Consent-first, human-gated inference for detecting upcoming birthdays from authorized text imports.

LavenderButWithCakes2 is a safety-oriented inference system that ingests explicitly provided text (SMS exports, email snippets, chat logs) and produces reviewable candidates for time-sensitive actions — never autonomous actions.

No scraping. No guessing. No surprises.

Most “helpful” automation fails because it:

• overreaches
• infers intent without consent
• acts before a human can intervene

LavenderButWithCakes2 takes the opposite approach:

Inference is allowed.
Action is gated.
Humans stay in control.

• Consent-first
  Only text explicitly imported by a human is processed.

• Audit-first
  Every observation, inference, decision, and outcome is logged.

• Human-in-the-loop
  The system suggests candidates; humans decide.

• No surprise actions
  Privacy flags and “do not surprise” flags block candidate creation.

• Explainable logic
  Deterministic rules + Bayesian fusion. No opaque ML.

1. Ingest authorized text

   • Unified batch ingestion (/api/import_batch)
   • Import-only connectors (SMS, email, Slack, Teams, calendar)

2. Normalize + minimize

   • Strip inputs to only what’s needed
   • Store minimal observations

3. Extract evidence

   • Auditable pattern rules (e.g. “happy birthday”, explicit dates)
   • Evidence is evidence, not truth

4. Fuse probabilities

   • Conservative Bayesian log-odds fusion
   • Time-windowed posteriors (today / 3d / 14d)

5. Generate candidates

   • Only if thresholds + confidence are met
   • Blocked by privacy or “do not surprise”

6. Human gate

   • Approve / defer / reject
   • Required checks surfaced (address, allergies, preferences)

7. Log outcomes

   • Delivered / not delivered
   • Delight score (-1 / 0 / +1)
   • Used for future calibration

• ❌ Scrape data
• ❌ Read messages without explicit import
• ❌ Autonomously message anyone
• ❌ Infer sensitive attributes
• ❌ Act without human approval

This is an inference assistant, not an autonomous agent.

app.py → thin Flask boot + wiring config.py → thresholds, priors, global policy db.py → SQLite schema + connection audit.py → append-only audit log models/ → all DB writes live here people.py observations.py evidence.py candidates.py outcomes.py connectors.py inference/ → thinking only (pure logic) patterns.py extract.py bayes.py candidates.py ingestion/ → the single intake pipe normalize.py ingest.py connectors/ → import-only connectors base.py registry.py api/ → REST surface routes.py ui/ → lightweight human gate UI routes.py

Key rule: logic never lives in Flask routes.

GET /health

POST /api/import_batch

{
  "person_key": "mike_smith",
  "connector": "teams_import",
  "items": [
    {
      "occurred_at": "2025-12-21T18:20:00+00:00",
      "text": "Happy birthday Mike!",
      "meta": {
        "source": "teams",
        "channel": "Engineering",
        "author": "alice",
        "is_dm": false
      }
    }
  ],
  "default_meta": {
    "ingest": "example_client"
  }
}
Posteriors
GET /api/posteriors/<person_key>
Candidates
GET /api/candidates?status=PENDING
Outcomes
POST /api/outcome
________________________________________
UI (optional)
A lightweight UI exists at:
/ui
It supports:
•	Person management
•	Viewing posteriors
•	Human review of candidates
The UI is intentionally thin and replaceable.
________________________________________
Running locally
pip install -r requirements.txt
python app.py
The system uses SQLite by default and requires no external services.
________________________________________
Design philosophy
LavenderButWithCakes2 is built around a simple idea:
Automation should reduce cognitive load without removing agency.
This system is intentionally conservative, inspectable, and reversible.
It is designed for environments where trust, safety, and accountability matter more than speed or scale.
________________________________________
Status
This is a working reference implementation, not a growth-hacked product.
It is meant to be read, audited, extended, and adapted — not blindly deployed.
________________________________________
License
MIT