Model Changelog

Every methodology change, published openly.

Quant models are only trustworthy when their changes are visible. This page records every material change to PRISM's scoring methodology, validation rules, and historical comparability — including the ones that made our numbers look worse.

July 2026 · Runtime observability & performance

Measured, cache-aware runtime — verified not to change any score

The pipeline now archives a full runtime audit for every run (not just the latest), with a durable index, per-phase timings, and explicit detection of machine-sleep / stall time so a 10-hour wall-clock is never confused with active compute. This makes monthly rebalances and daily runs separately comparable over time.

Alongside it, the system is being made cache-aware — reusing unchanged evidence and price history instead of recomputing — and a per-provider concurrency gate keeps upstream data sources from being overloaded. A fixed benchmark re-runs the same set with caches off and on.

Crucially, a cache-transparency check asserts that final scores are bit-identical (within tolerance) with caches on versus off. These are speed, reliability and measurement changes only — scoring methodology, model regime, and published performance are unaffected, and we verify that rather than assert it.

July 2026 · Data & signals expansion

Market Signals layer + independent research services

A new Market Signals page surfaces ownership, flow and short-interest leaderboards, and every company page now shows an Ownership & flow panel for its ticker. These are fed by a set of independent, read-only services: short interest & squeeze pressure (FINRA), insider flow, congressional trades, an FDA-catalyst calendar, and a macro snapshot (FRED) — plus native boards for buybacks, 13F smart money, capital flow, policy exposure, fundamental quality and risk flags. Each feed is optional: if one is offline it simply shows as unavailable and nothing else is affected.

An ArkenLabs MCP server was also added, exposing this public data (rankings, boards and feeds) read-only to AI tools such as Claude and Cursor.

These are data-surface and presentation additions built alongside — not inside — the PRISM model. Scoring methodology, model regime, and published performance are unaffected.

July 2026 · Rebrand & interface

Official name: ArkenLabs

The public research product is now published under the official name ArkenLabs (arkenlabs.eu). The underlying scoring model keeps the name PRISM — ArkenLabs is the brand, PRISM is the model.

Alongside the name, the interface received a full visual redesign: a new octahedral gem logo, a single light “etched blueprint” theme on a khaki / burnt-orange / plum palette, monospace technical labels, faint schematic watermarks, and refreshed navigation. This is a presentation change only — scoring methodology, model regime, and published performance are unaffected.

June 2026 · Validation hardening

Backtest realism upgrade (numbers revised down, on purpose)

A full audit of the research pipeline led to four methodology changes that make reported performance more honest and strictly harder to achieve:

1. Entry timing: simulated entries now use the first trading day after a scoring snapshot, never the snapshot day itself, removing a same-day information edge.
2. Transaction costs: backtests now support explicit per-rebalance trading costs instead of assuming free trading.
3. Benchmark alignment: portfolio and SPY metrics use actual common observation dates instead of stale forward-filled prices.
4. Point-in-time inputs: holdings and evidence remain anchored to the snapshot that was available at the decision date.

These changes intentionally make the research record more conservative and easier to reconcile.