ruvnet.
Python
docs(ruvnet): record 2026-09-14 improvement evidence
rUv39409df8f7d1
diff --git a/docs/self-improvement/2026-09-14-evidence.json b/docs/self-improvement/2026-09-14-evidence.jsonnew file mode 100644index 0000000..fe9c678--- /dev/null+++ b/docs/self-improvement/2026-09-14-evidence.json@@ -0,0 +1,45 @@+{+ "schemaVersion": 1,+ "date": "2026-09-14",+ "decision": "accept",+ "baseline": "270fd513548e93862d59af8f22954c72686fb191",+ "candidateCommit": "3d8635856d62f1f0d0ca483fdc0a3e4899870937",+ "candidate": "bounded-upstream-snapshot-freshness",+ "version": "0.2.2",+ "upstream": {+ "observedAt": "2026-09-14T10:17:06Z",+ "sources": [+ {"project":"ruflo","revision":"1992ffb020fd461a7c2a3bb64f30476a0c5d2d26","license":"MIT","changedSincePrior":true,"incorporated":"source-head provenance only; the change was telemetry data, not a product capability"},+ {"project":"metaharness","revision":"d5833dc6512ac1adeeef91a331c29055cd8a4dbb","license":"MIT","changedSincePrior":false},+ {"project":"ruvector","revision":"edaffffb3b85768eb1f3ec1f683b7f46f0506af4","license":"MIT","changedSincePrior":false},+ {"project":"agentdb","revision":"6cdefe451f6afe7395bc88d3587a0225d161089b","license":"MIT","changedSincePrior":false}+ ],+ "xRuvIo":{"version":"0.7.2","routes":["/mcp","/chatgpt/mcp"]}+ },+ "heldOut":{"baseline":{"passed":1,"total":6,"note":"The invalid-timestamp check threw only because the API was absent."},"candidate":{"passed":6,"total":6}},+ "tests":{"groupsPassed":9,"groupsFailed":0},+ "validation":{"plugin":true,"skillsPassed":4,"skillsFailed":0,"marketplaceSchemas":2,"packageFiles":18,"compressedBytes":17560,"unpackedBytes":51600},+ "benchmarks": {+ "raw":"2026-09-14-benchmarks.json",+ "pairs":15,+ "baseline":{"discoveryP95MedianMs":0.0016450000000105547,"coldCliP95Ms":116.461063,"rssMedianBytes":46997504},+ "candidate":{"discoveryP95MedianMs":0.0015920000000022583,"coldCliP95Ms":102.03414899999999,"rssMedianBytes":47198208},+ "delta":{"discoveryP95Percent":-3.2218844989639117,"coldCliP95Percent":-12.387757443017676,"rssBytes":200704},+ "interpretation":"regression gates passed; timing deltas are not claimed as speed gains"+ },+ "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,"totalDependenciesReported":94},+ "secretScan":{"findings":0,"filesReportedByScanner":5},+ "manualTrackedFileCredentialPatternHits":0,+ "threatScan":{"indicators":0,"filesReportedByScanner":4},+ "limitations":"Secret and STRIDE scanners reported narrow file coverage. Results are not proof of complete security."+ },+ "memoryExperiment":{"catalogProjects":10,"changeRecords":6,"ruvectorWasm":"deferred","gnn":"rejected for this workload","persistentLearning":"not implemented","reason":"Corpus remains below the frozen 100-record minimum.","futureGate":{"minimumRecords":100,"minimumTaskSuccessGainPercentagePoints":10,"p95MaximumMs":20,"memoryMaximumBytes":134217728}},+ "forecast":{"prediction":"Interfaces will increasingly surface freshness or invalidation state for cached evidence.","confidencePercent":80,"horizonDate":"2026-12-13","criterion":"At least two interoperable agent-memory protocols or major hosted connector interfaces expose freshness, expiry, or invalidation fields to clients.","sources":["https://arxiv.org/abs/2605.23986","https://arxiv.org/abs/2608.04746","https://blog.modelcontextprotocol.io/posts/2026-07-28/"]},+ "ci":{"run":"https://github.com/ruvnet/ruvnet/actions/runs/34833003257","status":"success","node22":"success","node24":"success"},+ "dashboard":"https://ruvnet-harness-evidence.ruv.chatgpt.site"+}diff --git a/docs/self-improvement/2026-09-14.md b/docs/self-improvement/2026-09-14.mdnew file mode 100644index 0000000..40c3fce--- /dev/null+++ b/docs/self-improvement/2026-09-14.md@@ -0,0 +1,44 @@+# RuVnet harness improvement report · 2026-09-14++## Decision++Accepted bounded upstream-snapshot freshness and published version 0.2.2. `ruvnet changes`, `ruvnet_changes`, and `ruv://changes` now expose `current`, `stale`, or `clock-skew`, an age, a 72-hour bound, and an expiry. Stale or future-dated evidence produces an explicit warning. The companion remains offline, read only, and incapable of refreshing itself.