docs: publish verified Graphify native-memory comparison

README.md

@@ -17,6 +17,12 @@ The most robust result is retrieval, which is measured without an agent and so i
 file the gold patch edits reaches the agent in **96.3%** of sessions versus **81.5%** for the
 closest comparable tool, at equal payload bytes and with fewer search calls.
 
+In a separate full conversational-memory comparison, Entire Graph leads Graphify on LOCOMO
+recall@10 (**0.918 vs 0.844**) but trails on final-answer accuracy: LOCOMO **51.0% vs 67.4%** and
+LongMemEval-S **44.7% vs 62.7%**. That split matters: better evidence-session retrieval did not
+translate into better answers under the shared reader. See [Numbers, honestly](#numbers-honestly)
+for the exact public-protocol boundary.
+
 Read [Numbers, honestly](#numbers-honestly) before quoting any of that. In short: the figure is
 Haiku-specific and we **cannot** measure it on stronger models at these sample sizes; a **≥35%**
 saving is *refuted*, not merely unproven; two earlier claims (55%, then 31.6%) were withdrawn after
@@ -144,6 +150,42 @@ You (a human) will mostly experience it *through* those surfaces. Your agents ca
 
 ## Numbers, honestly
 
+### Native conversational-memory retrieval
+
+The frozen candidate at `dea450b` was evaluated against Graphify v8/v0.9.34 at `07b9143` and
+Codebase Memory MCP v0.9.0 on the same official 300 LOCOMO and 50 cleaned LongMemEval-S cases,
+three repetitions per case. Kimi K3 was the shared reader and blinded primary grader; Opus 5
+audited a deterministic 20%. Every arm had top 10 and a 128,000-byte ceiling, separate caches,
+identical questions, and no shared source reread.
+
+| Full native-memory run | Entire Graph | Graphify | cmm |
+|---|---:|---:|---:|
+| LOCOMO recall@10 (n=300) | **0.918** | 0.844 | 0.000 |
+| LOCOMO QA accuracy (n=300) | 51.0% | **67.4%** | 22.3% |
+| LongMemEval-S QA accuracy (n=50) | 44.7% | **62.7%** | 6.0% |
+| graph-build LLM credits | 0 | 0 | 0 |
+
+There is no blanket winner. Entire Graph retrieves the evidence session more often; Graphify
+answers more questions correctly. All 3,150 raw cells, 3,150 Kimi grades, and 630 Opus audits
+passed the sealed gates with zero invalid attempts; audit agreement was 97.94% (kappa 0.9583).
+The list-price equivalent for reader, grading, and audit was $50.82. Provider-billed spend was not
+available for every Fireworks cell, so the smaller known-actual subtotal is not reported as total
+spend.
+
+This is a **public-protocol reimplementation**, not a reproduction of Graphify's historical
+claims. Graphify's advertised memory harness and selectors are not public. The Graphify reader
+received its native BFS-rendered graph text; a separate projection of the same traversal onto
+session files supplied recall@10. Entire Graph supplied only native search snippets and locator
+headers. cmm's shipped Markdown search excludes `Section` nodes from BM25, so its result describes
+that public Markdown path, not its quality on programming-language graphs.
+
+The full protocol, fairness explanation, exact artifact hashes, 1%/10%/100% sequence, and results
+are in [GraphMark PR #63](https://github.com/entirehq/graphmark/pull/63). The benchmarked binary is
+from `dea450b`; these documentation changes do not change it.
+
+See [docs/benchmarks.md](docs/benchmarks.md#external-native-memory-comparison) for the frozen
+identities and integrity record.
+
 All savings are measured against the **baseline**: the same agent, same model, same task, with
 no code tool at all. For scale, the same suites also ran
 [codebase-memory-mcp](https://github.com/DeusData/codebase-memory-mcp) ("cmm"), an open-source

