seedscope

SeedScope

AI-native idea validation pipeline — from raw signal to validated MVP spec in Pure Lua.

Built entirely on algocline bundled packages. No Python, no external APIs, no infra — just Lua + MCP.

Why

Solo developers waste months building products nobody wants. SeedScope answers "Should I build this, and if so, what exactly?" by running a multi-stage evaluation pipeline that combines LLM reasoning with quantitative simulation.

Pipeline

Web signals → Extract ideas → Screen & dedup → Bundle (market reference)
    → 5-persona debate + self-consistency EV scoring
    → ABM Monte Carlo market simulation (200 runs × 24 months)
    → Structured deliberation → MVP spec ready for Vibe Coding
    → Telemetry grounding (re-evaluate with real DL/revenue data)

Bundled Package Showcase

SeedScope demonstrates what you can build with algocline's bundled packages alone:

Bundled Package	Used In	What It Does
panel	evaluator	5-persona adversarial debate (pragmatist / skeptic / builder / customer / contrarian)
deliberate	designer	6-phase structured decision: abstract → consult → generate → debate → confidence → rank
contrastive	simulator	Incumbent absorption analysis with competitive landscape comparison
calibrate	simulator, evaluator	Probability calibration for risk confidence scoring and telemetry grounding
abm	simulator	Agent-based Monte Carlo market simulation with Agent.define + Model + mc.run_model
SwarmFrame	seed_scope_orch	Step-based pipeline orchestration with flow state persistence and resume

Packages (8 packages, ~1200 lines total)

Package	Role	Key Feature
seed_scope_orch	Orchestrator	4 modes: ingest / evaluate / re_evaluate / status
sourcer	Sourcing	Thompson Sampling theme arms + LLM idea extraction
screener	Filter	Fingerprint dedup → keyword rules → LLM batch screen
evaluator	Scoring	Panel debate + 5-lens self-consistency + EV formula + telemetry re-evaluation
simulator	Simulation	contrastive incumbent analysis + abm.mc.run_model (survival rate, Wilson CI, equilibrium)
designer	Design	Panel weakness analysis → feature classification → deliberate decision → MVP spec
portfolio	State	Eval history with EMA smoothing, Card emission, fingerprint dedup
bandit	Math	Thompson Sampling with Beta-Binomial (budget allocation + kill candidates)

Quick Start

# Install algocline
cargo install algocline

# Link the whole collection at once (collection mode — one call links every package)
alc pkg link packages

# Evaluate a single idea
alc run '
local orch = require("seed_scope_orch")
return orch.run({
    mode = "evaluate",
    task = "A browser extension that detects dark patterns on checkout pages and shows the real price",
    namespace = "demo",
})
'

Example: Evaluate Mode Output

-- Input
{ mode = "evaluate", task = "Slack bot that auto-summarizes threads > 10 messages" }

-- Result (abbreviated)
{
    decision = "SCAFFOLD",
    expected_value = 18.742,
    metrics = {
        Pain_level = 8, Willingness_to_Pay = 6, Purchase_Realism = 5,
        Defensibility = 4, Self_Serve_Fit = 8, Time_to_MVP = 14,
    },
    debate_synthesis = "Strong pain signal in remote teams. Slack API limits are manageable...",
    simulation = {
        survived_rate = 0.67,
        survived_ci = { lower = 0.58, upper = 0.75 },
        final_revenue_median = 2100,
        equilibrium = "contested",
    },
}

Example: Re-evaluate with Telemetry

After launching, feed real numbers back to ground the LLM evaluation:

alc run '
local orch = require("seed_scope_orch")
return orch.run({
    mode = "re_evaluate",
    task = "Slack bot that auto-summarizes threads > 10 messages",
    namespace = "demo",
    telemetry = {
        downloads = 340,
        revenue_usd = 1200,
        active_users = 180,
        churn_rate_pct = 8,
        months_live = 3,
    },
})
'
-- Result: EV 18.742 → 22.105 (+3.363), decision=SCAFFOLD, confidence=0.82

Example: Full Ingest Pipeline

alc run '
local orch = require("seed_scope_orch")
return orch.run({
    mode = "ingest",
    namespace = "hn_weekly",
    raw_posts = {
        { title = "Ask HN: What tools do you wish existed?", body = "...", source = "hn" },
        { title = "I spent 6 months building X and nobody uses it", body = "...", source = "hn" },
    },
})
'
-- 7-step pipeline: source → screen → bundle → evaluate → simulate → design → record
-- Produces: screened ideas, EV scores, market simulations, MVP specs, portfolio Cards

Specs

32 tests across 5 packages, all pure computation (no LLM calls in tests):

# Run all specs
for spec in packages/*/spec/*_spec.lua; do
    alc pkg test --code-file "$spec" --search-paths packages/
done

Claude Code Agents (plugin)