-tuning is earned later.
Why this belongs in the Lab
Capital OS is proof of a pattern Nimbus cares about: process + memory + human gates as the product, not raw model cleverness.
It shows domain intelligence with hard safety boundaries, selective interruption instead of dashboard noise, institutional memory in a decision journal and thesis store, and a deliberate stop where judgment remains human. The secondary audience is anyone building AI systems where the cost of a wrong action is real — and where “do nothing” must be a first-class outcome.
The adviser compounds. Every accepted recommendation, check-in, and closed thesis makes the next interruption more trustworthy — and the journal keeps the receipts.