Skip to content

ruvnet.

Python

docs(ruvnet): record 2026-09-17 improvement evidence

rUv7965670321cc
diff --git a/docs/self-improvement/2026-09-17-evidence.json b/docs/self-improvement/2026-09-17-evidence.json
new file mode 100644
index 0000000..013cc50
--- /dev/null
+++ b/docs/self-improvement/2026-09-17-evidence.json
@@ -0,0 +1,67 @@
+{
+ "schemaVersion": 1,
+ "date": "2026-09-17",
+ "decision": "accept",
+ "baseline": "846201b3baa898349f43a09348a7981353c27c6a",
+ "candidateCommit": "4a5a86084ad020e373367c7b3fbca812ff822871",
+ "candidate": "content-addressed-snapshot-contract",
+ "version": "0.5.0",
+ "snapshotId": "sha256:35f2144acebf730b1a434eddef6b30ebe36d166404921a710ff43814695a3edc",
+ "upstream": {
+ "observedAt": "2026-09-17T10:37:27Z",
+ "sources": [
+ {"project":"ruflo","revision":"6f0ed7112873eedc7cfe17281a2585188190b790","release":"3.42.3","license":"MIT","changedSincePrior":true,"incorporated":"release, quote-aware hook override, Windows root-publish compatibility and MCP policy-test audit isolation"},
+ {"project":"metaharness","revision":"d5833dc6512ac1adeeef91a331c29055cd8a4dbb","license":"MIT","changedSincePrior":false},
+ {"project":"ruvector","revision":"b336fbae8b15a5a487fc3e7c14a80dcb4984b04a","license":"MIT","changedSincePrior":true,"incorporated":"RVF/VectorDB boundary correction, fail-closed stub writes, optional MinCut WASM wrapper and artifact/NAPI compatibility records"},
+ {"project":"agentdb","revision":"6cdefe451f6afe7395bc88d3587a0225d161089b","license":"MIT","changedSincePrior":false},
+ {"project":"dream-machine","revision":"3edd426f6c9c4b1e80235f7447dc863e749345cc","license":"MIT","changedSincePrior":false,"incorporated":false},
+ {"project":"autogenous","revision":"905aa6cbe213392f8b3cab5d4f17bc3a48e0a509","license":"MIT","changedSincePrior":false,"incorporated":false},
+ {"project":"rvm","revision":"0973d77e5a9e99065376e8ef772a4a1d16e88c75","license":"no SPDX license detected by GitHub API","changedSincePrior":false,"incorporated":false}
+ ],
+ "newRecordsIncorporated": 8,
+ "xRuvIo": {"version":"0.7.2","routes":["/mcp","/chatgpt/mcp"],"changedSincePrior":false}
+ },
+ "heldOut": {"baseline":{"passed":1,"total":6},"candidate":{"passed":6,"total":6}},
+ "tests": {"groupsPassed":11,"groupsFailed":0,"node22":true,"node24":true},
+ "validation": {"plugin":true,"skillsPassed":4,"marketplaceSchemas":2,"packageFiles":18,"compressedBytes":23974,"unpackedBytes":75073,"lockfileSha256":"40b39122e48ea9bcfb42e188cb6a94ba58787e9750e7a713cb72165386c8eb8f"},
+ "benchmarks": {
+ "raw":"2026-09-17-benchmarks.json",
+ "pairs":15,
+ "baseline":{"discoveryP95MedianMs":0.0010520000000013852,"coldCliP95MedianMs":72.11870899999985,"rssMedianBytes":47722496},
+ "candidate":{"discoveryP95MedianMs":0.00100100000000225,"coldCliP95MedianMs":72.9613240000001,"rssMedianBytes":47968256},
+ "delta":{"discoveryP95Percent":-4.847908745158548,"coldCliP95Percent":1.1683722735528344,"coldCliP95Ms":0.8426150000002508,"rssBytes":245760},
+ "interpretation":"All frozen regression gates passed; timing changes are not claimed as product speed improvements."