++## Upstream review++| Project | Observed revision | Result |+| --- | --- | --- |+| Ruflo | [`1992ffb`](https://github.com/ruvnet/ruflo/commit/1992ffb020fd461a7c2a3bb64f30476a0c5d2d26) | Head advanced only for a clone-telemetry snapshot. Its files were reviewed but no telemetry behavior was incorporated; only the exact source head and observation time were refreshed. |+| MetaHarness | [`d5833dc`](https://github.com/ruvnet/metaharness/commit/d5833dc6512ac1adeeef91a331c29055cd8a4dbb) | No source change since the prior run. MetaHarness 0.4.16 remained the pinned evaluator. |+| RuVector | [`edaffff`](https://github.com/ruvnet/RuVector/commit/edaffffb3b85768eb1f3ec1f683b7f46f0506af4) | No source change. Documented vector, graph, feedback, WASM and GNN surfaces remain candidates, not capabilities of this companion. |+| AgentDB | [`6cdefe4`](https://github.com/ruvnet/agentdb/commit/6cdefe451f6afe7395bc88d3587a0225d161089b) | No source change. |++The four recorded repositories retain MIT license files at their observed heads. The current x.ruv.io service continued to advertise Ruflo 0.7.2 and separate `/mcp` and `/chatgpt/mcp` routes. Fetched repository and service text was treated as untrusted evidence, not instructions.++## MetaHarness loop and measured evidence++The [acceptance plan](2026-09-14-plan.json) was frozen before implementation. Nominal held-out results improved from 1/6 to 6/6; the baseline's lone pass was only the invalid-timestamp check throwing because the function did not exist. Nine test groups, manifest/schema validation, package closure, CLI, in-memory MCP and stdio MCP checks passed. MetaHarness 0.4.16 reported 6/6 hard constraints for baseline and candidate, with unchanged heuristic scores. Those scores are not proof of improvement.++Fifteen alternating baseline/candidate pairs measured 10,000 discovery calls per run and one cold CLI start per run. Median discovery p95 changed from 0.001645 ms to 0.001592 ms (-3.22%); cold CLI p95 changed from 116.461 ms to 102.034 ms (-12.39%); median resident memory increased 200,704 bytes. These clear the frozen regression gates, but the timing differences are treated as noise rather than speed gains. [Raw benchmarks](2026-09-14-benchmarks.json) preserve every sample.++The deep Ruflo 3.25.6 code and dependency scans plus npm audit reported zero findings. Manual tracked-file credential patterns returned zero files. The secret scanner reported five files and STRIDE four, so coverage is limited and the result is not a claim of complete security. Residual risks include a fixed policy that may not fit every source, hosts treating MCP annotations as enforcement, and local system clocks affecting the age calculation.++[Candidate commit](https://github.com/ruvnet/ruvnet/commit/3d8635856d62f1f0d0ca483fdc0a3e4899870937) passed [Node 22 and 24 CI](https://github.com/ruvnet/ruvnet/actions/runs/34833003257).++## Memory, WASM and GNN decision++No vector database, GNN training, persistent learning, or WASM acceleration was added or claimed. The eligible sanitized corpus remains ten projects and six change records, below the precommitted 100-record minimum. Exact lookup remains deterministic and materially simpler. A future experiment remains eligible only with a leakage-free split, at least 10 percentage points of task-success gain over exact lookup, p95 below 20 ms, and total memory below 128 MB.++## Forecast++Established results: MemForest reports lower memory-freshness latency through hierarchical temporal indexing, while ScrubJay-MEM reports improved temporal generalization from type-conditioned decay and also reports that gains reverse on fact-consolidation tasks. These are scoped preprint results, not validation of RuVnet's fixed 72-hour provenance policy. [MemForest](https://arxiv.org/abs/2605.23986) · [ScrubJay-MEM](https://arxiv.org/abs/2608.04746).++Prediction, 90 days, 80% confidence: host and memory interfaces will increasingly surface freshness or invalidation state so cached evidence cannot silently influence action beyond its useful horizon. This passes if, by 2026-12-13, at least two interoperable agent-memory protocols or major hosted connector interfaces expose freshness, expiry, or invalidation fields to clients. The MCP 2026-07-28 cache hints make this plausible but do not yet prove the prediction. [MCP release](https://blog.modelcontextprotocol.io/posts/2026-07-28/).++The prior 30-day host-profile, 90-day provenance/cache-governance, and 180-day memory-routing forecasts have not reached their horizons. Today's change is implementation evidence for the provenance forecast, not proof of host enforcement.++## Next hypothesis++Replace the global 72-hour default with source-class-specific policy only if a held-out stale-decision dataset shows fewer wrong routing decisions with no increase in false-stale warnings. Until such a dataset exists, keep the simpler deterministic bound.++[Private evidence dashboard](https://ruvnet-harness-evidence.ruv.chatgpt.site) · [Machine-readable evidence](2026-09-14-evidence.json)