AV Family EnterpriseResearch · Software · Accountable systems

Research & Demonstrations

The portfolio, by research area

Founder-led work published through The Temple of Two, presented under its own licenses and authorship. Every entry carries three labels so a reviewer knows what they are looking at: Type (instrument, study, RFC, hypothesis set), Lifecycle (living, published, closed, awaiting validation), and Evidence (what the record actually supports).

01

Governed AI & agent authorization

cosmic-cli terminal: a scripted mission ends verified only after an operator-supplied check exits 0Scripted mission · No model call

COSMIC-ALLOW

A cross-vendor proof-of-allow protocol for AI coding-agent cockpits: deny is the ground state, allow requires an unforgeable sentinel, and a kernel sandbox floor — not the policy layer — is the load-bearing boundary.

Type
RFC + reference implementation (cosmic-cli, Apache-2.0)
Lifecycle
Published · v1.1
Evidence
Supported within scope — the protocol holds across vendors; not a claim that any deployment is secure

What failed

Policy-layer-only enforcement: it can be talked past by the thing it governs.

DOI ↗cosmic-cli ↗
Compass Observatory live feed: shell commands classified OPEN, PAUSE or WITNESS before they runSynthetic session · Local dashboard

T2Helix

An editor plugin placing a governance check before a tool call rather than in the log afterward: destructive operations are hard-denied, credential use is gated, and method candidates are quarantined until a human promotes them. Local-only storage.

Type
Instrument
Lifecycle
Living · v0.13.1
Evidence
In use, local-only — a demonstration of placement, not a measured claim about how often it prevents harm

What failed

The advisory-banner version was ignored, by the human as often as the model.

Repository ↗
Project Epistemic Bound workroom showing one recorded scripted run: its budget, evaluation report and outcome checks, each traced to eventsScripted provider · Recorded run
The workroom separates a proposal, an authorization, and an observed effect.

Project Epistemic Bound

A local workroom and evaluation workbench recording an AI agent’s assignments, permissions, proposals, corrections, and observed effects as a hash-linked event chain — keeping a proposal, an authorization, and an effect apart.

Type
Instrument
Lifecycle
v0.1 integration state · release blocked
Evidence
Untested — no behavioral result about any model yet
Repository ↗

02

Persistent knowledge & research infrastructure

Sovereign Stack

A persistent memory and governance layer with a self-verifying record: each claim’s identity is derived from its content (sha256), carries verified-by receipts, and is superseded rather than edited. Continuity treated as an infrastructure property, not a model property.

Type
Instrument
Lifecycle
Living · v1.21.0 · 52 tools on the running surface (heartbeat 2026-09-19)
Evidence
In use, self-verifying — demonstrates a design; does not test a hypothesis about minds

What failed

The first design let a claim be edited in place, which made the record unfalsifiable. Supersession replaced mutation.

Repository ↗Public chronicle mirror ↗

temple-harness

A read-only access layer that gives a local model running on the machine controlled recall of the record — seven named read-only doors, no master credential, and a canary suite that mutates each gate in turn to prove the checks are still there.

Type
Release
Lifecycle
Published · shim 0.4.0 (2026-09-07)
Evidence
Publicly verifiable; second-seat review observed, not yet enforceable
Repository ↗

03

AI evaluation & human–AI workflows

The Black Box Boundary

A non-interventional, byte-exact reconstruction of the inference-event boundary around a small deployed language model. In the studied session the operator’s words were 0.56–14.00% of model-input bytes; the rest was enumerable application-layer material — inspectable, and never inspected.

Type
Study
Lifecycle
Published (2026-07-29)
Evidence
Bounded, n=1 session — measures the boundary, not the weights; not an interpretability claim
DOI ↗Code + evidence ↗

Governed co-creation case study

A documented investigation of how human direction, machine contribution, approval, attribution, and correction interact in one research practice — with a public retrospective of the occasions when claims exceeded the evidence.

Type
Case study + retrospective
Lifecycle
Published, maintained
Evidence
Single documented practice — not a population-level validation of human–AI collaboration
Retrospective ↗
Master Operations Agent dashboard withholding its assessment because a synthetic observation is too oldSynthetic observation · Offline baseline
Why a finding was released or held: freshness, required reads, envelope checks, content validation.

Master Operations Agent

A research workbench for an evidence-first operations advisor on synthetic observations: bounded read tools, an independent validator, and a hash-linked evidence log. v0.7: the frozen policy-only model candidate scored 7 of 24 useful answers on the private synthetic holdout; the deterministic baseline scored 24 of 24. The failed acceptance is retained.