+ },
+ "metaharness": {"package":"[email protected]","scope":"plugins/ruvnet","hardConstraints":"6/6","baseline":{"harnessFit":60,"compileConfidence":65,"taskCoverage":84,"toolSafety":90,"memoryUsefulness":38},"candidate":{"harnessFit":60,"compileConfidence":65,"taskCoverage":84,"toolSafety":90,"memoryUsefulness":38}},
+ "security": {
+ "rufloPackage":"@claude-flow/[email protected]",
+ "deepCodeScan":{"critical":0,"high":0,"medium":0,"low":0},
+ "deepDependencyScan":{"critical":0,"high":0,"medium":0,"low":0},
+ "npmAudit":{"critical":0,"high":0,"moderate":0,"low":0},
+ "secretScan":{"findings":0,"filesReportedByScanner":5},
+ "manualTrackedFileCredentialPatternHits":0,
+ "threatScan":{"indicators":0,"filesReportedByScanner":4},
+ "trustBoundaries":["CLI positional arguments","MCP tool arguments","opaque cursor and caller-supplied snapshot identifier","reviewed upstream JSON","host interpretation of annotations and linked evidence"],
+ "mitigations":["strict sha256 identifier grammar","canonical finite acyclic JSON hashing","cursor binding to snapshot, filters and offset","fail-closed expected-snapshot precondition","read-only offline implementation and MCP annotations","72-hour freshness state","64 KiB MCP frame bound"],
+ "residualRisks":["The snapshot digest provides consistency, not publisher authentication or a signature","MCP annotations are advisory to hosts","reviewed evidence can become stale","linked upstream claims were not independently reproduced","secret and STRIDE scanners reported narrow coverage"],
+ "limitations":"Zero reported findings is not proof of complete security."
+ },
+ "memoryExperiment": {
+ "catalogProjects":10,"changeRecords":36,"eligiblePublicRecords":46,
+ "ruvectorRevision":"b336fbae8b15a5a487fc3e7c14a80dcb4984b04a",
+ "supportedApisInspected":["ruvector VectorDB with flat and HNSW indexes","@ruvector/wasm VectorDB adapter","ruvector-gnn-rerank research surface","optional @ruvector/mincut-wasm wrapper with exact TypeScript fallback, DynamicMinCut and RoadRouter"],
+ "exactBaseline":"current allow-listed filters, snapshot-bound cursor paging and deterministic content digest",
+ "ruvectorWasm":"deferred","gnn":"rejected for this workload","persistentLearning":"not implemented","artifactSize":"not measured because no WASM or GNN dependency was adopted",
+ "reason":"The sanitized corpus remains below the frozen 100-record minimum. The newly inspected MinCut WASM surface solves graph cut/routing rather than provenance retrieval, so adopting it would add complexity without demonstrated held-out relevance benefit.",
+ "futureGate":{"minimumRecords":100,"minimumTaskSuccessGainPercentagePoints":10,"p95MaximumMs":20,"memoryMaximumBytes":134217728,"requiredAblations":["exact filters","vector only","exact plus vector fusion","GNN rerank","shuffled-graph control"]}
+ },
+ "forecast": {
+ "prediction":"Within 90 days, agent evidence and memory interfaces will increasingly expose content-addressed snapshot or receipt identifiers so caches, replay and audit can fail closed across state transitions.",
+ "status":"prediction, not established result","confidencePercent":82,"horizonDate":"2026-12-16",
+ "criterion":"At least two primary agent-memory, MCP, or run-evidence implementations newly publish digest-, receipt-, or revision-bound snapshot metadata used for cache invalidation or replay verification.",
+ "sources":["https://modelcontextprotocol.io/specification/2026-07-28/server/utilities/caching","https://arxiv.org/abs/2609.12582","https://arxiv.org/abs/2608.02843","https://arxiv.org/abs/2605.25002"],
+ "priorForecastsExpired":[]
+ },
+ "ci":{"run":"https://github.com/ruvnet/ruvnet/actions/runs/35212497317","status":"success","node22":"success","node24":"success"},
+ "dashboard":"https://ruvnet-harness-evidence.ruv.chatgpt.site"
+}
diff --git a/docs/self-improvement/2026-09-17.md b/docs/self-improvement/2026-09-17.md
new file mode 100644
index 0000000..29f42e7
--- /dev/null
+++ b/docs/self-improvement/2026-09-17.md
@@ -0,0 +1,33 @@
+# RuVnet harness improvement — 2026-09-17
+
+## Decision
+
+Accepted a content-addressed snapshot contract and published RuVnet entry point 0.5.0. Every reviewed-change page now exposes a deterministic SHA-256 identifier. CLI and MCP callers can send that identifier with a continuation cursor; a changed snapshot fails closed before returning records, preventing pages from different evidence states from being mixed. The digest is a consistency check, not a signature, publisher identity or authorization token. The companion remains offline, read only and unable to refresh or learn by itself.
