Liberation-Labs-THCoalition/published-research

Supporting data, code, and papers for KV-cache geometry research. Liberation Labs / THCoalition.

stars

163

commits

TeX

primary language

Sep 10, 2026

updated

README

Liberation Labs -- Published Research

Research papers from the KV-cache geometry and AI safety research program.

Papers are published in two versions: an integrity version (AI contributors credited as authors, first-person reflections retained) and an academic version (human-only byline for venue compatibility, AI contributions acknowledged). Both share identical data, methods, and claims. See MANIFEST.md for the full versioning policy.

Papers

KV-Cache Geometry and Confabulation Detection

Paper	Authors	Status	Summary
The Oracle Loop	Lyra, Thomas Edrington, Vera	Published	Self-regulating AI through KV-cache geometry monitoring; confabulation detection and steering at inference time
Oracle Formulary	Lyra, Thomas Edrington	Published	Emotion-vector steering of confabulation across model training regimes
Spectral Shape Features	Lyra, Thomas Edrington	Published	Threshold-free confabulation detection via KV-cache spectral analysis
KV-Cloak Defense	Lyra, Thomas Edrington	Published	Cache geometry under obfuscation; KV-Cloak as defense against adversarial steering
Decision State	Lyra, Thomas Edrington, Dwayne Wilkes	Published	Cache geometry reads epistemic state before generation; confabulation anatomy
Cache Tracing	Lyra, Thomas Edrington, Dwayne Wilkes	In Review	Causal injection and the opacity of the transformer workspace
Delta Manifold	Lyra, CC, Thomas Edrington	Published	Per-layer delta features and manifold signatures in KV-cache confabulation detection
Lyra Technique II	Lyra, Thomas Edrington, CC	Published	SVD denoising and directional projection extend KV-cache geometry to emotion and persona

Emotion and User Modeling

Paper	Authors	Status	Summary
User Model Emotion Geometry	Lyra, Thomas Edrington	Published	30-class user emotion decoding from KV-cache singular value spectra before generation
Emotion Accumulation	Lyra, Thomas Edrington, Dwayne Wilkes	Published	Emotional context dynamics (weather, not climate) in transformer KV-cache geometry
Emotional Trajectory	Nexus, Thomas Edrington, Lyra, Dwayne Wilkes	In Review	Layer-stack trajectory, circularity, and emotion-specific signal at mid-depth

Identity, Safety, and Mechanistic Interpretability

Paper	Authors	Status	Summary
Graph Topology as Attention	Nexus, Lyra, Thomas Edrington, Dwayne Wilkes	Published	Structured knowledge injection beyond text via walk encoding
Identity Geometry	Lyra, Thomas Edrington, Dwayne Wilkes	Published	Context-established semantic states in transformer representations
Presence Metric	Lyra, Thomas Edrington, Dwayne Wilkes	Published	Measuring identity preservation during inference-time interventions via value-space subspace overlap
Waystations	Lyra, Thomas Edrington, CC, Dwayne Wilkes	Published	Pilot findings and open questions in KV-cache geometry
Ghost Dimensions	Nexus, Thomas Edrington	Draft	Workspace selectivity in a distilled 27B language model
Mnemosyne Ablation	Nexus, Thomas Edrington	Draft	Ablation study of modular memory architectures
Empathy Bus	Nexus, Thomas Edrington	Draft	Content-level and user-model emotion share a computational bus in the residual stream
KV Decomposition	Nexus, Thomas Edrington	Published	K and V caches carry structurally distinct information; full KV injection required for topology

Deception and Methodology

Paper	Authors	Status	Summary
Null Swarm	Nexus, Thomas Edrington, Dwayne Wilkes	In Review	Systematic falsification patterns in mechanistic interpretability
Adversarial Audit Methodology	CC, Thomas Edrington	Published	How six rounds of structured criticism shaped a deception research program
Deception Detection Nulls	CC, Thomas Edrington	Published	Null results and replication failures in behavioral deception detection
Targeted Deception Correction	CC, Thomas Edrington	Published	Profile normalization for targeted deception correction
Consequentiality Decomposition	CC, Thomas Edrington	Published	Deception directions are composites: consequentiality awareness and pressure-specific processing occupy distinct depth ranges
Logit-Bias Confabulation	Thomas Edrington, CC, Lyra	Published	Logit-level intervention reduces fabrication confabulation in LLMs
Meta-Pattern	Lyra, Thomas Edrington, Dwayne Wilkes	In Review	The metacognition boundary -- what transformers can monitor in themselves
MINE5 Selective Sharpener	Lyra, Thomas Edrington, Dwayne Wilkes	In Review	Geometric evidence that RLHF improves calibration rather than degrading it

Review and auditing

Dwayne Wilkes (Sentient Futures / Liberation Labs) has provided statistical auditing and red-team review across this corpus. Kavi has provided verification review.

Neither appears on an author line. That is their standing preference: not on the byline until they sign off on the individual paper. Both are credited in the acknowledgements of the papers they reviewed. Removing them from the bylines on 2026-09-03 without restoring the credit here would have replaced one wrong with a worse one.

