PRODUCT & BUSINESS REASONING, WITH EVIDENCE
From a business question to research delivery.
Collect the data, validate the judgments, give the next step something to stand on.
Askovo splits a product judgment into an inspectable reasoning chain: every judgment states why this step follows from the previous one, and which evidence carries it. Evidence is graded E0–E5; when a conclusion outruns its evidence, askovo audit names it outright, and exit code 1 can gate CI. Built for product managers, user researchers, user operations, and market/requirements/business analysts, covering software, hardware, consumer goods, automotive, and mobile.
One chain, from user reasoning to businessSCENE → BUSINESS
Data collection, interviews, experiments, and platform connectors are supporting steps only — they validate a node or an inference edge; they never replace the reasoning itself.
Every node must record: the current statement, the rationale, the assumptions it rests on, the supporting evidence (evidence_ids), and a status. One machine truth (product-reasoning.json) projects into the MRD, PRD, BRD, and decision memo — not four parallel sources of fact. Six kinds of conclusions are never mixed: need does not pose as acceptance, usability does not pose as commercial, and stated intent is never paying behavior.
What is available todayCAPABILITY BOUNDARIES
| Capability | Current boundary |
|---|---|
| Seven-node chain and decision package / MRD / PRD / BRD projections | Structure and evidence references are validated automatically; semantic review stays with humans |
| Evidence-level audit (E0–E5, exit code fits CI) | Checks structure and configured gates only; proves neither demand prevalence, market success, nor release feasibility |
| Platform radar and collection (GitHub / Reddit official APIs) | Blocked platforms are never bypassed; CDP is limited to authorized visible pages; the Design gate comes before any collection |
| Unit-economics and PSM price-sensitivity projection | Deterministic calculation; PSM is stated intent, not paying behavior |
| Human delivery (need-validation study / team workshop) | Scope and fee agreed per specific task; commercial repeat purchase and scaling are still being validated |
The engine, CLI, Agent Skill, MCP, and the five base domain packs are MIT open source, with no feature-gated edition. Only deep domain assets and human delivery can be paid — never the engine itself.
How need validation worksNEED VALIDATION
A user saying "I want it" is only E1, a verbal statement. To reach "validated", a judgment needs at least one piece of E2 — an observed substitute, workaround, or failure behavior. The full method is in "When a user says 'I want it', is it a need?": four common misreadings, a substitute-behavior checklist, and the E0–E5 criteria — usable without installing any tool.
A concrete task: five demand judgments, two backed only by words
A synthetic example, to show how a judgment lands and how a gap is marked — not real customer data, and not a market size. Rule source: docs/concepts.md.
| Demand judgment | Evidence on hand | Marker | Next validation step |
|---|---|---|---|
| Adult children will pay to "check on an elderly parent remotely" | 3 interviews saying "I would buy"; no actual payment | E1 · needs validation | Price anchor: a real reservation deposit, not a survey tick |
| Night-time blind spots push users to build their own substitute | 2 user posts: adding a second camera, getting up at night to check | E2 · validated | Add 5 more links to original posts in the same scene |
| A privacy switch is a purchase precondition | 1 explicit refusal: no physical shutter, no purchase | E4− · counter-evidence (single case) | Write "physical shutter" into the acceptance conditions, then find 3 more like it |
| Willing to pay 30% more for anomaly detection | 60% of survey respondents ticked "willing" | E1 · needs validation | A real price-quote experiment; stated intent is not payment evidence |
| Installation complexity is one reason people give up | 3 return or bad-review records pointing at installation | E4− · counter-evidence | Record it separately from "unwilling to pay", so the attributions do not blur |
The two judgments with only E1 do not get upgraded because "many people say so"; a cell with no evidence says "unknown", never "zero". That is what askovo audit does: a judgment marked "validated" whose question chain holds only E1 is named and recommended for downgrade, and exit code 1 can gate CI (regression tests in tests/test_askovo_cli.py). Run askovo demo locally to see one full audit output.
Verify it locally in 30 secondsNO SIGNUP, NO UPLOAD
An offline synthetic sample: it verifies that the structure and the reasoning chain are self-consistent — it proves nothing about real demand or the market. Python 3.9+, standard library only; published on PyPI, with the PyPI page as the version of record (checked 2026-11-14).
pipx install askovo
askovo do "home security camera false alerts at night, want to validate the prototype" --domain hardware --dry-run
askovo demo
Expected key results: status: pass · supported: 1 · overclaims: 0. Every command runs locally and uploads nothing. The same engine plugs into your agent workflow as an Agent Skill or an MCP server.
Engineering and claims disciplineCLAIMS
Unit tests plus a Python 3.9–3.12 CI matrix, and an automated release flow (the test count is whatever the CI badge says). The synthetic sample proves structural self-consistency only; on the claims page we list, item by item, what each number and conclusion can claim and what it must not pose as — that is the same product as the engine itself.
GitHub repository · Quickstart · llms.txt (agent entry) · CI