Skip to content

positioning.md

Latest commit

 

History

History
79 lines (51 loc) · 7.2 KB

positioning.md

File metadata and controls

79 lines (51 loc) · 7.2 KB
Raw

QuantMind Positioning

Quick Summary

  • Purpose: State what QuantMind is, who its consumer is, and the two engineering dimensions (context and harness) that everything else — README, slides, skills, decks — derives from. This is the single canonical positioning source; when prose elsewhere disagrees with this page, this page wins.
  • Read when: Writing or reviewing any outward-facing description of QuantMind (README, a talk, a skill, an issue), or deciding whether a feature belongs to the shipped story or the roadmap.
  • Load next: For the harness mechanics, ../dev/harness-engineering.md; for operation naming used throughout, operations/naming.md.
  • Authority: Current as of July 2026. This page supersedes docs/superpowers/specs/2026-04-16-quantmind-positioning/ (the April "personal knowledge base / MCP product" framing), which is kept only for history. Do not cite the April spec.

Contents

Positioning and Hero

QuantMind is an agent-native workbench for financial knowledge extraction. The primary consumer of this repository is not a human importing a package — it is a coding agent working inside the checkout. You describe the pipeline you want; the agent builds it here, against the repo's contracts, skills, and verification.

Hero line: "Don't import it. Open it." The intended first move is cd quant-mind && claude (or codex), not pip install. QuantMind remains a perfectly good importable library — PaperFlow(cfg).build(input), collect_news, batch_run are real Python operations — but the framing is workbench-first, library-second. The workbench is where an agent turns a loose request into a typed, cited, reproducible extraction pipeline.

Two engineering dimensions structure the whole project — dimensions of one repository, not versions of a product. Context engineering turns any source into typed knowledge. Harness engineering turns any agent into a domain specialist. The first is what the library does; the second is what the repository is. (The July 2026 poster labels the pair "V1 / V2" as a contrast device; serious prose uses the dimension names, never the version labels.)

Context Engineering

Any source → typed knowledge. This dimension lands unstructured financial content in typed, cited, as-of-correct knowledge that downstream retrieval can trust.

  • Deterministic preprocess — fetch / parse / format + clean produce source-faithful values with no model in the loop, so the provenance is exact and replayable.
  • Config-driven operations — PaperFlow(cfg).build(input) binds an immutable build config once and applies it per input; collect_news collects a replayable source window; batch_run fans any operation across a list of inputs under one unified setting. Callers do not write asyncio.gather boilerplate.
  • Typed knowledge shapes — a Paper structure tree for whole-document artifacts, and flat cards for News / Earnings / Factor / Thesis. Every artifact is self-contained: it carries its own text, an as_of timestamp, and a light source ref, so it persists and time-queries standalone.
  • Retrieval over that knowledge — rag/ (deterministic chunking + BM25 / similarity), library/ (local persistence + meaning-based search), and mind/ (agentic, reasoning-based retrieval where an LLM decides). Together they serve RAG and Agentic RAG, deep research, and data-MCP serving.

Context engineering is the substance shipped as the NeurIPS 2025 GenAI-in-Finance workshop paper (arXiv:2509.21507).

Harness Engineering

Any agent → domain specialist. Harness engineering is the claim that the repository itself — its contracts, disclosure structure, skills, hooks, and verification — is what upgrades a general coding agent into one that reliably does QuantMind-quality work.

  • Repo-level contracts — AGENTS.md / CLAUDE.md state the always-on rules once, in one source, for every agent that opens the repo.
  • Progressive-disclosure contexts/ — agent-facing reference pages with a Quick Summary / Contents preview, so an agent loads only the one page a task needs instead of the whole design corpus.
  • Portable skills — quantmind-dev ships today (commit / PR / component-development workflow); quantmind-best-practice is planned.
  • Claude + Codex hooks — shared hook scripts give both agents identical hard guarantees (for example, no verification bypass) without maintaining two copies of a rule's content.
  • Deterministic verify — scripts/verify.sh runs lint + types + import boundaries + tests, fast-failing in a fixed order. CI runs the exact same script, so a green local run means a green PR.

The mechanics — which enforcement layer catches whom, how one rule reaches both Claude and Codex — are detailed in ../dev/harness-engineering.md.

The Bet

The wager behind harness engineering, stated plainly: a weak model in a good harness beats a strong model running bare. If that holds, the durable asset is not the model of the month but the repository that mounts around it — the contracts, contexts, skills, and verification that keep any agent on the rails.

Evaluation In Design

Evaluation is in the design phase. No results are claimed here; two benchmarks are being designed, one per dimension.

  • Knowledge eval (context engineering) — correctness (EM / F1), citation precision and recall, point-in-time correctness (no look-ahead), and task lift. Protocol anchors under consideration: FinanceBench, FailSafeQA, and RAGAS faithfulness.
  • Harness eval (harness engineering) — SWE-bench-style paired runs: the same model and the same tasks, run bare versus run inside the mounted repo. Metrics: cost-to-green, pass@1, pass^k across seeds, and wall-clock.

Run instrumentation — tokens, cost, duration, and a verify oracle — already ships. The benchmarks that consume it do not yet exist; treat any evaluation claim as a plan until this section says otherwise.

LLMQuant Data In Production

LLMQuant Data is QuantMind in production. The hosted data platform runs extraction pipelines powered by QuantMind; QuantMind is the open engine, LLMQuant Data is the operated product on top of it.

  • The dependency direction is one-way: llmquant-data imports quantmind, never the reverse. Nothing in this repo may depend on the hosted platform.
  • "The knowledge harness for AI-native finance" is LLMQuant Data's slogan, not QuantMind's. Do not attach it to QuantMind.

Roadmap

The following are planned, not shipped. Describe them as roadmap; never present them as current capability.

  • quantmind-best-practice skill (companion to the shipped quantmind-dev).
  • SEC / filings collection flow.
  • Prediction-market knowledge type.
  • Concrete GraphKnowledge types (today a placeholder).
  • The two evaluation benchmarks described in Evaluation In Design.