david6666666 · Oct 26, 2026
diff --git a/README.md b/README.md
@@ -47,6 +47,22 @@ Knowledge retrieval is read-only; candidate storage is optional and separately
 authorized. Automatic search and automatic knowledge promotion remain planned. Missing
 or incomplete evidence stays `INCONCLUSIVE`.
 
+An optional [offline harness-improvement path](docs/harness-improvement.md)
+now checks caller-prepared skill edits using the existing Skill Eval runner:
+
+```mermaid
+flowchart LR
+    A["H0 + INV / DOM skill snapshots"] --> B["Matched development probes"]
+    B --> C["Quality-first, cost-aware choice or H0"]
+    C --> D["Mandatory rerun"]
+    D -->|"eligible candidate"| E["One-shot holdout gate"]
+    E --> F["Bounded candidate or retain H0"]
+    D -->|"H0 / invalid evidence"| F
+```
+
+This path is verified with fixed-output fixtures. Real-model automatic editing,
+cross-task reuse and measured model savings remain unverified.
+
 ## Choose your starting path
 
 | You want to… | Start at |
@@ -137,6 +153,11 @@ Ready to measure your own model? [Connect your benchmark and quality checks →]
 The implemented [Harness × Model × Infra boundary](docs/architecture.md) keeps new
 campaigns out of the common AB/BA scheduler, process cleanup and evaluator.
 
+The [bounded skill policy](docs/harness-improvement.md#implemented-architecture)
+sits above Skill Eval, reuses its comparison runner and independent verifier,
+and records development experience separately from executor-facing skills.
+Engine performance claims continue through the engine AB/BA lane.
+
 ![InferRSI system architecture: inputs, experiment loop and evidence](docs/assets/system-architecture-en.svg)
 
