toolvingYou can see why it answered.
Conceiver works with your LLM to provide System Two thinking. It recalls evidence from what you taught it and what happened before, weighs that evidence, and shows why it answered the way it did. It is not an LLM. It has no weights and does not use backpropagation.
The pieces
| Object | What it is |
|---|---|
| Instance | The active context where you teach and query. |
| Learning | The examples and observations your account has accumulated. |
| Sequence | A saved, ordered record that can teach a fresh Conceiver the same capabilities. |
| Prediction | An answer or unresolved result returned for a query. |
| Evidence | Supporting and contradicting examples supplied with a decision. |
| Episode | Contextual memory that a memory-enabled query can recall in full. |
| Outcome | The answer recorded after you know what actually happened. |
| Operation | A request you can look up or resume if its response was lost. |
Teach, ask and resolve
Teaching provides an input and its expected output. One explicit example establishes the exact transformation when that input appears again.
A query returns ANSWER when the available evidence supports an answer. Otherwise it returns UNRESOLVED with relevant evidence for downstream decision-making. Your application can pass that evidence to an agent's LLM or a person. Conceiver does not invent an answer to fill the gap.
Resolving a prediction records what actually happened. Contradictions lower confidence. When answers disagree, Conceiver abstains and provides the available evidence instead of selecting one.
Examples become evidence
Independent supporting examples establish three evidence levels:
| Level | What it means |
|---|---|
| Hypothesis | One supporting example. |
| Confident | Two independent supporting examples. |
| High | Three or more independent supporting examples. |
These are evidence levels, not calibrated probability percentages. A multi-step answer is only as strong as its least-supported step. Check status before using a result, then apply the evidence threshold appropriate for your application.
Reading its reasoning
Every decision includes a short Markdown record. It shows where the answer came from, which examples support it, and which examples point the other way. "Lossless provenance" refers to this decision record, not guaranteed retrieval of every past episode.
Episodic memory
A memory-enabled query uses contextual evidence and can recall full episodes. It costs 2 credits. Standard predictions and retrievals cost 1 credit.
We recommend storage for episodic memory. Without storage, the instance and its episodic memories disappear when the instance retires. With storage, its memories and learned capabilities are saved as an artifact that a fresh session can load.
Training sequences
Copying or saving asks the private plastic trie which teaching evidence still supports learned constructions. Copying returns the assembled text without server-side storage. Saving writes those question-and-answer pairs to a compact .txt account file and therefore requires storage. Each question and answer occupies two adjacent lines; one empty line separates teachings. Queries that did not teach a construction are omitted.
A saved file can be ingested into the same compatible model. Toolving parses the line pairs and sends each pair through the ordinary teaching path. Repeated ingestion of the same file into one compatibility scope is idempotent. Each account and model scope remains isolated.