+
+## Upstream review
+
+Ruflo advanced to [3.42.3](https://github.com/ruvnet/ruflo/releases/tag/v3.42.3) and head `6f0ed7112873eedc7cfe17281a2585188190b790`. Four incorporated records cover the release, quote-aware hook parsing, a Windows root-publish fix and UUID-isolated policy-audit tests. RuVector advanced to `b336fbae8b15a5a487fc3e7c14a80dcb4984b04a`; four records cover its corrected RVF/VectorDB boundary, fail-closed stub writes, optional MinCut WASM wrapper, generated WASM packages and refreshed NAPI binaries. These are reviewed upstream claims, not locally reproduced performance claims. MetaHarness, AgentDB, Dream Machine and Autogenous were unchanged. x.ruv.io still reports 0.7.2 and the established `/mcp` and `/chatgpt/mcp` routes. RVM remains pinned with no SPDX license detectable through GitHub's license API.
+
+## MetaHarness loop and measurements
+
+The [frozen pre-implementation plan](2026-09-17-plan.json) defines the hypothesis and gates. Baseline `846201b3baa898349f43a09348a7981353c27c6a` passed 1/6 held-out snapshot tasks; candidate `4a5a86084ad020e373367c7b3fbca812ff822871` passed 6/6. All 11 test groups, Node 22 and 24, both marketplace schemas, four portable skills, MCP SDK subprocess/in-memory paths, package closure and negative boundaries passed. MetaHarness 0.4.16 remained at 6/6 hard constraints; its unchanged readiness score is not evidence of the measured improvement.
+
+Fifteen alternating pairs used 10,000 discovery calls and 20 cold CLI starts per run. Median discovery p95 moved from 0.001052 ms to 0.001001 ms (-4.85%), cold CLI p95 from 72.119 ms to 72.961 ms (+1.17%, +0.843 ms), and RSS rose 245,760 bytes. All frozen regression gates passed. These local microtimings are not product speed claims; [raw samples](2026-09-17-benchmarks.json) include a 223.85 ms baseline cold-start outlier.
+
+Pinned Ruflo 3.25.6 deep code and dependency scans, npm audit, and a tracked-file credential-pattern check reported zero known findings. The secret scanner covered five files and STRIDE four, so coverage is explicitly limited. Trust boundaries include CLI/MCP inputs, the caller-supplied snapshot ID, opaque cursors, reviewed JSON and linked untrusted evidence. Residual risks include stale evidence, advisory host annotations, an unsigned public digest and upstream claims not independently reproduced. This is not a claim of complete security.
+
+[Candidate commit](https://github.com/ruvnet/ruvnet/commit/4a5a86084ad020e373367c7b3fbca812ff822871) passed [Node 22 and 24 CI](https://github.com/ruvnet/ruvnet/actions/runs/35212497317).
+
+## RuVector, WASM and GNN decision
+
+At pinned RuVector `b336fbae8b15a5a487fc3e7c14a80dcb4984b04a`, the reviewed surfaces include flat/HNSW `VectorDB`, `@ruvector/wasm`, a GNN reranking research surface, and an optional `@ruvector/mincut-wasm` wrapper with exact TypeScript fallback, dynamic cut and directed routing. The new MinCut work is a graph-topology capability, not evidence that vector retrieval or GNN learning improves this entry point.
+
+This run has 10 catalog projects plus 36 reviewed changes: 46 sanitized public records, below the frozen 100-record gate. Exact filtering, snapshot-bound paging and a deterministic digest remain the baseline. Persistent vector memory, WASM acceleration, GNN training and persistent learning were not implemented or claimed. Artifact size and latency were not measured because the candidate rejected those dependencies before adoption. The future experiment remains exact filters versus vector-only, fusion, GNN reranking and a shuffled-graph control, requiring a 10-point held-out task-success gain.
+
+## Forecast
+
+Prediction, 82% confidence by 2026-12-16: agent evidence and memory interfaces will increasingly expose content-addressed snapshot or receipt identifiers so caching, replay and audit can fail closed across state transitions. It passes only if at least two primary implementations newly publish digest-, receipt- or revision-bound metadata used for cache invalidation or replay verification.
+
+Established signals are narrower. [MCP caching guidance](https://modelcontextprotocol.io/specification/2026-07-28/server/utilities/caching) defines cache validation semantics. [NovaFabric](https://arxiv.org/abs/2609.12582) reports signed run capsules and tamper rejection, but only 2/10 tool-using workloads completed. [MutMem](https://arxiv.org/abs/2608.02843) reports signed mutation transitions while explicitly not establishing content truth, and [MemMark](https://arxiv.org/abs/2605.25002) studies attribution surviving memory snapshots. The forecast goes beyond those established results. No previous forecast horizon expired today.
+
+Next hypothesis: add a verifiable publisher receipt only if a stable trust root and independent verification path can be pinned; an unsigned digest must not be relabeled as authenticity.