Business Report · 2026-04-17

Evolvable Signal Engine — V1 Shipped

A feedback-driven scoring system that replaces rigid, set-and-forget weights with evidence-based signal lifecycle management. Built end-to-end today for the Trades $15M+ revenue-prediction project.

44 signals registered33 assigned to trades-15m8 clusters$3.1420 per company$314.20 per 100 batch

Executive Summary

Prior scoring systems had a fatal flaw: weights were hardcoded in Python, locked on a date, and changed only by rewriting code. Signals never earned their place — they inherited it. There was no record of which signals actually predicted outcomes and no way to add or retire a signal without a rewrite.

V1 installs the foundation that makes scoring evidence-driven and self-evolving. Every signal is a row in a registry. Every captured value is appended to an observation log. Every ground-truth event (calls, meetings, deal outcomes) is appended to an outcome log. A review job joins the two and computes predictiveness. Weights change because data says so, not because anyone opined.

All 33 signals in the first list generator (trades-15m) enter as probationaryat weight 0. They earn weight by predicting outcomes; they lose weight — or get retired — when they don't.

What Shipped — Three Deliverables

Deliverable 1

Signal list of known signals

A Supabase-backed registry of every signal we measure. Each row: description, data type, status (probationary/active/deprecated/retired), cost per observation, fallback rule, source vendor.

44 signals · 44 probationary

Deliverable 2

Signal engine

Python module (backend/lib/signal_engine.py) with 5 primitives: select_signals, record_observation, record_outcome, score_target, review_performance. CLI + end-to-end smoke-tested.

5 primitives · 1 CLI · psycopg2 direct

Deliverable 3

Review webpage

This site. Server-rendered from Supabase — projects, signals, costs, predictiveness all live. Zero redeploy to publish a new signal.

3 routes · Next.js 16 · Tailwind

Architecture — Five Primitives

Signal Registry

signal_registry

Rows, not code. Each signal: status, current_weight, data_type, source_vendor, fallback_rule, cost_per_observation_usd.

Signal Observations

signal_observations

Append-only log of every captured value per target. Polymorphic (any entity type). Records weight_at_time and scoring_event_id so history is reconstructable.

Outcome Observations

outcome_observations

Append-only log of ground truth: call_connected, meeting_booked, deal_closed_won, revenue_truth, etc. The other half of the feedback loop.

Signal-Outcome Pairs

v signal_outcome_pairs

View joining signals to subsequent outcomes for the same target. Feeds the nightly review job; also ad-hoc queries for 'does X predict Y?'.

Project Assignments

project_signals

Links signals to list-generator projects with project-specific weights. Same signal can serve many projects with different weightings.

Cost View

v v_project_signal_cost

Joins project + signal + cost math for the review webpage. Cost per observation × batch size = cost per run.

Trades-15M Project — Live KPIs

Signals assigned

Cost per company

$3.1420

Cost per batch (100)

$314.20

Clusters

Cost breakdown by cluster

Cluster	Signals	$ / company	$ / batch (100)	% of total
Added by Claude	7	$0.8700	$87.00	27.7%
Social Media	6	$0.7600	$76.00	24.2%
Industry Involvement	1	$0.4000	$40.00	12.7%
Website Wealth	4	$0.3200	$32.00	10.2%
Business Maturity	5	$0.3050	$30.50	9.7%
Paid Marketing	3	$0.2700	$27.00	8.6%
Regulatory	4	$0.1500	$15.00	4.8%
Reviews	3	$0.0670	$6.70	2.1%
Total	33	$3.1420	$314.20	100%

Top-5 highest-cost signals

enrich.industry-involvement-score.composite
Composite 0-100 score based on presence of: conference sponsorship (0-12pts), chamber of commerce involvement (0-12), guest speaker appearances (0-12), chapter president history (0-12), BNI membership (0-12), industry awards received (0-13), industry orgs listed on own website (0-13), paid sponsorships of any kind (0-14)
$0.4000
enrich.instagram-project-posts.social
Count of posts in trailing 12 months classified by vision model as showing finished/completed work
$0.3500
enrich.press-mentions-12mo.exa
Count of news/press mentions in trailing 12 months
$0.3000
enrich.monthly-ad-spend-estimate.sem
Estimated monthly paid search spend in USD
$0.2500
enrich.tech-stack-indicators.builtwith
Count of enterprise-tier technologies detected on website (HubSpot Enterprise, Salesforce, Shopify Plus, etc.)
$0.2000

Who authored each signal?

Every signal is tagged with its author. The registry has three: Ewing (pre-existing Hermes/CRM signals), Mark (named today 2026-04-17), and Claude (proposed independently). Authorship is stored on signal_registry.added_by and shown on every view.

Ewing

8 registry · 0 in trades-15m

Pre-existing signals inherited from Hermes succession scoring and CRM buyer-fit research.

Mark

29 registry · 26 in trades-15m

Signals Mark named today across 7 clusters (Social Media, Website Wealth, Reviews, Paid Marketing, Business Maturity, Regulatory, Industry Involvement).

