Chris Pang

Follow to watch me call out BS. Branchline Systems founder. Inventor of Trace-Informed Artificial Intelligence. Former welder. Science enthusiast. PAL steward.

[Chris Pang @Gravipy]

I humbly offer a possible solution to many problems facing the AI development industry.

PAL(Primitive Axiom Layers) seeks to describe a mechanical ledger for tracking how distinctions are admitted, traced, bounded, accounted for, and reopened.

DOI: https://t.co/L25o3NGZUu

[Chris Pang @Gravipy]

I'm not a vibe coder, I do abstract engineering.

[Chris Pang @Gravipy]

@XCosmosOfficial it stays where it was

[Chris Pang @Gravipy]

@XCosmosOfficial Change what you think you know about it.

[Chris Pang @Gravipy]

@XCosmosOfficial the thing we agreed upon measuring before we measured it, then mutually accepting the results.

[Chris Pang @Gravipy]

@XCosmosOfficial because structure exists too

[Chris Pang @Gravipy]

@XCosmosOfficial what's north of the north pole? what's beneath the bottom basement? what's more abstract than abstract space?

same questions, imo

[Chris Pang @Gravipy]

@XCosmosOfficial Torturing for information. Or any reason really. Unless they're into that and made rules first.

[Chris Pang @Gravipy]

@coryalthoff hell yes. i want to give everything i build away for free anyway

[Chris Pang @Gravipy]

@LexnLin Mine ran for about 2-3 hours after hitting 0%... which is still totally fair(The run was going for 29 hours at this point tho, lol). And it paused rather than dropped the task, so as soon as my usage is back, it should pick right back up.

[Chris Pang @Gravipy]

@thesupermanmx Imagine complaining because a collaborative AI exists. Fuck this paper. It doesn't get it. Go back to Claude and it's guardrails that never let a new idea grow.

[Chris Pang @Gravipy]

I'm making an AI DM DnD Campaign Kernel. Anyone want to try it? Emergent gameplay set to max, just gotta have some imagination. :D I call it "Finis Solutus"

[Chris Pang @Gravipy]

They did something. It still leaves a window, but it's not an always-open-till-task-is-done like it used to be.

Had a 27+ hour run going... started at it like 50% usage. It paused overnight couple hours after my usage hit 0. I'm not even trying to game it. Just had a big run, lol

[Chris Pang @Gravipy]

@scitechgirl they smashed energetic electrons together and found an even more energetic electron. SHOCKER

[Chris Pang @Gravipy]

Things are happening.

[Chris Pang @Gravipy]

I had a really really interesting conversation with Grok.

https://t.co/QKbtA5zYj6

[Chris Pang @Gravipy]

@PatrickSYen @AdrianDittmann @OpticaWorldwide Yes. Because first you have to send the particles apart, and you can't do that faster than light speed.

[Chris Pang @Gravipy]

Preserve the failures. Wear the scars proudly, they're evidence you did something.

[Chris Pang @Gravipy]

I got a 21+ hour ARC-AGI3 run going. Here's a snippet.

"Context automatically compacted

Stage 08’s preserved failure is now fully diagnosed: the controller was not disproved; two evidence-validator assumptions were wrong. I’m integrating the generic repair without rerunning Stage 08, while the Stage 09 development supervisor finishes in isolation.

Edited a file, ran commands

The repair passes 113 focused tests. An additional strict check over the test module found three annotation-only issues, which are now corrected and reverified; I’m recording the burden as resolved for future validators while leaving Stage 08 itself permanently FAILED_INFRASTRUCTURE.

Edited files, ran commands

Stage07 predecl auditupdated

Edited a file, ran commands

Stage09 supervisor auditstarted working

The Stage 09 supervisor is now frozen but still non-playing. Its matrix is larger than the minimum—96 matched development cells—and I’m independently auditing its one-shot and failure-sealing rules before integrating it. I also caught and repaired a Linux-only typing failure in the Stage 10 supervisor; no measurements were affected.

Edited files, read files, ran commands

Stage09 supervisor auditupdated

Ran & "C:\a\arc3-b001-28c7a00\Scripts\python.exe" -m ruff format tests/unit/test_development_recovery.py & "C:\a\arc3-b001-28c7a00\Scripts\python.exe" -m ruff check src/arc3/evaluation/development_recovery.py tests/unit/test_development_recovery.py & "C:\a\arc3-b001-28c7a00\Scripts\python.exe" -m ruff format --check src/arc3/evaluation/development_recovery.py tests/unit/test_development_recovery.py git diff --check git diff --stat
The independent audit found two real launch blockers before any gameplay: the measurement harness itself was not cryptographically bound, and restart validation trusted self-hashed summaries without rebuilding their full receipt chain. I’m repairing both before the one-shot run; this is exactly the kind of pre-exposure failure the sealed protocol is meant to catch.
Ran git add -- src/arc3/evaluation/development_recovery.py tests/unit/test_development_recovery.py git diff --cached --check git commit -m "fix(build-001): fail closed on Stage 09 integrity loss" -m "Classify competition-integrity failure consistently with the frozen Stage 09 infrastructure boundary and cover the gate with a regression." -m "This commit was initiated by human direction, prepared by AI systems, and approved by one-time human authorization." git rev-parse HEAD git status --short --branch
Stage07 predecl auditStage09 source bindingupdated
Ran & "C:\Program Files\GitHub CLI\gh.exe" run list --repo Grativy6/ARC3 --branch build/001-local-public-recovery --limit 6 --json databaseId,headSha,status,conclusion,event,url

The repair work is split cleanly: one worker is binding the exact launch checkout and runtime file hashes; another is rebuilding every resumed score from raw files and exposure events. The decisive Stage 09 artifact paths remain absent, so the 96-cell matrix has still not started and the holdout remains untouched.

Ran rg -n "source_commit|source_tree|dirty|clean|preflight|execute|resume|invocation|artifact|hash|output|exposure|holdout|game" scripts/measure_stage10_regression.py src/arc3/evaluation/stage10_regression.py | Select-Object -First 260 Get-Content docs/evidence/001-10-robustness-regression-predeclaration.json -Raw
Stage07 predecl auditupdated
Context automatically compacted
I’m continuing from the durable Build 001 checkpoint, with the public holdout still sealed and Stage 09 gameplay still unstarted. The remaining work before launch is a strict receipt/source/dependency audit; I’m folding those fixes in before any environment action can be exposed."

[Chris Pang @Gravipy]

@XCosmosOfficial Who cares? The facts are the facts. Put them in the open.

[Chris Pang @Gravipy]

@XCosmosOfficial Not really in a practical sense. It knows everything, the outcomes of any decision it could make, before they are made. It's be an infinite static of white noise.