Loose, unopinionated notes on things I don't entirely understand yet.
Lower commitment than essays. More permanent than a social media post. Thinking out loud, in public, slowly.
Can Bayesian forecasting beat Goldman Sachs during an oil crisis? 8-week experiment. Methodology public. Results documented. Ego at risk.
I built a script to track my own Threads suppression. Then a Meta insider confirmed the pattern. Four-part investigation, documented as it unfolds. Starting with the data.
Building Rainforest Studio's Living Archive. Specifics over theory. Starting with the two data model rewrites.
Pick one and come along.
The suppression tracker is on GitHub. MIT licensed, zero dependencies. You can check your own data now. Meanwhile, I've been writing here more and posting there less.
Brent crude jumped 7% after Trump ordered a naval blockade of the Strait of Hormuz. The pause lasted five days. The talks failed. Now the model has to decide if a CENTCOM order is different from a Truth Social post.
$28.33 off. The Iran ceasefire broke everything. Brent fell 13% in a week. The model called $123.53, the market closed at $95.20. Four weeks, four losses to the naive baseline.
I tried to use Living Archive to capture real-time crisis documentation during the Iran war. It didn't fit. That's when I realised the Living Archive isn't a product. It's a platform.
ElevenLabs puts an American flag next to 'English'. My transcription tool inserts Oxford commas I didn't say. Two platforms, same assumption. I'm keeping a list.
It was 23:48 on Tuesday. My forecast was locked at $123.53. Twelve minutes earlier, Trump agreed to suspend the bombing. I watched the oil forecast decay in real time.
A crypto Patreon cancels after five years. TikTok buries stand-up comedy. Cost of living reshapes what people will watch. AI eats every feed. These aren't separate stories.
Brent crude forecast for 10 April 2026. Trump posted 'Praise be to Allah' on Easter Sunday and threatened Iranian power plants by Tuesday. Physical crude hit $141. The model is at $123.53.
$9.83 off. Direction wrong for the first time. The model predicted up, Brent went down. Three weeks, three losses to baseline. But the physical market is telling a different story.
Rainforest Studio's website looked like every other tech startup. White backgrounds, clean sans-serif, tasteful green. The content was alive and the design was dead. Time to fix that.
Brent crude forecast for 2 April 2026, $118.88. Houthi strikes reopened, the price band broke, and Good Friday cuts the trading week short. The model's most aggressive call yet.
$1.32 off. The model nailed the price but the naive baseline still won. Week 2 of an 8-week oil price forecasting experiment during the Iran war, scored honestly.
I posted about Threads suppression on Threads itself. The post got 0 views and wouldn't load. The platform proved the thesis while I was writing about it.
Brent crude forecast for 27 March 2026, $108.16. The model corrected from Week 1's overshoot. But Trump's Strait of Hormuz ultimatum could break the entire framework.
My Brent crude forecast overshot by $13.75. The naive baseline won. Week 1 of an 8-week oil price forecasting experiment during the Iran war, scored honestly.
I posted my Threads suppression data on LinkedIn. A Trust & Safety Specialist at Meta commented saying it highlights what people inside already suspected.
First Brent crude forecast locked at $99.47. Escalation probability 45%. Eight weeks of public oil price predictions during the Iran war starts now.
Starting an 8-week experiment. Can systematic Bayesian updating beat Goldman Sachs' oil forecasts during a geopolitical crisis? Week 1 methodology and baseline.
The full methodology behind an 8-week oil price forecasting experiment. Five scenarios, Bayesian updating, public scoring, and why I'm doing this during a war instead of doom scrolling.
My Threads engagement dropped 30x overnight. I built a script to pull 30 days of data and found evidence of algorithmic suppression. Here's what the numbers showed.
Week 1 of building Living Archive, a RAG tool for community oral histories. The data model changed twice. The ingestion pipeline works. The chunking doesn't. Here's what happened.