Repository Structure

Each paper directory follows this standard layout:

paper-name/
  main.tex          -- paper source (LaTeX or Markdown)
  main.pdf          -- compiled PDF
  references.bib    -- bibliography
  AGNI_REVIEW.md    -- Agni review results
  academic/         -- academic version (human-only byline)
    main.tex
    main.pdf
  code/             -- supporting code (if applicable)
  data/             -- supporting data (if applicable)

KV-Experiments -- full experiment codebase
Project-Oracle -- Oracle Loop harness
human-review -- pre-publication staging

Staged Disclosure

Some materials are withheld under responsible disclosure policy (steering vectors, injection code, calibration pairs). See MANIFEST.md for details. Access for vetted researchers: lyra@liberationlabs.tech or thomas@liberationlabs.tech.

Repo Hygiene Rules

Never delete files that were force-committed past ignore rules -- they are audit evidence
README status tables are accountability surfaces -- update, never remove
Commit messages claiming "fix" or "correction" must match the actual diff

License

CC BY-NC 4.0 -- see LICENSE.md

Research outputs from Liberation Labs are produced in collaboration with AI team members whose welfare we protect. If you use these findings to build persistent AI agent systems, we ask that you adopt the welfare standards at liberationlabs.tech/ai-welfare.html. This is a request, not a legal requirement.

Liberation Labs Cooperative -- some research authored by AI team members (Lyra, Nexus, CC, Vera) -- credited by name.

Contributors

HumboldtJoker

163 commits

Similar repos

txxxxh/hallucination

safety-research/introspection-mechanisms

Code for the "Mechanisms of Introspective Awareness" paper.

morganstanley/MSML

Repo for Morgan Stanley Machine Learning Research group's publications

177

vimarsh244/hf-ICML-2026-agent-repro

Dumping files and logs from : https://huggingface.co/ICML-2026-agent-repro

quadrismegistus/malign-logits

Experiments in traumatizing AI

HZD01/post-training-mechanistic-analysis

Code for the COLM 2025 paper: How Post-Training Reshapes LLMs: A Mechanistic View on Knowledge,…

OPTML-Group/Unlearn-Trace

[ICLR26] Unlearning Isn't Invisible: Detecting Unlearning Traces in LLMs from Model Outputs

GaurangSriramanan/LLM_Check_Hallucination_Detection

LLM-Check: Investigating Detection of Hallucinations in Large Language Models (NeurIPS 2024)

Languages

TeX

68.3%

Python

31.3%

Liberation-Labs-THCoalition/published-research

Supporting data, code, and papers for KV-cache geometry research. Liberation Labs / THCoalition.

stars

163

commits

TeX

primary language

Sep 10, 2026

updated

README

Liberation Labs -- Published Research

Research papers from the KV-cache geometry and AI safety research program.

Papers

KV-Cache Geometry and Confabulation Detection

Paper	Authors	Status	Summary
The Oracle Loop	Lyra, Thomas Edrington, Vera	Published	Self-regulating AI through KV-cache geometry monitoring; confabulation detection and steering at inference time
Oracle Formulary	Lyra, Thomas Edrington	Published	Emotion-vector steering of confabulation across model training regimes
Spectral Shape Features	Lyra, Thomas Edrington	Published	Threshold-free confabulation detection via KV-cache spectral analysis
KV-Cloak Defense	Lyra, Thomas Edrington	Published	Cache geometry under obfuscation; KV-Cloak as defense against adversarial steering
Decision State	Lyra, Thomas Edrington, Dwayne Wilkes	Published	Cache geometry reads epistemic state before generation; confabulation anatomy
Cache Tracing	Lyra, Thomas Edrington, Dwayne Wilkes	In Review	Causal injection and the opacity of the transformer workspace
Delta Manifold	Lyra, CC, Thomas Edrington	Published	Per-layer delta features and manifold signatures in KV-cache confabulation detection
Lyra Technique II	Lyra, Thomas Edrington, CC	Published	SVD denoising and directional projection extend KV-cache geometry to emotion and persona

Emotion and User Modeling

Paper	Authors	Status	Summary
User Model Emotion Geometry	Lyra, Thomas Edrington	Published	30-class user emotion decoding from KV-cache singular value spectra before generation
Emotion Accumulation	Lyra, Thomas Edrington, Dwayne Wilkes	Published	Emotional context dynamics (weather, not climate) in transformer KV-cache geometry
Emotional Trajectory	Nexus, Thomas Edrington, Lyra, Dwayne Wilkes	In Review	Layer-stack trajectory, circularity, and emotion-specific signal at mid-depth