Type
Research workbench
Lifecycle
v0.7 · ongoing
Evidence
Development results, not model promotion — not a trained operations expert, autonomous controller, or plant-ready product
Repository and receipts ↗

04

Local-first AI & architecture R&D

Phenomenological Compass

A two-stage inference architecture: a small conditioning model reads the epistemic weight of a question and conditions a larger action model. Ships as a native macOS app, fully local on Apple Silicon. Token-level entropy profiling shows the conditioning restructures the probability distribution rather than restyling the output.

Type
Study + instrument
Lifecycle
Published (2026-04-02)
Evidence
Supported within scope — 96% signal accuracy (101/105) and ΔH = +0.47 nats on one architecture family and one evaluation set; not yet replicated across model scales

What changed

The first framing treated the compass as an output filter. A filter cannot move token-level entropy before the tokens exist; the mechanism is conditioning, a different claim.

DOI ↗Repository ↗

Phase-Modulated Attention

A 176M-parameter hybrid SSM-attention language model in which Kuramoto oscillators modulate attention routing. System-prompt framing induces attention-dependent entropy regime switching; placement — gating information flow rather than accompanying it — is the variable that matters.

Type
Study
Lifecycle
Published (2026-02-28)
Evidence
Supported within scope, standing on a published negative result

Negative result retained

Liminal K-SSM: oscillators in hidden state were epiphenomenal — no language-modelling effect. That failure located the causal entry point.

DOI ↗Repository ↗

05

Scientific modeling & simulation

Entropy as a Tunable Equilibrium

Four pre-registered computational campaigns (v4.0–v4.3), each committing in advance to the result that would falsify the prior version. The horizon knob is real; the geometry-scaling prediction that would have made entropy an “engine” failed; an apparent short-horizon effect dissolved under matched-estimator control.

Type
Study
Lifecycle
Closed · v4.3 (2026-07-10)
Evidence
Bounded null — the correction is published under its own DOI, not folded into a later version
DOI ↗Kinetics follow-up ↗
Unit 03 graphic: fired preheater and exothermic fixed-bed reactor during a drillSynthetic training environment

Independent operator training simulator

A browser-based training simulator: four synthetic process units, an ISA-18.2 alarm lifecycle, PID faceplates, trends, an event journal, scored drills, an instructor station with snapshots and replay, an architecture-training view, and a self-contained offline build.

Type
Independent training aid
Lifecycle
Released · v3.1.0 (2026-09-03)
Evidence
Shipping — not a Honeywell product; drill scores independent of any vendor certification; not a diagnostic display or live control
Repository ↗

Temple Lab

An interactive 3D chemistry workbench — 118 elements, reference molecules with published geometries, an electron lab, a bonding coach, and a reaction lab — in which each tool states on its face what it does not do. A missing value reads “Not available”, never zero.

Type
Release (MIT)
Lifecycle
Published · v1.3.0 (2026-09-10)
Evidence
Shipping, not a predictor — nothing in it predicts products, rates, or conditions
Releases ↗

06

Computational biology & pharmacology research

VDAC1 rendered as a gate, annotated with hexokinase-II, Bcl-xL, and the cholesterol ratioScientific illustration
Scientific illustration from the Atlas — a hypothesis diagram, not a measurement.

VDAC1 Pharmacology Atlas

The voltage-dependent anion channel modelled as a druggable mitochondrial decision gate: a cofactor equation for how hexokinase-II, Bcl-xL, and cholesterol jointly set the apoptotic threshold, expanded into six atlas layers and 22 mechanisms, plus the Gate-Jamming Score as a candidate biomarker for immune evasion.

Type
Biomedical hypothesis set
Lifecycle
Awaiting wet-lab validation
Evidence
Untested in vivo — 24 testable predictions, 0 tested at the bench. Two of three transcriptomic cohorts returned nulls; the surviving signal is stratum-specific (n=209). Literature concordance is not validation.

Next validation step

Bench assays with a wet-lab collaborator; electrophysiology for channel recording; independent pharmacology review.

Atlas · OSF ↗Monograph DOI ↗Cohort preprint ↗

Descriptions and standing labels are quoted from each project’s public record and were last reviewed on 26 September 2026. A project’s appearance here does not change its license or ownership. Metrics carry their scope and date; read the linked source before relying on one.

Propose a collaboration