Browse inference setups with qualified baselines and sampler details

.github/workflows/dashboard_refresh.yml

@@ -36,6 +36,7 @@ on:
     paths:
       - "registry.yaml"
       - "campaigns.yaml"
+      - "sampler_candidates.json"
       - "tasks/**"
       - "migration.yaml"
       - "insight/**"

.github/workflows/lint.yml

@@ -54,7 +54,7 @@ jobs:
 
       - name: Install tools
         run: |
-          python -m pip install --quiet ruff pytest pyyaml
+          python -m pip install --quiet ruff pytest pyyaml playwright
           curl -sSfL https://github.com/lycheeverse/lychee/releases/latest/download/lychee-x86_64-unknown-linux-gnu.tar.gz \
             | tar -xz -C /usr/local/bin --strip-components=1 lychee-x86_64-unknown-linux-gnu/lychee
 
@@ -74,6 +74,13 @@ jobs:
           PYAUTO_BRAIN: ${{ github.workspace }}/_brain
         run: bin/pyauto-insight check --offline --mind _mind
 
+      - name: Chromium setup navigation and responsive layout
+        env:
+          PYAUTO_BRAIN: ${{ github.workspace }}/_brain
+        run: |
+          python -m playwright install --with-deps chromium
+          python tests/browser_setup.py
+
       - name: lychee (markdown link-rot)
         env:
           # github.com serves 503 to unauthenticated page fetches (rate limit);

REFERENCE.md

@@ -245,3 +245,31 @@ and state; links continue to point into project repositories.
 Before shipping: Ruff, formatting, full pytest and offline contract validation.
 Integration CI validates Brain's feed contract. Actual sender/receiver success and
 receipt revision are deployment evidence, distinct from mocked workflow tests.
+
+## Setup browser and candidate handoff
+
+Version 2 captures expose project → dataset family → model setup navigation.
+Each setup opens a new tab with a stable `?view=setup&instance=…&setup=…`
+address. The headline shows only the producer-selected, validator-qualified
+reference baseline. Missing acceptance remains explicit; historical runtime is
+never promoted into an expected runtime. Sampler records show initialization,
+hardware, compilation/cache state and measured clock/work units separately.
+Baseline manifests, archived results, comparison refusals and provenance remain
+available in disclosures. Unmapped records remain reachable from the project.
+Without JavaScript, setup content remains accessible as ordinary disclosures.
+Images, when declared, resolve at the captured producer revision.
+
+`sampler_candidates.json` is a manually reviewed public handoff from literature
+curation. Each entry records primary paper and official code URLs, review date,
+expected uses and limitations, separately stated integration/benchmark status,
+and an editable investigation prompt. Update it after an on-demand literature
+review; verify source claims and installed search coverage before changing status.
+Do not copy private notes or infer measured superiority from a paper. Candidate
+prompts propose work through the sampler pipeline; they do not authorize compute.
+The candidate file participates in the dashboard input digest. Regenerate the
+board after changing it and run the offline check.
+
+The required Chromium navigation check (`python tests/browser_setup.py`, after
+installing Playwright and Chromium) covers new-tab routing, mobile width,
+unknown routes, history, editable clipboard payloads and the no-JavaScript
+fallback alongside the full Python suite.

insight/board.py

@@ -9,7 +9,7 @@
 from datetime import UTC, datetime
 from pathlib import Path
 
-from insight import ORGAN_ROOT, campaigns, summary
+from insight import ORGAN_ROOT, campaigns, candidates, setup_browser, summary
 from insight.theme import CSS, JS, theme
 
 PAGES_URL = "https://pyautolabs.github.io/PyAutoInsight/"
@@ -89,7 +89,9 @@ def input_marker(views, data):
     ]
     return (
         "<!-- insight-inputs:"
-        + hashlib.sha256(json.dumps([body, data], sort_keys=True).encode()).hexdigest()
+        + hashlib.sha256(
+            json.dumps([body, data, candidates.load()], sort_keys=True).encode()
+        ).hexdigest()
         + " -->"
     )
 
@@ -135,6 +137,9 @@ def _detail(s, now=None):
         parts.append(
             '<p class="warn">Local preview, not a published capture. No receipt written.</p>'
         )
+    if d.get("version") == 2:
+        nav, _ = setup_browser.render(s, now)
+        return "".join(parts) + nav
     if not d:
         return "".join(parts)
     parts += [
@@ -256,18 +261,21 @@ def _captured_at(views):
 def render_html(views, now=None, campaign_data=None):
     data = campaign_data if campaign_data is not None else campaigns.load()
     shared = theme()
+    browser_css, browser_js = setup_browser.assets()
     return shared.section_layout(
         '<!doctype html><html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">'
         "<title>PyAutoInsight dashboard</title><style>"
         + shared.css("insight")
         + CSS
+        + browser_css
         + "pre{white-space:pre-wrap;overflow-wrap:anywhere}details{margin:1rem 0}dd{margin-left:1rem}</style></head><body>"
         + shared.hero(
             "insight",
             "Inference dashboard",
             navigation=[
                 {"href": "#campaigns", "label": "Active campaigns"},
                 {"href": "#evidence", "label": "Inference evidence"},
+                {"href": "#sampler-candidates", "label": "Sampler candidates"},
                 *(
                     {
                         "href": "#" + v.instance.instance,
@@ -277,7 +285,7 @@ def render_html(views, now=None, campaign_data=None):
                 ),
             ],
         )
-        + "<main>"
+        + '<main><p id="setup-route-status" role="status"></p><div id="inference-home">'
         + input_marker(views, data)
         + campaigns.render_html(
             data,
@@ -289,9 +297,14 @@ def render_html(views, now=None, campaign_data=None):
         )
         + "<p>Execution completion is separate from convergence and scientific acceptance. Cortex retains scientific conclusions. Evidence remains in project storage.</p>"
         + "".join(_detail(s, now) for s in views)
+        + candidates.render(candidates.load())
+        + '</div><div id="inference-pages">'
+        + "".join(setup_browser.render(s, now)[1] for s in views)
+        + "</div>"
         + "</main><script>"
         + shared.JS
         + JS
+        + browser_js
         + "</script></body></html>\n"
     )
 
