Operating principles

If you’re new to investing, this page is the right starting point. If you’re not, it’s still the shortest description of what cents is trying to do that the rest of the docs assume you already understand.

Start with one idea: the thesis

A thesis is a written reason for owning a stock. Not “I think NVDA will go up” — something more specific, like “NVDA will go up because demand for AI chips will keep growing for the next 12 months.” That sentence is testable: a year later, you can look back and decide whether the reasoning held.

Cents makes you write down a thesis whenever you take a position, and attaches a conviction score (0–100) — how confident you are. A bunch of automated research agents help refine that conviction: one looks at fundamentals (earnings, debt), one at technical price patterns, one at the macro economy, one at news sentiment, and so on. Each one adjusts your conviction up or down based on what it finds.

That’s the boring core. The rest of cents exists to solve one specific problem.

The problem: the rules keep changing

The political and policy environment right now is unusually chaotic. Tariffs get announced, paused, struck down by the Supreme Court, then re-imposed under a different law — sometimes within weeks. Executive orders pile up. The dollar swings. In a calm period, you could buy a stock for a good reason and hold it for a year. Right now, the reason you bought might evaporate overnight because the government changed the rules.

This breaks a hidden assumption in normal investing tools. They assume the backdrop — the regime — is stable, so the only thing you need to track is price. In a regime like this one, that’s wrong. You also need to track whether your reasoning still holds.

The fix, in three pieces

1. The EventAgent: watch the rules, not just the company

A regular news agent reads articles about a specific stock — like searching “NVDA” on a news site. That’s fine for company-specific stuff (earnings, scandals), but it misses the big policy events that move everything. A new executive order on Chinese tariffs doesn’t show up in a search for “NVDA” — but it absolutely affects every chip company.

The EventAgent fixes that gap. It pulls from the Federal Register (the official US publication of every executive order, agency rule, and regulation), and tags each item with what it’s about — tariffs.china, fed_policy, energy_policy, etc. It maintains a running list of what’s actually happening at the policy level, drawn from a fixed vocabulary (EVENT_TAGS) of about 30 themes.

2. Premise tags: declare what your idea depends on

When you write a thesis (or when the factory writes one for you), it also gets tagged with the regime variables it depends on. For an NVDA thesis: ["tariffs.china", "ai_capex"] — “this idea works only if China tariff policy stays roughly where it is and companies keep spending heavily on AI.” For a JPM thesis: ["fed_policy", "rates"] — “this works only if Fed and rates stay roughly in the regime they’re in.”

The tags come from the same fixed vocabulary the EventAgent uses. That shared vocab is what makes the mechanical part possible.

3. Premise-invalidation alerts: get warned when the world shifts

When the EventAgent ingests a new event tagged tariffs.china, the system automatically cross-references it against every open thesis. Any thesis whose premise tags include tariffs.china gets an alert: “hey, an event just happened that may have killed the reason you bought this.”

Without this, you’d find out the same way you always do — the stock drops a few days later, you read the news, and you connect the dots. With this, the system warns you the moment the event is published, before the market has fully processed it.

A new outcome category goes with this: in addition to target hit (you were right, take profits), stop hit (you were wrong, cut losses), and horizon expired (your time window ran out), there’s now premise invalidated (the reason your idea worked just stopped being true). That distinction matters when you later look back — “I was right about the stock but the rug got pulled” is a different lesson than “I was just wrong.”

Why a regime snapshot too?

Every time a thesis is created, cents records a snapshot of the broader environment at that moment: how many policy events were happening recently, which themes dominated, the overall mood (bullish/bearish). It’s not used to make any one decision — it’s stored for later.

Why bother? Because eventually, with enough closed theses to learn from, you don’t want to just average them and say “my theses win 55% of the time.” You want to ask: “my theses win 75% of the time when the policy backdrop looks like Y, but only 35% when it looks like Z.” That kind of insight is only possible if you recorded what the backdrop was at the moment each thesis was opened. The snapshot is the seed.

Two cohorts: the control group matters

If cents is going to produce theses autonomously (via the factory), there’s a measurement problem: how do you know whether your wins are skill, or whether you just happened to run the factory during a friendly tape?

