A personal notebook on things built & questions still open

Jay Building things. Following questions.

I'm Jay. I spent years doing HVAC and construction work, where you learn fast that a system lying to you about its own state gets someone hurt — so when I moved into software, I brought that with me. Everything I build comes down to one belief: intelligence may be probabilistic, but authority should be explicit. Right now that's three doors: Investigate (Black Lantern), Decide (SHAD0W), and Orchestrate (Automated Social).

Explore with me Email Jay ↗ Engineering archive ↓

Perspective

From the field to software.

I also spend time with Scripture, philosophy, and the wider world. This is a place for the things I make and the questions I keep returning to.

Model behavior, AI agents, and inference infrastructure are current interests. So are Scripture, the humanities, philosophy, astronomy, and emerging technology. Some interests become software; others stay open questions.

Jay’s hand-drawn Sinai scene: two tablets on a stone platform, a calf below, flames at the sides and beams of light.

Original drawing · Jay

Sinai & the golden calf

A personal visual exploration of the tablets of Mount Sinai and the golden calf, and their symbolic duality.

View the original drawing (PNG, 1.6 MB)

Current explorations

What has my attention.

BibleMap · A reading experience

JonahThe Man Who Ran

Four chapters. Room to linger.

Marble sculpture of a bearded man reclining beneath a curling vine with broad leaves.
Jonah Under the Gourd Vine 280–90 CE. Anatolia, late Roman–early Christian. A later artistic interpretation of Jonah 4, not a portrait or reconstruction of the story.

These stories echo through the doorways of time. It is vital to explore them, and explore them deeply.

Jay · A personal reflection ·

Read all four chapters of Jonah, with Scripture at the center and optional historical context and philosophical or theological perspectives alongside it. Move through the story at your own pace.

Enter the Jonah experience

Scripture · Geography · Humanities

Read Jay’s full Jonah reflection

Jay’s personal interpretation, in his original words. Historical context is sourced separately in the Jonah experience.

I'm developing Biblemap as I explore my own spirituality, the human story through a Biblical lens. Because if nothing else, these are historical documents with stories and lessons we see repeated throughout our human history- us simple creatures that we are. Jonah captures a beautiful human element- he wants to hide from an Eternal being, God. Jonah hops a ship to run and tries to hide; and the unthinkable happens. The way Jonah presents himself through the work of God aligned with there pagan beliefs at the time, and thus Ninevah changed there ways.

These stories echo through the doorways of time. It is vital to explore them, and explore them deeply.

Artwork: Cleveland Museum of Art, 1965.239. John L. Severance Fund. Public domain / CC0.

Continue to Curiosity

Questions, instruments & observations

Curiosity

My deepest curiosity on this exponential trend of intelligence is we do not need a chat interface. We need observability and an orchestrator of intelligence. What does this paradigm shift mean for humanity and how will it affect us all?

Jay · A personal thesis & open question ·

TEMPER below is one limited exploration of model decisions. It does not test this broader thesis about intelligence and humanity.

Model behavior · TEMPER

When should a model decline to answer?

A confident answer can still be wrong. I’m interested in what happens when a system can abstain or pass a decision on.

The inspected TEMPER source measures coverage, errors among accepted predictions, and calibration. Those are tools for asking the question; they don’t establish that a model knows when to stop.

  • How many decisions remain when the confidence threshold rises?
  • How often are the accepted decisions wrong?
  • Does reported confidence match observed correctness?

Source inspected at 0747730. Tests and experiments not rerun here; no cost or reliability advantage established.

An observatory instrument · Synthetic test inputs

Confidence is a cutoff, not a guarantee.

A 0.95

Correct

Prediction 1 · Reference 1

Accepted

B 0.90

Incorrect

Prediction 0 · Reference 1

Accepted

C 0.60

Correct

Prediction 1 · Reference 1

Deferred

D 0.40

Correct

Prediction 0 · Reference 0

Deferred

Coverage accepted / all decisions
50%
Selective risk errors / accepted decisions
50%

At threshold 0.80, 2 of 4 decisions are accepted; 1 accepted decision is incorrect.

Four synthetic inputs from TEMPER’s selective-risk and risk/coverage unit tests. This illustrates the metric definitions, not model performance. At 0.80, the incorrect 0.90-confidence decision stays while two correct decisions are deferred. Raising the threshold does not guarantee lower risk. With none accepted, risk is undefined.

A note from field work

Check what reached the component.

A stopped unit and a missing signal can look like the same failure from the outside. I want to know what the component received, what it did, and what the next person was told. I ask the same questions at a software boundary.

How I think

Working principles

A few habits I bring to technical work. Current questions ↓

  1. Make authority explicit.

    Separate what a system proposes from what it is allowed to do.

  2. Keep the evidence.

    Decisions and outcomes should remain understandable outside the process that produced them.

  3. Learn from the failure.

    Leave enough context to understand what happened and decide what comes next.

  4. Check the handoff.

    Make consequential exchanges explicit without unnecessarily constraining intelligence inside the boundary.

Three doors into the work

Follow the mechanism.

Investigate a boundary. Decide when to act. Coordinate without surrendering authority.

01 / Investigate

What actually enforces an agent’s boundaries?

Black Lantern

Harness-bound identity reaches a final authorization check; a separate observer records the synthetic resource effect.

Published synthetic tests expose an unchecked-route blind spot. They do not establish real-world containment.

Inspect the boundary

02 / Decide

When is an action eligible to happen?

SHAD0W

A chronological timeline admits opportunities strictly after signal availability. Explicit lifecycle state and a separate paper-risk gate govern what follows.

Pinned source and regressions inspected; no profitability, live trading or broker execution is established.

Read the decision study

03 / Orchestrate

Can intelligence operate without publishing authority?

Automated Social

Ollama drafts locally. SQLite binds review to an exact revision and payload hash; Telegram carries the human decision.

Delivery, approval and publishing remain separate. X publishing is disabled; mocked tests do not prove production reliability.

Follow the review handoff