Evidence
Every claim links to the record that produced it, in the repository. A card tagged measured · negative records something that did not work — those are the ones worth reading. PROVEN was measured here; BUILT is tested but not yet exercised on live work.
What holds
Zero runtime dependencies
A fresh clone served the dashboard with no install step, in seconds.
Binds loopback only
Remote access is Tailscale's job; the request guard was verified with spoofed Host/Origin headers (403).
A real three-node fleet
A 10 GB desktop, a 6 GB laptop, and a CPU-only box, spanning a private network with one node in another state.
Routing on the measured budget
A model is refused on the node's measured budget, not its nameplate — the work that found a VRAM estimate 3.7× wrong.
Independent review, enforced
Every finding is re-judged by a different model family in a fresh context; 27 pairings, zero same-family.
A judge earns admission or is refused
A judge that cannot discriminate is refused; today every local model measured is refused.
A degenerate judge is caught
One answered REJECT on 35 of 36 cases and was flagged in the act.
Writes stay inside the contract
A write outside a task's declared paths quarantines the task; the merge is human, fast-forward only.
Local context is token-budgeted
Each role gets one declared token budget, with what was shortened or excluded recorded.
Hermes has one door
One live read-only retrieval query; eight tools refused by name.
A document, bound and contained
One PDF ingested, verified by hash, bound to one repository, and did not leak to another.
What the measurements refused
Three small models matched one 30B
Equal defect recall (0.708), 19× less author time — and the two arms found different defects (union 0.875).
Every panel below its best member
0.972 alone against 0.924 for the best three-member vote. The adjudicator is a cascade of one.
A keyword scan beat the embedder 6–4
On the memory fixture, weighted word presence found the labelled location in six of ten cases; the embedder found four.
9% of the corpus was unreachable
The chunker truncated long sections for the life of the index. Fixed, with drift now reported.
Built, not yet live
The egress gate
Eleven material classes, one may leave, default refuse, enforced at one chokepoint; 18 tests and three structural guards verified by breaking them.
The cloud call chain
Every branch settles the money; twelve staged attacks held, and the one that did not — an advisory gate — was found by trying and closed.
A provider, declared and off
One provider is a disabled peer; a missing key is inert. No cloud call has ever been made.
The spend ledger
Micro-dollars, reserve/commit/release, three caps; held under six concurrent processes after a cross-process race and a crash were found and fixed.