The answer is the same answer every science uses: a control group. For investing, that means policy-neutral theses — paired trades where you’re long one stock and short its sector ETF at matched dollar size. The pair cancels most of the regime exposure, leaving roughly just the company-specific signal.

Cents runs both cohorts in parallel by default:

• Directional cohort — single-leg theses that take a clear regime view (long X because tariffs / rates / AI capex).
• Neutral cohort — paired theses that hedge out the regime exposure, capturing whether the underlying name actually outperforms its peers.

The gap between the two cohorts’ returns is the value of taking regime bets, holding skill constant. If the directional cohort returns 0.8 Sharpe and the neutral cohort returns 0.7, the actual edge is 0.1, not 0.8. Without the neutral cohort, you can’t tell skill from tape. With it, you can.

Two arms: the random orchestrator

The paired-neutral cohort controls for regime beta. It does not control for signal value — the question of whether the LLM-driven orchestrator did anything useful at all, or whether opening theses on the same cadence with random conviction would have produced the same cohort spread.

Cents addresses that with a second control: a RandomOrchestrator arm. cents factory run --orchestrator random runs the loop with a uniform-random conviction_delta and no LLM calls. Every thesis it opens is tagged orchestrator_label = "random"; the default LLM arm is tagged "llm". Run them on the same universe at the same cadence and the gap between the two arms is the value of the LLM signal, isolated from everything else.

This is the design move that makes the pipeline a research experiment rather than a story. If the LLM arm doesn’t beat the random arm at matched N, the headline conclusion is that the LLM signal didn’t earn its operating cost — which is a useful conclusion, just not the one the system was for.

Pre-registration makes it falsifiable

Cohort spreads and per-arm win rates are vulnerable to a different kind of cheating: post-hoc selection. Run the factory ten different ways, look at all the splits, pick the one that looks best, and write it up. The result will always look good — and won’t replicate.

cents experiment register <spec.yaml> blocks that move. A spec declares the hypothesis, the primary metric, and a minimum_n_per_arm target before any theses open. Registration snapshots the current factory.toml SHA so the parameters are frozen too. Every thesis opened while the experiment is active carries the experiment_id. At the end, cents experiment finalize <name> locks the verdict against the original spec — not against whatever cut of the data flatters it.

The autonomous factory

The factory is the loop that runs all of this autonomously. It walks a universe of symbols (often produced by a screener), asks the orchestrator who looks interesting, opens paper theses (paired by default) for those that clear an entry threshold, watches them, and closes them on target / stop / horizon / premise-invalidation. Every thesis it creates gets the full regime-aware treatment: premise tags, regime snapshot, sector-ETF hedge, and a record of the discovery source (the universe — and hence the screener — that produced its symbol).

The factory’s goal is not P&L. It’s to produce a labeled outcomes dataset, stratified by cohort and regime, so that after enough resolved theses you can mine the data for what kinds of trades cents-the-system actually generates well in which regime conditions. The eventual “income algorithm” comes out of that map — it doesn’t exist yet, and shouldn’t, until the data is big enough to support a real conclusion.

This is why the factory tracks:

• Outcomes by closure reason, not just win / lose
• LLM cost per run, so you know the operating bill
• Per-thesis regime snapshot, so future analytics can stratify
• A PREEMPTED outcome (excluded from win-rate math) so theses that got displaced by higher-conviction candidates don’t poison the stats
• A concentration cap per premise tag, so three financials all depending on fed_policy don’t get counted as three independent bets when they’re really one Fed call

The mental model, condensed

• Old way: “I bought stock X for reason Y. Did the stock go up?”
• New way: “I bought stock X for reason Y, which depends on policy environment Z being stable. Did Y still hold by the time the position resolved? Did Z shift? If yes/no, what does that tell me about my next decision?”

It’s more bookkeeping upfront. But the bookkeeping is what eventually lets you tell the difference between a good investment process and being lucky to live in a good time to be invested.

Where to go next

• Quickstart — install and run your first scan
• Architecture — how the pieces fit together, with diagrams
• Agents — what each research agent does, including the EventAgent
• Factory — the autonomous loop in detail
• Roadmap — what’s done, what’s queued, what’s deliberately out of scope