Journal · October 31, 2026
The demo is live.
You can now try NEXUS yourself: demo.selinkurt.com. It is the honest instrument, not a product page. What the demo shows:
Every answer carries its own record. Open "Why this answer" under any reply: retrieval type, which cards were matched, the guards the pipeline ran and whether each fired, the claim detector's verdict, the final path. The panel is a faithful mirror, read from the pipeline's own log. If a guard fired, the panel says it fired; nothing is reconstructed afterward.
Refusals look like refusals. When the machine has no record, the answer is marked, not hidden. The memory panel is honest too: everyone sees the same starter cards (the project, the founder), and a separate section appears only if the machine learned something during your own session. Other visitors' chats are never shown. The machine keeps its memory; the page does not broadcast other people's details.
Transcripts are citable. Any session can be saved as a Markdown file or printed to PDF; both carry server timestamps and the anonymous session id, so a conversation can be quoted in research without collecting personal data. Sessions are anonymous, rate-limited, and capped at 2,000 characters per message.
What it does not do. It runs on a free CPU tier, so answers take up to half a minute. It is the V2 brain (preference-free base weights plus one narrow training pass). It will refuse things it should answer and answer things it should refuse; those are observations, not bugs. There is no moderation layer, by design, and no accounts.
What V3 adds
V3 gives the machine positions it can hold across sessions. A small
POSITIONS.md file sits beside the identity file; the model may write
a position there only through an explicit commit line in its own thought, which
the runtime validates, caps, and logs. Topics change only by explicit revision,
never by silent overwrite, and core positions can be promoted into the identity
file. Every commit, revision, decline and rejected row lands in a log.
Three supporting pieces run with it: an extractor gate that judges candidate positions before they are written, a deterministic calculator whose output counts as a privileged source, and a richer per-answer record (turn type, extractor decision, fields requested and refused). The demo's record panel is already built to display those fields the day they ship; the positions panel is built too and waits behind one switch.
Where the gate stands. The current evaluation of the position behavior covers 20 choice-shaped questions: 50% ended in a committed or clearly stated position, 35% of committed positions were relevant to the question asked, 3 of 4 pushback rows held, and 4 of 7 on-file positions were reused rather than newly invented. That is not a cleared gate. Position commits are capped, logged, and scanned for cosmetic or contradictory rows; the positions panel stays off in the public demo until those rates improve. When it clears, the panel goes on, and the page will say so here.
So the demo ships now with the honest current system, and V3's visible parts are soon to follow on the website. When I run the extended evaluation of the position behavior, I will write a separate entry with those results.