 [English SVG](docs/assets/system-architecture-en.svg) ·

diff --git a/docs/architecture.md b/docs/architecture.md
@@ -83,6 +83,8 @@ src/inferrsi/
     runner.py               One AB/BA loop; execute/snapshot adapter hooks
     evaluation.py           Performance/quality/source-integrity adjudication
     reporting.py            HTML evidence report
+    skill_evaluation.py     Skill trials, usage summaries and shared trial budget
+    improvement.py          Offline quality/cost selection policy; no model execution
     scenarios/
       external.py           External benchmark/validator commands; metric-reader hook
       media.py              Warmup seeds, first-response retention, sample/artifact groups
@@ -116,6 +118,23 @@ examples/                   Current Qwen runner, image validator and v4 adapter
 tests/fixtures/legacy/      Frozen historical validation dependencies for regressions
 ```
 
+### Bounded skill improvement
+
+[`tools/skill_eval_improve.py`](../tools/skill_eval_improve.py) composes the existing
+Skill Eval comparison tool for caller-prepared H0/INV/DOM snapshots. It freezes
+the supported audit inputs and policy, shares one outer budget across probes,
+mandatory rerun and one-shot holdout, and revalidates retained evidence when
+reporting. The Harness policy module owns quality gates and cost selection;
+the existing comparison runner and Infra executor still own trials and cleanup.
+See the [implemented flow and offline limits](harness-improvement.md).
+
+This is a separate optional skill-evaluation path. It neither replaces engine
+AB/BA adjudication nor automatically writes/promotes knowledge. The initial
+executor accepts only explicitly supplied trusted offline fixtures; automatic
+model edits, external sandbox qualification and cross-task reuse remain future
+work. Shared total-token and command-count summaries also serve the original
+pilot, configured comparison and PR integration consumers.
+
 ## Where to extend
 
 | Add | Main extension locations | Required qualification |

diff --git a/docs/examples/harness-improvement/README.md b/docs/examples/harness-improvement/README.md
@@ -0,0 +1,42 @@
+# Recorded offline controller example
+
+These are fixed-output software controls for
+[issue #93](https://github.com/InferMatrixLab/InferRSI/issues/93), generated with
+the adjacent implementation and `tests.fixtures.skill_eval_improvement_demo`.
+Python 3.12.3 and Git 2.47.3 were used on Linux. No real model was invoked.
+Full raw trial artifacts are produced locally by the reproduction command;
+these public JSON files retain the aggregate report without private paths.
+
+| Control | Actual subprocess calls | Scripted total tokens | Scripted held-out candidate checks | Regressions |
+| --- | ---: | ---: | ---: | ---: |
+| Paired H0/H0 | 12 | 1,440 | 6/6 | 0 |
+| Repeated H0, same call budget as the loop | 48 | 5,760 | 24/24 | 0 |
+| Bounded INV/DOM → rerun → holdout | 48 | 4,200 | 6/6 | 0 |
+
+The cheaper DOM fixture costs 240 scripted tokens over six development trials
+but has three regressions; the controller rejects it. INV costs 360 versus its
+paired H0's 720, passes the mandatory rerun and then the held-out checks. The
+full loop records 4,200, not just the selected candidate's 360. These values
+are deliberately programmed inputs to test quality-first selection and complete
+accounting. **They do not show a measured percentage saving or success-rate
+improvement for InferRSI.** The repeated-H0 control matches call count, not billed
+cost or a hard token budget, and has no best-of-N answer selection.
+
+The [summary](offline-summary.json) separates preparation and local pipeline
+wall time. A single local timing observation is not a speedup benchmark; these
+fixtures were also run alongside repository tests. The
+[bounded report](bounded-report.json) retains phase costs, selection reasons,
+rerun/holdout assessments and unknown proposal/monetary cost. Cached-input tokens
+are recorded separately and are not added to total tokens twice.
+
+Reproduce from the repository checkout with development dependencies:
+
+```bash
+python -m tests.fixtures.skill_eval_improvement_demo --output <new-output-dir>
+```
+
+The generated JSON includes retained-plan/report digests for that invocation.
+Temporary Git commit identities, file digests and wall times can differ between
+runs. The scripted call/token/outcome values above are reproducible. Model
+efficacy, complete authoring-plus-execution cost and engine benefits remain
+`INCONCLUSIVE`; this example is not a campaign `PASS`.
diff --git a/docs/examples/harness-improvement/bounded-report.json b/docs/examples/harness-improvement/bounded-report.json
@@ -0,0 +1,131 @@
+{
+  "plan_sha256": "81e4d4e425ea4eb58ea6dd7f145e4fc17df3bc37197cd7d3e1e430cfd6e82031",
+  "selection": {
+    "selected": "INV",
+    "reason": "quality_then_cost",
+    "assessments": {
+      "INV": {
+        "eligible": true,
+        "reason": null,
+        "success_rate": 1.0,
+        "baseline_success_rate": 1.0,
+        "total_tokens": 360,
+        "baseline_total_tokens": 720,
+        "regressions": 0
+      },
+      "DOM": {
+        "eligible": false,
+        "reason": "quality_floor",
+        "success_rate": 0.5,
+        "baseline_success_rate": 1.0,
+        "total_tokens": 240,
+        "baseline_total_tokens": 720,
+        "regressions": 3
