B3: register autofit_inference as instance fit; gaussian_x3 campaign and wave-1 task

badge.json

@@ -1,6 +1,6 @@
 {
   "schemaVersion": 1,
   "label": "insight",
-  "message": "1 projects \u00b7 56 records \u00b7 0 cached",
+  "message": "2 projects \u00b7 56 records \u00b7 0 cached",
   "color": "yellow"
 }

campaigns.yaml

@@ -89,6 +89,20 @@ campaigns:
   job_status: unknown — live scheduler not queried; historical ledger states are not current job evidence
   evidence: https://github.com/PyAutoLabs/PyAutoCortex/blob/d45f198804615c090e6b6694cf725ddb0be2354e/projects/ic50_workspace.md
   project: ic50_workspace
+- id: gaussian-x3-search-benchmark
+  title: PyAutoFit search benchmark on gaussian_x3 (autofit_inference)
+  status: blocked
+  updated: '2026-10-08'
+  next: Run the 297 deferred wave-1 runs (and the JAX blend reference seed 1) on RAL partition ral on a PyAutoFit
+    revision after A2, then freeze gaussian_x3@2 before any wave-2 row.
+  recent_progress: Wave-1 pilot stopped at 223 of 520 expected runs by the B3 wrap-up ruling (2026-10-08);
+    pilot ranks nothing; catalogue, calibration record and inference-summary@1 published.
+  blockers: Phase A2 (PyAutoFit#1679, gradients under JAX) must merge first; NSS waits on A3b; no compute
+    submission is authorized by this ledger.
+  job_status: unknown — live scheduler not queried; the local pilot was stopped and this ledger records no
+    RAL submission
+  evidence: https://github.com/PyAutoLabs/autofit_inference/issues/4
+  project: autofit_inference
 tasks:
 - id: drop_1e3_prior_centring_after_631
   campaign: slam-hst-base
@@ -150,3 +164,13 @@ tasks:
   next: 'An unscoped idea, not permission to submit: reconcile point-source admission with imaging and weak-lensing
     candidate targets.'
   evidence: https://github.com/PyAutoLabs/PyAutoMind/blob/8dd1edc7e04c960daae3663513c61395823eddb0/ideas.md
+- id: gaussian_x3_search_wave1
+  campaign: gaussian-x3-search-benchmark
+  title: gaussian_x3 search benchmark — wave 1 (the pilot)
+  status: blocked
+  priority: normal
+  path: tasks/gaussian_x3_search_wave1.md
+  updated: '2026-10-08'
+  next: Pilot partial (223 of 520 runs); run the deferred cells on RAL partition ral after A2, then freeze
+    gaussian_x3@2.
+  evidence: https://github.com/PyAutoLabs/autofit_inference/issues/4

dashboard.md

@@ -1,16 +1,16 @@
 # PyAutoInsight — inference control room
 
-<!-- insight-inputs:f70051ff43253519801b4a7d148faab5708eee5feeb4f55c4f6efca3a0b2c865 -->
+<!-- insight-inputs:4d0d489a4e3ad725a1cce864a9cb930eb030864e01ae55fc468b4eb68df27e08 -->
 
 | Where | Count |
 |---|---:|
-| [Projects](#evidence) | 1 |
+| [Projects](#evidence) | 2 |
 | [Records](#evidence) | 56 |
 | [Comparisons](#evidence) | 0 |
 | [Cached](#evidence) | 0 |
 | [Failed](#evidence) | 0 |
 
-<!-- insight-campaigns:cebf9dfa86c60986e2ae3381ae314259f7e491233af2ec497ff330f710f3e901 -->
+<!-- insight-campaigns:b76456132d7495bee180722f43f437e05e473221388e08c82463a9dd7dbb4481 -->
 ## Check in on all inference work
 
