Cursiv
Direction
Trustworthy first. Then bigger.
Cursiv is meant to become a local intelligence system that keeps reliable knowledge even when connectivity, hardware, information or the AI itself are unreliable. It gets there one level at a time, and each level has to be shown to work before the next one is built.
Connectivity is optional. Trust is not.
What "trustworthy" means here
Honest about what it knows.
Not "always right". No AI is. The goal is that Cursiv is always clear about what it knows, what it inferred, what it's assuming, and what it can't tell, and that it says "I don't have enough evidence" instead of inventing an answer.
Kinds of knowledge kept apart
- Memory: what a person told Cursiv
- Reference: curated material, with its source
- Observation: something a person or instrument recorded
- Inference: a conclusion drawn from the above
- Hypothesis: a possible explanation, not established
- Unknown: not enough evidence
Today Cursiv separates memory origin and Known / Inferred / Assumed / Unknown in answers. Observations and hypotheses as records are planned.
What has to be demonstrated
Five things, each shown with evidence.
Not more agents or features: proof that it works.
| Goal | Status | Where it stands |
|---|---|---|
| 1. Reliable answers A repeatable benchmark shows how often Cursiv is right, when it invents things, and whether it recognizes missing evidence. | Baseline done | Historical baseline: suite 1.3 had 133 cases; U65 scored 126/133 and U66 scored 132/133. The current suite 1.5 has 158 cases. The latest published result is U75 on 153 shared cases; it is not a full current-suite run. The repository changelog's latest documented app release is U77, with no U77 benchmark result. See the current method and results. |
| 2. Genuine offline operation Disconnect the computer and show what still works. | Demonstrated | One documented run on October 18, 2026 tested 15 listed scenarios on a fresh copy with qwen3:4b-instruct, offline mode and socket-level outbound blocking. All listed checks passed and there were no outbound attempts. This is a subset of scenarios, not a feature-by-feature test of every capability or configuration. See the run record. |
| 3. Traceable knowledge Tell a recorded observation from an AI inference, and show where claims came from. | Started | Memory origin tags; 37 reference entries written from named public sources, each showing its source and check date. Observation records and a provenance ledger are planned. |
| 4. Resilience Survive interrupted connections, restarts, conflicting records and sync failures without silently corrupting information. | Planned | Not yet tested. |
| 5. A field demonstration Someone who didn't build Cursiv uses it to solve a real problem and evaluates the result independently. | Planned | After 1–4. |
The build order
Each level depends on the one below it.
Specialized applicationsFor example field science or mission support. Only after everything below is proven. Would need independent evaluation and validation.
▲
Local hubGathers authorized information from several nodes; larger models; timeline; backups. Works without internet.
▲
Independent nodes and synchronizationEach Cursiv works alone; when a link appears (network, USB, anything) records sync, with duplicate, conflict and integrity checks.
▲
Structured knowledgeObservations, an append-only record of events with provenance, corrections and retractions; "why do you believe this?"
▲
Reliable local AI — nowAnswer checking, offline reference, memory with origin, offline mode, and the truth benchmark to measure it.
Next: field science
A dependable offline tool for research in the field.
Where recording data correctly matters more than clever answers. Each step ends with something a reviewer can check.
| Step | What | Status |
|---|---|---|
| Field notebook | Structured observations kept apart from AI inferences; corrections never overwrite; export to CSV, JSON, GeoJSON. | Built and tested |
| Calculations in code | Units, dilutions, coordinates, statistics: computed, not guessed. | Built and tested |
| The team's documents | Protocols and field guides offline, answers quoting the page. | Built and tested |
| Integrity and chain of custody | Tamper-evident entries and sample IDs with printable labels (built); signed exports next. | Started |
| Hardware and endurance | Battery, no-GPU laptops, power loss, the SSD kit on clean machines. | Started |
| Team sync without internet | Laptops share field logs by USB drive or shared folder; changed copies refused, conflicts shown, never silently resolved; personal memory never syncs. | Built and tested |
| Safety and discipline packs | Sourced expedition safety, then ecology, geology, hydrology and atmospheric references from open sources. | Started |
| Field pilot | A real team on a real trip, with the results published, including failures. | Planned |
Not the goal
What Cursiv isn't trying to be.
- Another general chatbot competing on personality or size.
- Features added because they sound futuristic.
- A system that claims perfect accuracy, or certification it doesn't have. Cursiv is not certified for medical, safety-critical, aerospace or spaceflight use.
- A product that needs the cloud. Online services stay optional.
- A distributed system built before the single-computer version is shown to be reliable.