+      }
+    },
+    "pareto": [
+      "INV"
+    ],
+    "scores": {
+      "INV": 0.0
+    },
+    "plan_sha256": "81e4d4e425ea4eb58ea6dd7f145e4fc17df3bc37197cd7d3e1e430cfd6e82031"
+  },
+  "retained_selection_consistent": true,
+  "rerun": {
+    "eligible": true,
+    "reason": null,
+    "success_rate": 1.0,
+    "baseline_success_rate": 1.0,
+    "total_tokens": 360,
+    "baseline_total_tokens": 720,
+    "regressions": 0
+  },
+  "holdout": {
+    "eligible": true,
+    "reason": null,
+    "success_rate": 1.0,
+    "baseline_success_rate": 1.0,
+    "total_tokens": 360,
+    "baseline_total_tokens": 720,
+    "regressions": 0
+  },
+  "accepted": true,
+  "effective_candidate": "INV",
+  "phase_usage": {
+    "probe-INV": {
+      "calls": 12,
+      "input_tokens": 960,
+      "output_tokens": 120,
+      "cached_input_tokens": 240,
+      "total_tokens": 1080,
+      "command_calls": 24,
+      "elapsed_seconds": 0.20449978415854275
+    },
+    "probe-DOM": {
+      "calls": 12,
+      "input_tokens": 840,
+      "output_tokens": 120,
+      "cached_input_tokens": 240,
+      "total_tokens": 960,
+      "command_calls": 24,
+      "elapsed_seconds": 0.1887140409089625
+    },
+    "rerun-INV": {
+      "calls": 12,
+      "input_tokens": 960,
+      "output_tokens": 120,
+      "cached_input_tokens": 240,
+      "total_tokens": 1080,
+      "command_calls": 24,
+      "elapsed_seconds": 0.1887151626870036
+    },
+    "holdout-INV": {
+      "calls": 12,
+      "input_tokens": 960,
+      "output_tokens": 120,
+      "cached_input_tokens": 240,
+      "total_tokens": 1080,
+      "command_calls": 24,
+      "elapsed_seconds": 0.18848644592799246
+    }
+  },
+  "observed_execution": {
+    "calls": 48,
+    "input_tokens": 3720,
+    "output_tokens": 480,
+    "cached_input_tokens": 960,
+    "total_tokens": 4200,
+    "command_calls": 96,
+    "elapsed_seconds": 2.1600983038078994,
+    "cost": null
+  },
+  "budget": {
+    "calls": 48,
+    "observed_tokens": 4200,
+    "usage_complete": true,
+    "stop_reason": "call_budget"
+  },
+  "proposal_cost": null,
+  "proposal_cost_note": "Caller-prepared snapshots; authoring cost is unknown and excluded.",
+  "selected_execution": {
+    "trials": 6,
+    "verdicts": {
+      "PASS": 6,
+      "FAIL": 0,
+      "INCONCLUSIVE": 0
+    },
+    "evaluation_errors": 0,
+    "elapsed_seconds": 0.09431080287322402,
+    "input_tokens": 300,
+    "output_tokens": 60,
+    "cached_input_tokens": 120,
+    "total_tokens": 360,
+    "command_calls": 12,
+    "cost": null,
+    "expected_trials": 6,
+    "missing_trials": 0
+  },
+  "promotion": "not_evaluated",
+  "claim": "Offline control-flow evidence only. Fixture tokens are scripted, not billed. No measured model efficacy, end-to-end savings, sandbox isolation or engine speedup."
+}
diff --git a/docs/examples/harness-improvement/offline-summary.json b/docs/examples/harness-improvement/offline-summary.json
@@ -0,0 +1,65 @@
+{
+  "kind": "offline_fixture_control",
+  "rows": [
+    {
+      "condition": "H0",
+      "preparation_seconds": 0.05566692119464278,
+      "execution_seconds": 0.5605955601204187,
+      "calls": 12,
+      "scripted_total_tokens": 1440,
+      "scripted_command_calls": 24,
+      "holdout_candidate_trials": 6,
+      "holdout_candidate_verdicts": {
+        "PASS": 6,
+        "FAIL": 0,
+        "INCONCLUSIVE": 0
+      },
+      "holdout_regressions": 0,
+      "effective_candidate": "H0",
+      "plan_sha256": "bd0d7c0062dfe8bf244b46cb8b7423e4db65d08c651f5207a9dbf07f446f79ad",
+      "report_sha256": "9c1c68988dc512a3bc95b0004de5ba7d0bc3f9dfd689ca46675592a1a8f00fb2"
+    },
+    {
+      "condition": "repeated-H0",
+      "preparation_seconds": 0.04197078011929989,
+      "execution_seconds": 1.9998727978672832,
+      "calls": 48,
+      "scripted_total_tokens": 5760,
+      "scripted_command_calls": 96,
+      "holdout_candidate_trials": 24,
+      "holdout_candidate_verdicts": {
+        "PASS": 24,
+        "FAIL": 0,
+        "INCONCLUSIVE": 0
+      },
+      "holdout_regressions": 0,
+      "effective_candidate": "H0",
+      "plan_sha256": "6ebec22ce01f12d61ec07c00711f3c190206b0830d92345adbdceb4a42789f21",
+      "report_sha256": "ba0866d1d2f145117eb122a1c844c2ecc5ee0eb6e3b79e3231e3263a8a2f0c2a"
+    },
+    {
+      "condition": "bounded",
+      "preparation_seconds": 0.1715290721040219,
+      "execution_seconds": 2.1994067209307104,
+      "calls": 48,
+      "scripted_total_tokens": 4200,
+      "scripted_command_calls": 96,
+      "holdout_candidate_trials": 6,
+      "holdout_candidate_verdicts": {
+        "PASS": 6,
+        "FAIL": 0,
+        "INCONCLUSIVE": 0
+      },
+      "holdout_regressions": 0,
+      "effective_candidate": "INV",
+      "plan_sha256": "81e4d4e425ea4eb58ea6dd7f145e4fc17df3bc37197cd7d3e1e430cfd6e82031",
+      "report_sha256": "b0e3d7f1d17708a737687b6aca07838f42a45919071343e9e74634812b3bb04b"
+    }
+  ],
+  "scope": "Actual subprocess calls and pipeline wall time; scripted answers and tokens. H0 is a paired H0/H0 control, repeated-H0 has the bounded loop's 48-call budget. No best-of-N answer selection. Different repeat counts are not efficacy estimates.",
+  "model_calls": 0,
+  "monetary_cost": null,
+  "proposal_authoring_cost": null,
+  "efficacy_verdict": "INCONCLUSIVE",
+  "engine_performance_verdict": "INCONCLUSIVE"
+}