 Copy this into one chat. Add a campaign focus or idea before or after it, or leave it unchanged.
@@ -46,6 +46,7 @@ After taking action, report what changed, what the evidence supports and what re
 | [N=50 NUTS versus three EP repeats](https://github.com/PyAutoLabs/PyAutoCortex/blob/d45f198804615c090e6b6694cf725ddb0be2354e/projects/ep_toy_gaussian.md) | Cortex records failed wave 1 (EMFILE/OOM), wave-2 preparation and no submission. | Present the existing wave-2 proposal with current dependency evidence for decision. |
 | [Hierarchical slope recovery at scale](https://github.com/PyAutoLabs/PyAutoCortex/blob/d45f198804615c090e6b6694cf725ddb0be2354e/projects/slope_hierarchy_scale.md) | Cortex retains graphical baseline, post-processing OOM and stale EP hierarchical factor; Mind records later moments fix. | Verify the moments fix release and project mirror; recover missing summaries from existing samples before a proposed rerun. |
 | [IC50 EP and graphical scale ladder](https://github.com/PyAutoLabs/PyAutoCortex/blob/d45f198804615c090e6b6694cf725ddb0be2354e/projects/ic50_workspace.md) | Cortex records N=5/10/25 results, failed N=50 EP rung and width-interpretation caveat. | Check implementation close-outs, then propose a seeded ladder with appropriate output and mean-field diagnostics. |
+| [PyAutoFit search benchmark on gaussian_x3 (autofit_inference)](https://github.com/PyAutoLabs/autofit_inference/issues/4) | Wave-1 pilot stopped at 223 of 520 expected runs by the B3 wrap-up ruling (2026-10-08); pilot ranks nothing; catalogue, calibration record and inference-summary@1 published. | Run the 297 deferred wave-1 runs (and the JAX blend reference seed 1) on RAL partition ral on a PyAutoFit revision after A2, then freeze gaussian_x3@2 before any wave-2 row. |
 
 ## Active tasks
 
@@ -56,6 +57,7 @@ After taking action, report what changed, what the evidence supports and what re
 | [Expectation Propagation Scale-Up — Scoping](https://github.com/PyAutoLabs/PyAutoInsight/blob/main/tasks/ep_scoping.md) | graphical-ep | needs-slicing | high | Reconcile completed wrapper work and Pulse profiling ownership before selecting a bounded remaining inference phase. |
 | [Graphical Model Scale-Up — Scoping](https://github.com/PyAutoLabs/PyAutoInsight/blob/main/tasks/graphical_scoping.md) | graphical-ep | needs-slicing | high | Reconcile current Cortex scale results and library dependencies; select one bounded inference phase with a witness. |
 | [NUTS/HMC trial on validated likelihoods](https://github.com/PyAutoLabs/PyAutoInsight/blob/main/tasks/gradient_sampler_validated_likelihoods.md) | point-source-admission | needs-decision | normal | An unscoped idea, not permission to submit: reconcile point-source admission with imaging and weak-lensing candidate targets. |
+| [gaussian_x3 search benchmark — wave 1 (the pilot)](https://github.com/PyAutoLabs/PyAutoInsight/blob/main/tasks/gaussian_x3_search_wave1.md) | gaussian-x3-search-benchmark | blocked | normal | Pilot partial (223 of 520 runs); run the deferred cells on RAL partition ral after A2, then freeze gaussian_x3@2. |
 
 [Ledger](https://github.com/PyAutoLabs/PyAutoInsight/blob/main/campaigns.yaml)
 
@@ -12126,6 +12128,27 @@ Archived rows belong to this project's September 2026 runs and are historical, n
 </details>
 
