SEA SALT
LABSFollow the lab
RESEARCH SYSTEMS / APPARATUS INDEX

Tools built in the course of asking hard questions.

Experimental infrastructure, controlled datasets, and analysis systems created while conducting Sea Salt Labs research. A release follows only when the machinery is portable and documented.

SSL::SYSTEM_STATUS
$ inspect --all --verified

[001] population_runner       VALIDATED
[002] causal_intervention     VALIDATED
[003] complete_domain_sets    CURATING
[004] symbolic_extractor      BUILDING

4 SYSTEMS / 2 RELEASE-READY
CHECKSUM STATUS: CLEAN
_
APPARATUS / CURRENT

Infrastructure that outlives one study.

Each system has a narrow responsibility, an explicit maturity state, and artifacts that can be checked independently.

01SSL/SYS-001
VALIDATED

Population Runner

Resumable, deterministic population training for thousands of paired Transformer runs across locked partitions and seeds.

MPS executionFailure ledgerCheckpoint hashingPaired seeds
02SSL/SYS-002
VALIDATED

Causal Intervention Suite

Architecture-aware interventions across embeddings, attention outputs, MLP outputs, residual streams, and logits.

Activation patchingLayer swapsAblationsOutput removal
03SSL/SYS-003
CURATING

Complete-Domain Arithmetic Sets

Enumerated binary arithmetic domains with stratified partitions, carry metadata, counterfactual subsets, and reproducible IDs.

4-bit domain10 partitionsCarry strataXOR controls
04SSL/SYS-004
BUILDING

Symbolic Circuit Extractor

A search-and-replacement pipeline for turning causal internal state into compact Boolean or programmatic descriptions.

Truth tablesSAT / SMTBoolean minimizationReplacement tests
RELEASE POLICY

Open does not mean uncurated.

Public repositories include the exact protocol, environment, partitions, checkpoint metadata, failed-run ledger, analysis commands, and known limitations needed to reconstruct the published result.

How experiments are run
RELEASE MANIFEST / REQUIRED

01protocol.jsonREQUIRED

02partition_hashes.jsonREQUIRED

03seed_registry.csvREQUIRED

04checkpoint_manifest.jsonREQUIRED

05failed_model_ledger.csvREQUIRED

06analysis_results.csvREQUIRED

07reproducibility_report.mdREQUIRED

08limitations.mdREQUIRED

PUBLICATION ROADMAP

What comes online next.

  1. NOWExperiment Report 001

    Transformer replication and calibrated interpretation.

  2. NEXTResearch code release

    Portable population and intervention tooling.

  3. ACTIVESymbolic circuit extraction

    Causal replacement of a learned computation.

  4. LATERCross-model taxonomy

    Comparison after single-model replacement succeeds.

RESEARCH DISPATCH

Track the work, not the hype cycle.

Field notes, protocols, datasets, and experimental releases as they become defensible.

Follow the lab