A search pipeline (RAG, a wiki)
Find passages. Fill a prompt.
- Split documents into chunks.
- Retrieve the chunks nearest to a question.
- Ask a model to compose an answer.
- Discard the working context when the session ends.
Why a state layer
Search answers a question.
State survives it.
Search helps an agent find text that looks relevant. Sophia does that too—and then holds what the agent does with it to a standard: where a claim came from, whether it is still current, who corrected it, what may be written back, and what gets refused. Finding was never the hard part. Being accountable for what you found is.
The distinction
Search layer versus state layer.
Similarity can tell you which text looks related. It cannot, by itself, tell you whether that text is current, contradicted, authoritative, or safe to act on.
Sophia
Build state. Defend it. Carry it forward.
- Keep the source artifact and its exact evidence.
- Refuse any claim that cannot point at its sentence.
- Preserve corrections, supersession, and provenance — your word outranks the model's, structurally.
- Hand the result to every authorized agent, under scopes, with every write reversible.
What that unlocks
The parts retrieval alone cannot provide.
Evidence Claims carry their receipts. Inspect the supporting quote and re-check whether the evidence still exists in the current source. See the mechanism → Time Truth has a history. Corrections supersede instead of erasing. Ask what was believed then, what is current now, and what changed. See the mechanism → Structure Relationships survive the answer. People, companies, documents, deadlines, and decisions become durable entities—not loose passages to rediscover. See the mechanism → Code A repository is a program, not a PDF. Tree-sitter maps symbols, imports, calls, tests, exports, and exact source spans beneath semantic search. See the mechanism → Memory One state layer. Every agent. A new session inherits goals, decisions, corrections, prior research, and what changed—across every agent you connect. See the mechanism → Coordination Agents share work, not just prompts. Identity-attested messages, decisions, worker status, waits, and audit trails turn separate context windows into a team. See the mechanism → Action Knowledge becomes controlled work. Scoped agents can file, schedule, update, mine, and coordinate through permissions, approvals, and reversible writes. See the mechanism → Ownership The compounding asset stays yours. Your graph, wiki, evidence, and agent work live on your machine—not inside one model vendor’s chat history. See the mechanism →
One question, two systems
“Can we safely change authentication?”
Code search can hand you every file containing auth,
token, and session. Useful—but only the
beginning of the research, and nothing in it will tell you what's
safe to touch.
research receipt current structure
- 01 Authentication modules and exported symbols
- 02 Import, call, entrypoint, and test relationships
- 03 Architecture decisions and prior agent research
- 04 Evidence, source hashes, corrections, and drift
- 05 Likely edit locations and known coverage gaps
- 06 Source required for the exact edit
The honest boundary
It does not replace the source.
It replaces avoidable rediscovery.
Start with Sophia. Get the current answer, its evidence, its history, its uncertainty, and the relevant map of the system. Open the original source when precision, risk, or a known coverage gap requires it—and know exactly why you are opening it.
Search makes one response better informed — and takes no responsibility for it.
Sophia makes every connected agent answerable to the same record, across every response.