cmd/graph-bench/main_test.go

@@ -6,6 +6,7 @@ import (
 	"encoding/json"
 	"os"
 	"path/filepath"
+	"runtime"
 	"strings"
 	"testing"
 	"time"
@@ -55,6 +56,9 @@ func TestValidateExecutionModeRejectsParentOnlyCPUProfile(t *testing.T) {
 }
 
 func TestMaxRSSGuardFailsRunEvenWhenViolatingRowIsExcludedFromAggregates(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("process RSS is not available on this platform")
+	}
 	dir := t.TempDir()
 	manifestPath := filepath.Join(dir, "manifest.json")
 	if err := os.WriteFile(manifestPath, []byte(`{"languages":{"Go":["owner/repo"]}}`), 0o644); err != nil {

docs/benchmarks.md

@@ -171,6 +171,56 @@ Treat the numbers as historical; re-run with the current streaming benchmark
   tiny constant per relation rather than the full payload, which is what kept
   the in-memory path from finishing on the largest repos.
 
+## External native-memory comparison
+
+GraphMark's `memory-native-v2` suite evaluated this repository's candidate
+commit `dea450b924e0ba4e9b5a7dc3ae5db72cf3aa857a` alongside Graphify and
+Codebase Memory MCP. This is an external public-protocol reimplementation of
+Graphify's advertised memory surfaces, not a `cmd/graph-bench` run and not a
+reproduction of Graphify's unpublished harness.
+
+Frozen product identities:
+
+- Entire Graph candidate: `dea450b924e0ba4e9b5a7dc3ae5db72cf3aa857a`
+- Entire Graph stable reference: `90a3346a624d76f7fee21bd894721e5438dd9ac2`
+- Graphify v8/v0.9.34: `07b9143d4b90b1e1cb88dc71423f742a501efd29`
+- Codebase Memory MCP v0.9.0: `b637e3330c96cfe452da623db068c241aaa3ec01`
+- GraphMark harness: `09d14f2f149e8459d98b30ecf6bf1a31c757f04d`
+- protocol SHA-256: `524e64ce2b0e8c888169d68d990466bed71336e2d7cd30b2723a78ba679fdb1b`
+- full completion SHA-256: `30c8c19c24d7077abc28b30bb5853391b3f86700801f5e9a371acec0c1938a76`
+
+The full run used 300 LOCOMO and 50 LongMemEval-S cases, three repetitions,
+Kimi K3 reader/grader, a deterministic 20% Opus 5 audit, top 10, and a shared
+128,000-byte context ceiling. It followed a sealed 1% validation and untouched
+10% validity holdout; the full run was outcome-independent. Each product has
+exactly 1,050 raw cells with identical case/question/repetition topology.
+
+Entire Graph receives no benchmark-specific query rewrite. For each original
+question the adapter executes `entire graph search --format json --top-k 10`
+and gives the shared reader only the returned native snippets and neutral
+locator headers. It does not open surrounding files after search. The build and
+query cache is isolated from every other arm.
+
+The verified full result is:
+
+| Metric | Entire Graph | Graphify | cmm |
+|---|---:|---:|---:|
+| LOCOMO recall@10 | **0.918** | 0.844 | 0.000 |
+| LOCOMO QA accuracy | 51.0% | **67.4%** | 22.3% |
+| LongMemEval-S QA accuracy | 44.7% | **62.7%** | 6.0% |
+| graph-build LLM credits | 0 | 0 | 0 |
+
+Integrity record: 3,150/3,150/630 raw/grade/audit cells, zero invalid
+attempts, promotion passed with zero findings, final artifact verification
+passed, and Opus agreement was 97.94% (kappa 0.9583). Five transient raw-reader
+timeouts and two transient grader timeouts were recorded and recovered within
+the sealed retry policy; no audit retry occurred.
+
+The publication bundle and review are in
+[entirehq/graphmark#63](https://github.com/entirehq/graphmark/pull/63). Raw
+artifacts remain outside Git because they occupy about 420 MB; the bundle
+publishes their SHA-256 values and the exact reproduction contract.
+
 ## Notes
 
 - A full-tier run is heavy (the manifest includes linux, tensorflow, vscode,