strategy-studio

Deterministic strategy routing: a business question goes in, a named strategy method comes out, with a scored explanation, an audit log entry and a worksheet.

Outcome: bash scripts/smoke.sh routes six sample questions to six different methods, checks the hash-chained audit log, and runs the accuracy gates on a labelled set and a held-out set.

What it does

Given a question such as "Are we profitable per customer once you count acquisition cost?", strategy-studio:

1. Scores the question against every method in src/strategy_studio/data/methods.json (keyword phrases with weights; no LLM, no network).
2. Picks the winner by score, then priority, then id, so the same question always gets the same answer.
3. Explains the choice: matched phrases, score, margin over the runner-up, and a confidence label (clear, narrow, tie-break, fallback).
4. Appends a hash-chained entry to a JSONL audit log (question hash, all scores, runner-up, methods-file version).
5. Emits a structured worksheet for the chosen method, with a calculator for unit economics and expected value.

Methods shipped as data:

id	Method
future-back	Future-back planning
ten-types	Ten Types of Innovation
jobs-to-be-done	Jobs to be done
pre-mortem	Pre-mortem
wardley-mapping	Wardley-style mapping
unit-economics	Unit economics (with calculator)
build-vs-buy	Build vs buy
reference-class-forecast	Reference-class forecast
real-options	Real options portfolio (with expected-value calculator)
org-design	Org design
frame-problem	Problem framing (fallback when nothing matches)

Add or tune a method by editing the JSON; strategy-studio --methods my.json ... loads a custom file.

Quick start

git clone https://github.com/mrodgersjs-web/strategy-studio.git
cd strategy-studio
python3 -m pip install -e ".[test]"
bash scripts/smoke.sh

CLI

strategy-studio route "We launch the new billing system next month. What could go wrong?"
#   Pre-mortem [pre-mortem] score=8 confidence=clear
#   Pre-mortem scored 8 from 'what could go wrong' (+5), 'launch' (+3). No other method matched any keyword.
#   audit: .strategy-studio/audit.jsonl #1

strategy-studio explain "Should we build or buy, and what could go wrong?"   # why every other method lost
strategy-studio worksheet "Are we profitable per customer?" \
  --set revenue_per_unit=40 --set variable_cost_per_unit=25 --set cac=120 --set retention=18
strategy-studio worksheet "..." --method pre-mortem --markdown              # force a method
strategy-studio audit verify                                                # recompute hashes and links
strategy-studio audit show
strategy-studio methods
strategy-studio eval --min-accuracy 0.95                                    # labelled (tuning) set
strategy-studio eval --holdout                                              # held-out set

All commands take --json. route takes --log PATH, --no-log and --ts (fixed timestamp for reproducible logs).

The original archetype router is still available as strategy-route "question" (A1-A4 cell and gates).

Accuracy, honestly

• Labelled set (data/labelled_questions.jsonl, 47 hand-written questions): 47/47. The keyword weights were tuned against this set, so it is a regression check, not evidence of generalisation.
• Held-out set (data/holdout_questions.jsonl, 22 questions written after the last weight change and not used for tuning): 20/22 on first measurement. Tests fail below 0.85.
• It is a keyword router. Questions phrased without any of a method's vocabulary fall back to frame-problem or land on the wrong method. explain shows exactly which phrases drove the choice.

Architecture

question -> score all methods (data/methods.json) -> rank -> decision + explanation
         -> audit entry (hash chain) -> worksheet (+ calculator)

Details: docs/routing-model.md, docs/architecture.md.

Public boundary

See docs/public-boundary.md. Sample questions are synthetic.

Roadmap

• Routing beyond keywords (for example a small embedding or classifier stage) gated on the held-out accuracy check.
• Gate checking: today worksheet gates are listed with passed: false; nothing marks them passed yet.
• Decision records that link a filled worksheet back to its audit entry.

Related

fde-portfolio · doctrine

Video walkthrough

• Script: docs/video-script.md
• Recording: assets/demo.mp4 (75s captioned; recorded against the earlier archetype router)
• Preview: docs/strategy-studio-demo.gif

Practices in this repo

Practice	Here
Smoke proof	bash scripts/smoke.sh (also runs in CI)
Tests	pytest -q
Lint	ruff check .
Public boundary	docs/public-boundary.md
Claim under test	same question -> same method, logged and verifiable
Fleet	profile · resume · patents

If scripts/smoke.sh fails, treat README claims as false until fixed.