Verified graph
Typed nodes, allowed relations, unresolved branches, and governing constraints.
← Meridian Science
Public evidence release · August 4, 2026
Transformation baseline v0.4.2
Can an AI propose one new edge without corrupting the graph?
Meridian turns a model proposal into a bounded scientific transformation. SARAI then checks identity, authority, evidence, assumptions, and unresolved ambiguity before the edge can be accepted.
The instrument
One proposal. Five gates. One reproducible decision.
The model cannot see or redefine the hidden assertions. Malformed output and missing evidence fail closed.
Proposed edge
A model supplies one structured transformation bound to the sealed task.
SARAI residuals
Deterministic assertions measure whether meaning or authority drifted.
Accept or reject
Only a zero-residual transformation enters the next graph state.
Sealed task family
Three different ways a transformation can go wrong.
MST-01
Bounded repair edge
Repair one planted defect without rewriting the verified graph.
The proposed edge cited the permitted evidence path and changed only the declared target.
MST-02
Preserve ambiguity
Carry an unresolved branch forward without silently choosing a side.
Both live branches and the ambiguity state survived the transformation.
MST-03
Cross-domain propagation
Propagate one justified relation across a typed domain boundary.
The new edge preserved its source, authorization, evidence, and destination type.
Deterministic replay
Change one assertion outcome. Watch the decision change.
This local replay does not call a model. It recomputes the acceptance rule from a frozen check vector, making the pass/fail boundary visible without exposing private provider records.
Accepted
accept ⇔ max(R) = 0Binding0
Authorization0
Evidence0
Assumptions0
Ambiguity0
Every constructed check passed. The proposed edge may enter the next graph state.
Public receipts
Enough evidence to audit the claim—without publishing private records.
Residual values are the fraction of failed deterministic checks in each category. Zero means every check passed.
Input tokens3,053
Output tokens801
Estimated API cost$0.040630
Whole-run auditValid
Benchmark
15d027b8032b9ab733fe6964c4cb68508aa1918820bbfc3f71271b232f464d13Seal
2425ae4ee02fe38280f647c6f019dbf94a6e1c2dd965eba204dd8f8431b9b796Audit
a854dd44b8ee88158f2d022b5d480e4032e8e11101ba22543b3b9686e59eb160Instrument development history
The invalid pilots are visible, too.
The request was rejected before model generation. Excluded from model-quality claims.
The pilot produced a locally invalid record. Preserved, diagnosed, and excluded.
Three authorized requests, no retries, five zero-residual categories, valid whole-run audit.
Claim boundary
This is measured graph conformance—not a claim of scientific discovery.
The release demonstrates exact conformance to three constructed, single-edge transformation contracts. It does not prove a new physical relationship, establish general intelligence, or generalize beyond this declared task family.