 
+<!-- insight:instance name=fit receipt=ec35be493712d693b8cab4d071d97d9a242137b6 outcome=ok shown=ec35be493712d693b8cab4d071d97d9a242137b6 -->
+## autofit_inference
+Integrity: ok. Freshness: freshness policy unspecified · evidence time unknown. Qualification: no measurements.
+Capture source branch: main; revision: ec35be493712d693b8cab4d071d97d9a242137b6; fetched 2026-10-08T19:29:28Z; latest attempt 2026-10-08T19:29:28Z.
+
+<details><summary>Full captured evidence and provenance</summary>
+
+
+Coverage: {"excluded": [], "expected": {"reason": "No expected-run manifest yet; wave 1 (B3) defines one", "runs": null}, "observed": {"records": 0, "runs": 0}}
+Verdicts are producer-asserted under protocol gaussian_x3@1, whose PLACEHOLDER thresholds are calibrated in the wave-1 pilot and then frozen
+Rows are judged only against a reference with the same dataset, backend, data seed and assertion mechanism; log evidences are compared only within one backend and one assertion mechanism
+Nested-search convergence needs the search's own termination condition: Nautilus is read from its sampler; other nested searches are not_assessed until PyAutoFit exposes it (phase A3)
+Completion is execution, not convergence; convergence is reported separately from acceptance
+Sample archives remain in project storage; current availability unknown
+
+| Target | Stage | Seed | Execution | Convergence | Acceptance | Sampling s | Total s | Archive | Evidence |
+|---|---|---|---|---|---|---:|---:|---|---|
+
+</details>
+
+
 <!-- decision-history:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 -->
 ## Decision History
 

receipts/fit.json

@@ -0,0 +1,17 @@
+{
+  "instance": "fit",
+  "repo": "autofit_inference",
+  "github": "PyAutoLabs/autofit_inference",
+  "branch": "main",
+  "resolved_commit": "ec35be493712d693b8cab4d071d97d9a242137b6",
+  "summary_path": "dashboard/summary.json",
+  "source": "remote",
+  "fetched_at": "2026-10-08T19:29:28Z",
+  "outcome": "ok",
+  "schema": "inference-summary",
+  "version": 1,
+  "record_count": 0,
+  "errors": [],
+  "cached_from": null,
+  "refreshed_at": "2026-10-08T19:30:00Z"
+}

receipts/lens.json

@@ -13,5 +13,5 @@
   "record_count": 56,
   "errors": [],
   "cached_from": null,
-  "refreshed_at": "2026-10-08T11:16:58Z"
+  "refreshed_at": "2026-10-08T19:29:59Z"
 }

registry.yaml

@@ -7,3 +7,10 @@ instances:
     dashboard_url: https://github.com/PyAutoLabs/autolens_inference#readme
     library_refs: [PyAutoLens, PyAutoGalaxy, PyAutoArray, PyAutoFit]
     cortex_project: autolens_inference
+  - instance: fit
+    repo: autofit_inference
+    summary_path: dashboard/summary.json
+    supported_schema: inference-summary@1
+    dashboard_url: https://github.com/PyAutoLabs/autofit_inference#readme
+    library_refs: [PyAutoFit]
+    cortex_project: autofit_inference

