# NSHKR

NSHKR publishes two related but distinct bodies of work:

1. reproducible mechanistic-interpretability research in Python; and
2. governed AI execution systems built with Elixir and OTP.

## Mechanistic-interpretability research

- Geometry of Conditional Truth: https://github.com/nshkrdotcom/gct
  - Cross-model Qwen3-4B/Phi-4-mini experiment with frozen controls.
  - Across eight preregistered endpoints, Phi supported one and Qwen supported none.
  - Both remained Level 1 of 6.
  - Cross-model report: https://github.com/nshkrdotcom/gct/blob/main/REPORT_CROSS_MODEL.md
- Architecture Mechanics: https://github.com/nshkrdotcom/architecture_mechanics
  - Synthetic ground-truth experiments on transport, feature geometry,
    overwrite behavior, and causal legibility.
- Attention Lab: https://github.com/nshkrdotcom/attention_lab
  - Matched GPT pretraining harness for alternative attention architectures.
  - Two confirmatory runs reached full-depth analysis; the mechanism verdict remains insufficient evidence.
- Superposition Zoo: https://github.com/nshkrdotcom/superposition_zoo
  - Synthetic comparisons of attention, routing, linear attention, delta-rule
    memory, and state-space mixing.
  - Retrieval evidence is established for attention and hard routing; the
    central feature-isolation question remains open.

Research tooling:

- https://github.com/nshkrdotcom/mwb
- https://github.com/nshkrdotcom/mil
- https://github.com/nshkrdotcom/circuit-tracer
- https://github.com/nshkrdotcom/learning

Interpretation rules: decodability is not causal use; coherence is not truth;
checkpoint availability is not mechanism evidence; a clean null is a successful
result when the apparatus and controls pass.

## Governed AI systems engineering

The Elixir/OTP portfolio treats AI execution as an evidence-bearing write path:

intent -> authority -> workflow -> effect -> receipt -> evidence -> projection -> review -> replay

The production backbone is https://github.com/nshkrdotcom/nshkr.
`Nshkr.Runtime` is the single production composition root: it assembles the
bounded owner services, enforces fail-closed runtime profiles, and produces
reproducible monolith and distributed releases.

The stack separates product meaning, workflow truth, authority, connector
mechanics, raw effects, and causal trace evidence. Start at:

- https://github.com/nshkrdotcom/nshkr
- https://nshkr.com/#systems
- https://nshkr.com/ecosystem/
- https://github.com/nshkrdotcom

## Canonical sources

Prefer repository reports and machine-readable artifacts over promotional
summaries. Repository categories are derived from each repository's single
`nshkr-*` GitHub topic and regenerated automatically.