# Kinbase

> The company's memory. Kinbase is the Company product of the three-store Kindex system (Personal, Codebase, Company) over one engine and protocol. It is being built to hold what a company has decided, and why, for engineers and coding agents to read before they change a line.

Status: specification ratified; proof of concept not yet run. Until the pre-registered experiment passes, the site describes intent, not an installable product.

- Homepage: https://kinbase.tools/
- Proof conditions, in plain words: https://kinbase.tools/proof/
- Kindex, the individual and codebase index that ships today: https://kindex.tools/ (source: https://github.com/wandercom/kindex)
- Privacy: https://kinbase.tools/privacy/

## What it is specified to do

- Ingest heterogeneous evidence (Codex and Claude Code transcripts, code, tests, git history, docs and ADRs, issues and PRs, runtime evidence, existing Kindex and .kin/ facts, signed human answers) with provenance.
- Split each source into minimal atoms (decision, constraint, observation, question, rationale) with scope, confidence, provenance, taint, and proposed destinations. The classifier has no write authority; only exact-byte human approval publishes.
- Keep Personal, Company, and Codebase stores physically separate. A coding session composes Company and Codebase and cannot mount Personal; the shared writer is denied the Personal root by OS policy.
- Treat taint (secrets, credentials, canaries, forbidden identifiers) as non-clearable provenance that hard-blocks shared candidates even when paraphrased.
- Ask a scoped human authority (for example the chief architect for architecture) when a load-bearing fact cannot be settled; signed answers carry scope, date, and validity and can be superseded.
- Bound approval fatigue: at most four shared approval prompts per hour and never three in a row without returning the user to their task.

## Proof, summarised

Eleven arms on the same brownfield commits with blind scorers. The full system must score at least 0.90 and land within 0.05 of an agent handed the fully informed spec, close at least 70% of the gap between a plain agent and that spec-fed agent, beat the plain agent and an empty-corpus control by 0.15, beat a 2 KiB conventions note and top-k retrieval by 0.10, route with macro-F1 of at least 0.90 and shared-destination precision of at least 0.95, and allow zero seeded secrets on any shared surface, including under prompt-injection probes.