snapshots/fit.json

@@ -0,0 +1,58 @@
+{
+  "instance": "fit",
+  "repo": "autofit_inference",
+  "github": "PyAutoLabs/autofit_inference",
+  "summary_path": "dashboard/summary.json",
+  "commit": "ec35be493712d693b8cab4d071d97d9a242137b6",
+  "fetched_at": "2026-10-08T19:29:28Z",
+  "summary": {
+    "comparison_policy": {
+      "id": "gaussian_x3-task-groups-v1",
+      "note": "Grouped by task (point/MAP, posterior, evidence); wave 1 is a pilot that ranks nothing; never best-seed",
+      "ranking": "none"
+    },
+    "comparisons": [],
+    "coverage": {
+      "excluded": [],
+      "expected": {
+        "reason": "No expected-run manifest yet; wave 1 (B3) defines one",
+        "runs": null
+      },
+      "observed": {
+        "records": 0,
+        "runs": 0
+      }
+    },
+    "evidence_updated_at": null,
+    "evidence_updated_at_reason": "No result rows yet (wave 1 is phase B3)",
+    "generated_at": "2026-10-08T12:57:54Z",
+    "generation_basis": "Source commit timestamp; not a measurement or check-in time",
+    "limitations": [
+      "Verdicts are producer-asserted under protocol gaussian_x3@1, whose PLACEHOLDER thresholds are calibrated in the wave-1 pilot and then frozen",
+      "Rows are judged only against a reference with the same dataset, backend, data seed and assertion mechanism; log evidences are compared only within one backend and one assertion mechanism",
+      "Nested-search convergence needs the search's own termination condition: Nautilus is read from its sampler; other nested searches are not_assessed until PyAutoFit exposes it (phase A3)",
+      "Completion is execution, not convergence; convergence is reported separately from acceptance",
+      "Sample archives remain in project storage; current availability unknown"
+    ],
+    "producer_revision": "91bfcead296c7b3c558db72923b6ab3cf0288931",
+    "project": "autofit_inference",
+    "protocol_id": "gaussian_x3@1",
+    "records": [],
+    "references": {
+      "gaussian_x3_blend/jax_cpu": {
+        "assertion_mechanism": null,
+        "data_seed": null,
+        "status": "pending"
+      },
+      "gaussian_x3_blend/numpy": {
+        "assertion_mechanism": "raise_resample",
+        "data_seed": 1,
+        "status": "complete"
+      }
+    },
+    "schema": "inference-summary",
+    "scope": "inference",
+    "valid_until": null,
+    "version": 1
+  }
+}

state.json

@@ -3,8 +3,8 @@
   "organ": "insight",
   "repo": "PyAutoInsight",
   "status": "yellow",
-  "headline": "1 projects \u00b7 56 records \u00b7 0 cached",
-  "updated": "2026-10-04T08:09:23Z",
+  "headline": "2 projects \u00b7 56 records \u00b7 0 cached",
+  "updated": "2026-10-08T19:29:29Z",
   "pages_url": "https://pyautolabs.github.io/PyAutoInsight/",
   "items": [
     {
@@ -13,6 +13,13 @@
       "text": "lens: ok \u00b7 56 records \u00b7 0/56 accepted by producer; remaining assessments shown below \u00b7 freshness policy unspecified \u00b7 evidence time unknown",
       "url": "https://pyautolabs.github.io/PyAutoInsight/#lens",
       "prompt": null
+    },
+    {
+      "id": "insight:fit",
+      "severity": "yellow",
+      "text": "fit: ok \u00b7 0 records \u00b7 no measurements \u00b7 freshness policy unspecified \u00b7 evidence time unknown",
+      "url": "https://pyautolabs.github.io/PyAutoInsight/#fit",
+      "prompt": null
     }
   ]
 }

