> ## Documentation Index
> Fetch the complete documentation index at: https://docs.abacuslabs.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Data, freshness and limits

> Where the numbers come from, how fresh they are, how to read provenance, and what Abacus Labs is not.

## Not investment advice

Abacus Labs is software that computes. It does not know the person's goals, other assets or circumstances beyond what you send. It does not forecast returns, recommend securities, hold assets or place orders, and it is not a broker, an investment adviser or a tax adviser. Options are consequences of the rules in the request. Tax figures are estimates at the rates you send. Show the person every option and let them decide.

## Freshness

| Data | Source | How fresh |
| - | - | - |
| Prices | Databento market data | Cached up to 15 minutes within one market state. Outside market hours, the last close. Each quote carries its session (`as_of`, for example `"2026-11-13 close"`), time and market state. A quote older than 5 trading sessions is dropped. |
| Risk model | Abacus Labs' factor model: market, industry groups and 13 style factors; ETFs looked through to their holdings | `provenance.risk_model.data_end` is its last day of data. |
| Fund holdings (ETF benchmarks) | SEC N-PORT reports, rolled forward with prices | Public about 60 days after the month they describe. |
| Manager books (`{"manager": ...}` benchmarks) | SEC Form 13F, rolled forward with prices | Up to 45 days after the quarter; trails the real portfolio by up to about four and a half months. |

Symbols with no price are named in `warnings` and left out of value, weights and risk.

## Provenance

Every analyze, proposal and trade check answer has `provenance` and `as_of`:

```json theme={null}
"as_of": "2026-11-13 close",
"provenance": {
  "risk_model": { "version": "us_fund_v1", "vintage": "us_fund_v1@gen:cb849b0e8d02d6a0", "data_end": "2026-11-13" },
  "prices": "databento; each quote carries as_of (session label), quote_time (ISO, America/New_York), market_state (open|pre|post|closed) and source; cached up to 15 min within one market state",
  "engine": { "name": "layered", "solver": "thoropt.solve_layered", "seconds": 8.67 },
  "data_quality": { "unknown_cost": [], "unknown_term": [], "unresolved_symbols": [] } }
```

`engine` (proposals) says which engine ran. `data_quality` lists holdings with unknown cost, unknown purchase date, or no price; `warnings` says what each gap means for the numbers.

## Engines

The layered solver (exact rules and a proven bound) answers harvesting, investing cash, raising cash, cutting a position or sector, lowering risk, tracking a benchmark and rebalancing. Rule-based templates answer themes, more risk, and accounts the risk model does not fully cover; their options have no `guarantee` block. Every option passes the same rule checker.

## Reading goals and questions

A language model with a JSON schema reads goals into a policy and chooses the analyses for a question. `understood.policy` shows what it read; `understood.read_from_text` lists the fields it took from the text, and `understood.confidence` how sure it was. Check `understood.policy` before you show options. To remove doubt, send the fields as `policy` yourself.

## Wash sales

30 days before and after a loss sale, across every account in the request and the saved accounts linked to it (IRAs included), using lot dates, `recent_trades` and committed loss sales. Share classes of one company (GOOG and GOOGL) count as the same security. Funds that track the same index count as identical unless `wash_sale_policy` is `exact`. Abacus Labs cannot see accounts you do not send.

## Tax

Estimated with the rates in `policy.tax.rates`. Without them: top federal rates, 40.8% short term and 23.8% long term, including the 3.8% investment surtax; `warnings` says so.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.