HoundtidLabs

Research

Notes from the work.

Houndtid Labs publishes research as it happens — fine-tuning runs, infrastructure decisions, and the reasoning behind how the model constitution evolves. This is a working log, not a highlight reel.

Fine-tuningJune 2026

Domain adaptation for Uganda land survey documents

A QLoRA fine-tuning pipeline built on Gemma 4, trained against 991 distilled examples sourced from Institute of Survey, Entebbe materials — teaching Houndtid the structure, precedent, and vernacular of Uganda's land record in a way no general-purpose model has attempted.

InfrastructureMay 2026

Row-level security as a default, not a configuration

Houndtid Tensor's 41-table ontology is built with RLS policies enforced at the database layer for every actor role — meaning tenant isolation cannot be bypassed by an application-layer defect, a property most spatial platforms only add after an incident.

Model BehaviorOngoing

A versioned constitution for grounded uncertainty

Rather than tuning Houndtid to sound confident, our constitution work focuses on calibration — teaching the system when to say it doesn't know, and to distinguish a retrieved fact from an inferred one at the sentence level.

Conflict DetectionOngoing

Spatial conflict detection across overlapping land claims

Houndtid Tensor's spatial engine flags overlapping parcel claims automatically — a capability aimed at surfacing disputes before they escalate, built on top of Kampala's own GIS dataset.