KeilerHirsch-Labs is an independent engineering lab for AI metrology, high-assurance software and local-first systems.

We do not start with a feature backlog. We start with the system underneath it:

• What exactly is being measured?
• Which invariants must remain true?
• Where is the real trust boundary?
• Which failures are separate defects, and which are symptoms of one shared cause?
• What evidence would falsify the current explanation?

The common method is measurement → reproduction → evidence → root-cause isolation → small reviewable change → independent verification.

Research-first infrastructure for reproducible, uncertainty-aware and auditable measurement of AI-system behaviour.

BRONCO starts with metrology and experimental validity rather than a leaderboard: measurands, construct validity, repeatability/reproducibility, uncertainty, provenance and evidence come before benchmark features.

Its research foundation uses DIN/ISO/IEC and JCGM metrology work as engineering references. A deliberately small Ada/SPARK trusted core is being developed for measurement-critical deterministic logic. Provider APIs, model execution, orchestration and other rapidly changing components remain outside that trusted boundary.

Status: Foundation / Research · License: PolyForm Noncommercial 1.0.0

Standards references describe engineering alignment and traceability. They do not imply DIN/ISO/IEC affiliation, certification or laboratory accreditation. The presence of SPARK source likewise does not imply that planned proof gates have already been completed.

Field engineering against MemPalace, driven by real workloads rather than synthetic feature requests.

A systematic review produced many apparent defects that repeatedly clustered into a smaller set of failure families. Work therefore concentrated on shared invariants and recovery boundaries: single-writer semantics, rebuild/repair safety, native-index divergence, ingest completeness, platform encoding and Windows concurrency.

The public contribution trail includes upstream-merged fixes for atomic repair behaviour, large-palace MCP startup, HNSW-divergence preflights, re-mine completeness and writer-lease protection. Open work remains separate until current upstream behaviour is re-validated.

Status: Upstream hardening / research track · Working fork: KeilerHirsch/mempalace

Local-first export tooling for your own Claude conversations, project documents and memory.

The current supported path is Windows + Claude Desktop + a real Chrome/CDP session, producing local Markdown without telemetry or a hosted archive dependency. Security-sensitive credential handling and its limitations are documented explicitly in the repository threat model.

Status: Early Access · License: AGPL-3.0

• Evidence before claims — preserve raw observations, provenance and reproducible procedures.
• Root cause before patch count — a large symptom list may be a small architecture problem wearing many disguises.
• Research before features — define what is being measured before building a dashboard around the number.
• Invariants before convenience — identify properties that must stay true across platforms, failures and maintenance paths.
• Local control where practical — user-owned data should remain inspectable and portable.
• Small trusted computing bases — isolate high-assurance logic instead of assigning equal trust to an entire changing stack.
• Tests before confidence — automated checks are evidence for the properties they actually exercise, not a universal quality certificate.
• Formal methods where they buy real assurance — contracts and proofs belong around critical deterministic logic, not in marketing copy.
• No silent uncertainty — limitations, assumptions, dead hypotheses and unresolved ambiguity stay in the engineering record.
• Controlled technical language — use canonical terms, explicit evidence qualifiers and low-ambiguity prose. See KTCP.

Issues are the preferred public entry point for bug reports, hostile design review, research criticism, reproducibility findings and technical proposals.

For BRONCO during the foundation phase, a strong counterexample, standards correction or argument that the trust boundary is wrong is generally more valuable than another feature idea.

For upstream hardening work, a minimal reproducer and a shared failure mechanism are more useful than a long list of superficially unrelated symptoms.

_{Org avatar built from five icons by Lorc via game-icons.net ("Stag Head", "Boar Tusks", "Erlenmeyer", "Test Tubes", "Round Bottom Flask"), recolored and recomposed, CC BY 3.0. Full credits in CREDITS.md.}