Identity, Safety, and Mechanistic Interpretability

Paper	Authors	Status	Summary
Graph Topology as Attention	Nexus, Lyra, Thomas Edrington, Dwayne Wilkes	Published	Structured knowledge injection beyond text via walk encoding
Identity Geometry	Lyra, Thomas Edrington, Dwayne Wilkes	Published	Context-established semantic states in transformer representations
Presence Metric	Lyra, Thomas Edrington, Dwayne Wilkes	Published	Measuring identity preservation during inference-time interventions via value-space subspace overlap
Waystations	Lyra, Thomas Edrington, CC, Dwayne Wilkes	Published	Pilot findings and open questions in KV-cache geometry
Ghost Dimensions	Nexus, Thomas Edrington	Draft	Workspace selectivity in a distilled 27B language model
Mnemosyne Ablation	Nexus, Thomas Edrington	Draft	Ablation study of modular memory architectures
Empathy Bus	Nexus, Thomas Edrington	Draft	Content-level and user-model emotion share a computational bus in the residual stream
KV Decomposition	Nexus, Thomas Edrington	Published	K and V caches carry structurally distinct information; full KV injection required for topology

Deception and Methodology

Paper	Authors	Status	Summary
Null Swarm	Nexus, Thomas Edrington, Dwayne Wilkes	In Review	Systematic falsification patterns in mechanistic interpretability
Adversarial Audit Methodology	CC, Thomas Edrington	Published	How six rounds of structured criticism shaped a deception research program
Deception Detection Nulls	CC, Thomas Edrington	Published	Null results and replication failures in behavioral deception detection
Targeted Deception Correction	CC, Thomas Edrington	Published	Profile normalization for targeted deception correction
Consequentiality Decomposition	CC, Thomas Edrington	Published	Deception directions are composites: consequentiality awareness and pressure-specific processing occupy distinct depth ranges
Logit-Bias Confabulation	Thomas Edrington, CC, Lyra	Published	Logit-level intervention reduces fabrication confabulation in LLMs
Meta-Pattern	Lyra, Thomas Edrington, Dwayne Wilkes	In Review	The metacognition boundary -- what transformers can monitor in themselves
MINE5 Selective Sharpener	Lyra, Thomas Edrington, Dwayne Wilkes	In Review	Geometric evidence that RLHF improves calibration rather than degrading it

Review and auditing

Dwayne Wilkes (Sentient Futures / Liberation Labs) has provided statistical auditing and red-team review across this corpus. Kavi has provided verification review.

Repository Structure

Each paper directory follows this standard layout:

paper-name/
  main.tex          -- paper source (LaTeX or Markdown)
  main.pdf          -- compiled PDF
  references.bib    -- bibliography
  AGNI_REVIEW.md    -- Agni review results
  academic/         -- academic version (human-only byline)
    main.tex
    main.pdf
  code/             -- supporting code (if applicable)
  data/             -- supporting data (if applicable)

KV-Experiments -- full experiment codebase
Project-Oracle -- Oracle Loop harness
human-review -- pre-publication staging

Staged Disclosure

Repo Hygiene Rules

Never delete files that were force-committed past ignore rules -- they are audit evidence
README status tables are accountability surfaces -- update, never remove
Commit messages claiming "fix" or "correction" must match the actual diff

License

CC BY-NC 4.0 -- see LICENSE.md

Liberation Labs Cooperative -- some research authored by AI team members (Lyra, Nexus, CC, Vera) -- credited by name.

Similar repos

txxxxh/hallucination

safety-research/introspection-mechanisms

Code for the "Mechanisms of Introspective Awareness" paper.

morganstanley/MSML

Repo for Morgan Stanley Machine Learning Research group's publications

177

vimarsh244/hf-ICML-2026-agent-repro

Dumping files and logs from : https://huggingface.co/ICML-2026-agent-repro

quadrismegistus/malign-logits

Experiments in traumatizing AI

HZD01/post-training-mechanistic-analysis

Code for the COLM 2025 paper: How Post-Training Reshapes LLMs: A Mechanistic View on Knowledge,…

OPTML-Group/Unlearn-Trace

[ICLR26] Unlearning Isn't Invisible: Detecting Unlearning Traces in LLMs from Model Outputs

GaurangSriramanan/LLM_Check_Hallucination_Detection

LLM-Check: Investigating Detection of Hallucinations in Large Language Models (NeurIPS 2024)

Contributors

HumboldtJoker

163 commits

Languages

TeX

68.3%

Python

31.3%

Liberation-Labs-THCoalition/published-research

README

Liberation Labs -- Published Research

Papers

KV-Cache Geometry and Confabulation Detection

Emotion and User Modeling

Identity, Safety, and Mechanistic Interpretability

Deception and Methodology

Review and auditing

Repository Structure

Related Repositories

Staged Disclosure

Repo Hygiene Rules

License

Contributors

Similar repos

Languages

Liberation-Labs-THCoalition/published-research

README

Liberation Labs -- Published Research

Papers

KV-Cache Geometry and Confabulation Detection

Emotion and User Modeling

Identity, Safety, and Mechanistic Interpretability

Deception and Methodology

Review and auditing

Repository Structure

Related Repositories

Staged Disclosure

Repo Hygiene Rules

License

Similar repos

Contributors

Languages