{lede}
: null}{eyebrow}
: null}{lede}
: null}Principles
Most money apps manufacture urgency: flashing tickers, confetti on gains, red alerts on dips. Talunavo does the opposite. Nothing here can express "hurry". Deadlines are dates. Red is spent on exactly one thing: a goal that is short, where a decision is wanted.
A single projected number, "you will have \$3.4M", is false precision that invites false confidence. The unit of value is the difference between two plans under identical assumptions, where the shared modelling error largely cancels out.
Any figure can be drilled to the year, the line item, the rule that produced it, and, for tax figures, the primary source the rule was derived from.
Never buried in code. You can see, in one place, every assumption a result rests on, and change any of them. Every scenario input carries its plausible band, and every tax figure its confidence tier.
Adopting is the beginning, not the end. The system's job afterwards is to tell you, calmly, annually, not daily, whether the assumptions you bet on are still holding.
Talunavo models arithmetic under stated assumptions. It never phrases output as a recommendation, only as a comparison. The export target is a document you take to a meeting, and the language is comparative rather than imperative everywhere.
No forms, no e-file, no attempt at completeness on edge cases. The tax engine is a planning-grade approximation and says so. What each strategy does not model is written down rather than left to be discovered.
It evaluates plans you write. Automated search over the strategy space can come later, once evaluation is trusted. A tool that hands you an answer before you trust its arithmetic is worse than no tool.
Explicitly the opposite. The goal is a better-informed client who asks sharper questions. Its outputs are an education, a shortlist, a set of sharper questions, and, after a decision is made elsewhere, an early warning system.
You type your data in; no brokerage credentials, no account scraping. Planning reads the portfolios you track and never writes to them. Self-hosting is a feature, not a chore, for the privacy-conscious household that wants its own model on its own machine.
A test walks every shape in the plan comparison and fails if a score, rank, or composite field appears.
The store refuses to write a payload containing that vocabulary, and both test suites walk the payloads hunting for it.
A funding track is measured against the path stored at adoption, never a fresh run. A target recomputed at read time would move to meet you.
When a watch's condition binds, it produces a finding. The six verbs that would do more are refused by name, with a test walking every served field.
Funded, tight, short, needs input are glyphs, legible in greyscale, in sunlight, printed, and to a colourblind reader. If removing all colour destroys the meaning, the design is wrong.
No number without the plan, scenario, and real-or-nominal it was computed under. Nominal roughly triples every figure over a horizon, so the frame is never implied.