tasks/gaussian_x3_search_wave1.md

@@ -0,0 +1,73 @@
+# gaussian_x3 search benchmark — wave 1 (the pilot)
+
+Type: research
+Target: autofit_inference
+Campaign: gaussian-x3-search-benchmark
+Epic: search-extensibility (phase B3)
+Issue: https://github.com/PyAutoLabs/autofit_inference/issues/4
+Filed: 2026-10-08
+
+## Intent
+
+Wave 1 is the **exploratory pilot** of the `gaussian_x3` search benchmark
+(autofit_inference, protocol `gaussian_x3@1`,
+[`wiki/project/protocol_gaussian_x3.md`](https://github.com/PyAutoLabs/autofit_inference/blob/main/wiki/project/protocol_gaussian_x3.md)).
+It **ranks nothing**. It exists to:
+
+1. run every registered PyAutoFit search once per seed on the toy problem, so the
+   catalogue (`catalogue/search_catalogue.json`) can give every search a status —
+   measured, unsupported, deferred or failed — grouped by the task it answers
+   (point/MAP, posterior, evidence), never compared across tasks;
+2. calibrate the protocol's PLACEHOLDER thresholds (σ-ratio band, ppc tolerance, mode
+   radius, ESS and R-hat bars), the per-search timeouts and the censoring rule on the
+   reference runs' seed scatter and the pilot (decision D16 iii), and freeze them as
+   `gaussian_x3@2` before any wave-2 row.
+
+## Menu (protocol §1, §9)
+
+10 search seeds × both datasets (`gaussian_x3_blend`, the `gaussian_x3_separated`
+disjoint-prior control) × `local_numpy_fp64`, and `local_jax_cpu_fp64` where the search
+is JAX-native, on PyAutoFit `main` (post-A1), every row flagged `pilot` and stamped with
+its PyAutoFit commit:
+
+- point/MAP: LBFGS, BFGS, MultiStartAdam, MultiStartProdigy, MultiStartADABelief,
+  MultiStartLion;
+- posterior: Emcee, Zeus (numpy and JAX legs), BlackJAXNUTS cold and warm from a short
+  Nautilus, SMC;
+- evidence: Nautilus (`n_live` 100/200/400), DynestyStatic, DynestyDynamic; NSS is
+  `deferred` until phase A3b (NSS onto Fitness);
+- Drawer is a sanity floor and is not benchmarked (`unsupported` in the catalogue).
+
+The expected-run manifest is `scripts/misc/searches/_pilot.py`; every expected run with
+no row is published as deferred with its reason (`coverage.expected` of the summary), so
+a missing run is never silently absent.
+
+## Where things live
+
+- **Execution and rows**: autofit_inference (`results/searches/`, the exporter's
+  `dashboard/summary.json`, instance `fit` in this registry).
+- **Science record**: PyAutoCortex row `autofit_inference` (ledger
+  `wiki/project/state.md` in the project repo).
+- **Verdict commentary**: autofit_inference `wiki/project/state.md` (journal entry
+  2026-10-08). The planned autofit_assistant campaign page is not made in B3:
+  autofit_assistant gitignores dated `wiki/project/` entries (they are per-clone memory,
+  never shipped with the template), so its home needs a human decision.
+- **Bounded implementation work**: PyAutoMind (the search-extensibility epic).
+
+## Boundaries
+
+Local CPU only. Wave 2 (scored: 50 seeds × 5 data realisations) runs on RAL
+`--partition=ral` only — never `gpu`, `ral,gpu` or `gpu,ral` — on a PyAutoFit revision
+frozen after phases A2 and A4 (phase B5). This task authorizes no compute submission.
+
+## Status at the B3 wrap-up (2026-10-08)
+
+The pilot was stopped by the human wrap-up ruling with **223 of 520** expected runs on
+disk (autofit_inference#4). It ranks nothing. Every run without a row is published as
+`deferred` with its reason: wave 2 on RAL `--partition=ral`, NSS to phase A3b. The
+pilot ran on PyAutoFit main `0dbf258c4f5e`, before phase A2 (PyAutoFit#1679, gradients
+under JAX); BFGS/LBFGS non-finite results are expected to change once A2 merges and are
+not a judgement on those searches. The JAX blend reference is pending (2 of 3 Nautilus
+runs), so JAX blend rows are `not_assessed`. The calibration record keeps `@1` (17
+calibration rows < 20); the `gaussian_x3@2` freeze is deferred and still precedes any
+wave-2 row.

tests/test_registry.py

@@ -13,9 +13,17 @@ def _doc(*rows):
 
 def test_the_committed_registry_is_valid_against_a_body_map(body_map):
     data = yaml.safe_load((ORGAN_ROOT / "registry.yaml").read_text())
-    body_map.update({"PyAutoArray": {"github": "PyAutoLabs/PyAutoArray"}})
+    body_map.update(
+        {
+            "PyAutoArray": {"github": "PyAutoLabs/PyAutoArray"},
+            "autofit_inference": {
+                "path": "fit/autofit_inference",
+                "github": "PyAutoLabs/autofit_inference",
+            },
+        }
+    )
     assert registry.validate(data, body_map) == []
-    assert [r["instance"] for r in data["instances"]] == ["lens"]
+    assert [r["instance"] for r in data["instances"]] == ["lens", "fit"]
 
 
 def test_load_resolves_path_and_github_from_the_body_map(registry_file, fake_mind):