Research: where the Harness Knowledge Graph lands, and what structure fits instead

AGENTS.md

@@ -60,7 +60,9 @@ each tier is a superset of the confidence of the one above it, so a deep change
 runs all three.
 
 **1. cheap — always, before every commit that touches `plugins/**` or `evals/**`.**
-Deterministic, offline, free, under a second:
+Deterministic, offline, free. ~14s for 1296 checks across 25 plugins, measured
+2026-09-18 — it scales with the plugin count, so re-measure rather than trusting
+this number:
 
 ```sh
 evals/cheap/run.sh      # exit 0 required to commit

docs/research/agentic-patterns-corpus.md

@@ -94,7 +94,7 @@ Adoption alone is not fit. These were rejected with reasons:
 - Per-turn predictive model routing (GPT-5-style routers) — universal in consumer products, but a mis-route has no recovery path and per-turn switching destroys cache affinity (up to 12.5x prefix swing). Pin model and effort at BEGIN, hold to END
 - Semantic/vector response caching — widely marketed, but the coding-agent adoption claim has no primary source and agent steps are not repeated queries; the real practice is prefix/KV stability, already absorbed by criteria-pin
 - TDD-Agent dual-track refinement — test-first is confirmed and already Red Gate's BEGIN, but letting the agent refine the tests alongside the code is the exact reward-hacking mutation control exists to block. Name the rejection in the protocol
-- Graph/knowledge-graph memory at round level — 3-5x the cost of flat RAG and needs a hand-built ontology; a repo already has git, grep and a type checker as a better graph. Keep it out of rounds entirely
+- Graph/knowledge-graph memory at round level — 3-5x the cost of flat RAG and needs a hand-built ontology; a repo already has git, grep and a type checker as a better graph. Keep it out of rounds entirely [Revisited 2026-09-10 in [`harness-knowledge-graph.md`](harness-knowledge-graph.md): upheld, and now the load-bearing reason — but scoped to git-native repo data, not universal.]
 - Self-modifying agent archives (Darwin Gödel Machine) — adopt the two invariants it proves by violating them, never the mechanism; a harness that can rewrite its own checker has no gate
 - Concurrent fast-path/CoT racing — an inference-layer technique with no agent-orchestration deployments; racing two MIDDLE writers breaks single-writer and doubles cost to shave seconds off a human-gated round
 
@@ -677,12 +677,12 @@ Scout claims that did not survive verification are called out in the implication
 **Knowledge-graph + agentic RAG (GRAG-ProSafe class)**  
 *Mechanism:* Four-stage LLM extraction turns unstructured reports into a dynamic knowledge graph, then multi-hop retrieval plus chain-of-thought reasoning answers causal questions over it. GRAG-ProSafe built 1637 nodes / 2285 edges from 198 iron-and-steel accident reports, scoring 0.868 faithfulness, 0.824 answer relevancy, 0.805 factual correctness.  
 *Why leaders use it:* Multi-hop causal questions that flat vector RAG cannot answer in knowledge-dense, audit-bound domains such as industrial safety and root-cause analysis.  
-*Failure mode:* Adoption evidence does not hold up: this is a single Expert Systems with Applications paper on one 198-document corpus, not deployed production practice across leaders.  
+*Failure mode:* Adoption evidence does not hold up: this is a single Expert Systems with Applications paper on one 198-document corpus, not deployed production practice across leaders. [Revisited 2026-09-10 in [`harness-knowledge-graph.md`](harness-knowledge-graph.md): this evidence-quality objection no longer holds — Harness ships a production software-delivery knowledge graph at enterprise scale. The domain-fit objection in the rejected list stands.]  
 *Red Gate fit:* Should NOT enter Red Gate's loop — graph construction cost dwarfs the payoff at 24-skill scale. The only plausible use is DETECT recurrence over accumulated exhaust, and a flat index over dev-diary entries reaches that far more cheaply.  
 *Sources:* https://www.sciencedirect.com/science/article/abs/pii/S0957417425035626
 
 **Implications:**
-- Nothing here is outright vapor, but two are demoted. Knowledge-graph agentic RAG rests on one 198-document academic system, not leader adoption — treat as research, not roadmap. Circuit breakers are a blog-sourced restatement of what stop-rule and the budget pool already do.
+- Nothing here is outright vapor, but two are demoted. Knowledge-graph agentic RAG rests on one 198-document academic system, not leader adoption — treat as research, not roadmap. [Revisited 2026-09-10 in [`harness-knowledge-graph.md`](harness-knowledge-graph.md): this evidence-quality objection no longer holds — Harness ships a production software-delivery knowledge graph at enterprise scale. The domain-fit objection in the rejected list stands.] Circuit breakers are a blog-sourced restatement of what stop-rule and the budget pool already do.
 - Three scout citations do not hold up and were corrected: AgentTrace is arXiv 2602.10133 (not 2604.26152); STORM is a May 2026 research system, not a shipping OpenAI Agents SDK feature — the scout conflated it with SDK handoffs; and the NVIDIA/Gretel price was reported as nine figures above a $320M valuation, terms undisclosed.
 - The two highest-value absorptions are structural, not additive. (1) Make the between-round human gate a durable checkpoint and adopt LangGraph's replay discipline — read-only before the gate, writes after — so a resumed round cannot double-apply side effects. (2) Turn EMIT exhaust into OTel-shaped spans keyed to the pinned verifier id, so CONSOLIDATE and DETECT operate on structured traces instead of diary prose.
 - Cost is a harness property, not a model choice: 41% blended cost and 38% token reduction came from swapping orchestration alone. Red Gate's pointer envelopes should be governed by an explicit cache contract — stable prefix for criteria and skill text, 4 breakpoints budgeted, and an acknowledgement that human gates exceed the 5-minute TTL.