Claude

7 registry · 7 in trades-15m

Bias-breakers Claude proposed: tech stack, job postings, press mentions, customer logos, marketing staff, financing options, fleet visibility.

Cost reality — tiered funnel, not flat scan

The $3.14/company number is the naive upper bound— running every signal on every company. That's never the real cost. Real cost uses a tiered funnel: cheap signals filter the population, then only survivors get expensive ones.

Tier 1 — disqualify first

≤ $0.05 per obs

$0.3620

18 signals

Run on full population. Cheap enough that false positives don't matter. Includes: whois, google-rating, team-page, BBB, yelp, ad-transparency.

Tier 2 — score survivors

$0.05–$0.15

$1.0800

9 signals

Run only on targets that pass tier-1 filters. Includes: instagram activity, facebook, linkedin employee count, state license, fleet street view.

Tier 3 — deep dive

> $0.15

$1.7000

6 signals

Run only on top-ranked survivors. Includes: IG project vision, SEMrush ad spend, BuiltWith, press mentions, industry-involvement composite.

Example: scoring 10,000 companies

Tier 1 on full population (10,000 × $0.3620)	$3620
Tier 2 on survivors (1,000 × $1.0800)	$1080
Tier 3 on top ranked (200 × $1.7000)	$340
Tiered total	$5040
Naive (flat scan)	$31420
Saved vs. naive	$26380 (84%)

Assumes 10% pass Tier 1 and 20% of those pass Tier 2. Actual rates come from real outcome data after the first few runs.

Signal Library — 3 Final Lists

✓ USED

from Mark's 26

25 signals · 7 clusters

• Social Media (6)
• Website Wealth (4)
• Reviews & Reputation (3)
• Paid Marketing (3) inc. revenue-floor
• Business Maturity (5)
• Regulatory (4)
• Industry Involvement (1 composite)

+ ADDED

by Claude

7 signals · bias-breakers

• Tech stack (BuiltWith)
• Active job postings (Indeed)
• Press mentions 12-mo (Exa)
• Named customer logos (site+vision)
• Marketing staff on LinkedIn
• Financing options offered
• Fleet visible count (Street View)

⏸ / ✗ NOT USED

deferred + rejected

3 items · with reasons

⏸ Permits pulled (county). High value, but 50-state coverage is multi-quarter. Pilot 1 county first.
⏸ Equipment/truck photos. Vision cost + inconsistent sources. Revisit after M7 (fleet branding) proves worth.
✗ Generic gov-website mentions. Signal-to-noise too low; subsumed by state license + permits.

The highest-value signal in the library is score.revenue-floor-from-adspend.derived — the only causal signal (ad spend × 12 ÷ 0.25 margin = revenue floor). Everything else is correlational.

How signals evolve — the lifecycle

Step 1

Probationary

Enters at weight 0. No contribution to score until enough outcome data.

Step 2

Active

Earned weight from outcomes. Point-biserial r > threshold across ≥10 pairs.

Step 3

Deprecated

Lost predictiveness or superseded. Weight drops but observations keep flowing for comparison.

Step 4

Retired

Permanently off. Historical observations preserved; no new ones captured.

Signal Engine — CLI

# list all list-generator projects
python -m backend.lib.signal_engine list-projects

# list the signals assigned to a project w/ cost + weight
python -m backend.lib.signal_engine list-signals trades-15m

# run the nightly performance review (point-biserial r per signal)
python -m backend.lib.signal_engine review --window-days 30

# end-to-end smoke test (score demo target + outcome + review)
python -m backend.lib.signal_engine demo

Review pages

/projects

Index of all list generators with totals

/projects/[key]

Full signal table, costs, weights, predictiveness

/signals

Cross-project registry view

What's next (ordered)

1
Wire Hermes score_and_rank.py to record_observation()
Every scoring run starts writing append-only rows. Turns on the feedback-loop input side.
2
Build gatherers — cheapest first
whois, google-places, bbb-scrape, domain-age. Ship the free and near-free signals first so we generate volume for the review job.
3
Deploy nightly review via scheduled-tasks MCP
review_performance runs nightly; results flow into signal_registry; morning digest surfaces candidates for status transitions.
4
First weight assignment (not opinion-based)
After 2–4 weeks of data, move signals with r > 0.2 from probationary to active and assign a real weight. First time weights change from zero.

Code & commit references

backend/data/20-SIGNAL-REGISTRY.sql — registry + observations + outcomes + pairs view

backend/data/21-SIGNALS-TRADES-15M.sql — 34 INSERT statements for trades verticals

backend/data/22-SCORING-PROJECTS.sql — cost columns + projects + project_signals + cost view

backend/data/23-TRADES-15M-PROJECT-SEED.sql — cost estimates + project + 33 assignments

backend/lib/signal_engine.py — the 5 primitives + CLI

next-chapter-os/src/app/{projects,signals,report}/* — this site

Shared Supabase: dwrnfpjcvydhmhnvyzov · Generated live on each page load.