@@ -321,6 +334,13 @@ def render_markdown(views, now=None, campaign_data=None):
             out.append("Cached: original evidence and capture times retained.")
         if not s.doc:
             continue
+        if s.doc.get("version") == 2:
+            for setup in s.doc["setups"]:
+                out.append(
+                    f"- [{md(setup['dataset_family'])} / {md(setup['model_family'])} / {md(setup['label'])}]"
+                    f"({PAGES_URL}{setup_browser.route(s.instance.instance, setup['id'])})"
+                )
+        out += ["", "<details><summary>Full captured evidence and provenance</summary>", ""]
         out += [
             "",
             "Coverage: " + _dump(s.doc["coverage"]),
@@ -363,6 +383,17 @@ def render_markdown(views, now=None, campaign_data=None):
                 f"Comparison {md(c['id'])}: "
                 + md(summary.comparison_refusals(s.doc, c) or c["protocol"])
             )
+        out += ["", "</details>", ""]
+    out += [
+        "## Sampler candidates",
+        "",
+        f"[Browse candidates and copy investigation prompts]({PAGES_URL}#sampler-candidates)",
+        "",
+    ]
+    for candidate in candidates.load():
+        out.append(
+            f"- [{md(candidate['name'])}]({candidate['paper_url']}): {md(candidate['integration_status'])}. {md(candidate['benchmark_status'])}."
+        )
     return "\n".join(out) + "\n"
 
 

insight/candidates.py

@@ -0,0 +1,78 @@
+"""Curated public literature handoff; expectations are not benchmark evidence."""
+
+import json
+from datetime import date
+from urllib.parse import urlparse
+
+from insight import ORGAN_ROOT
+from insight.setup_browser import disclosure, e
+
+
+def load(path=None):
+    data = json.loads((path or ORGAN_ROOT / "sampler_candidates.json").read_text())
+    if not isinstance(data, list):
+        raise ValueError("sampler candidates must be a list")
+    ids = set()
+    for row in data:
+        if not isinstance(row, dict):
+            raise ValueError("sampler candidate must be object")
+        for key in (
+            "id",
+            "name",
+            "family",
+            "integration_status",
+            "benchmark_status",
+            "investigation_prompt",
+            "reviewed_at",
+        ):
+            if not isinstance(row.get(key), str) or not row[key].strip():
+                raise ValueError(f"candidate needs {key}")
+        if row["id"] in ids:
+            raise ValueError("duplicate sampler candidate id")
+        ids.add(row["id"])
+        for key in ("paper_url", "code_url"):
+            value = row.get(key)
+            url = urlparse(value) if isinstance(value, str) else None
+            if not url or url.scheme != "https" or not url.netloc or url.username or url.password:
+                raise ValueError(f"candidate needs public HTTPS {key}")
+        for key in ("expected_uses", "limitations"):
+            if (
+                not isinstance(row.get(key), list)
+                or not row[key]
+                or not all(isinstance(v, str) and v.strip() for v in row[key])
+            ):
+                raise ValueError(f"candidate needs {key} list")
+        date.fromisoformat(row["reviewed_at"])
+    return data
+
+
+def render(data):
+    out = [
+        '<h2 id="sampler-candidates">Sampler candidates</h2>',
+        "<p>Literature expectations, not measured PyAuto recommendations. Review a candidate and copy its investigation prompt.</p>",
+    ]
+    for i, row in enumerate(data):
+        body = f"<p>{e(row['family'])} · reviewed {e(row['reviewed_at'])}</p>"
+        body += f'<p><a href="{e(row["paper_url"])}" target="_blank" rel="noopener">Paper ↗</a> · <a href="{e(row["code_url"])}" target="_blank" rel="noopener">Official code ↗</a></p>'
+        for label, key in (("Expected uses", "expected_uses"), ("Limitations", "limitations")):
+            body += (
+                f"<h3>{label}</h3><ul>" + "".join(f"<li>{e(v)}</li>" for v in row[key]) + "</ul>"
+            )
+        body += f"<p>Integration: {e(row['integration_status'])}</p><p>Benchmark evidence: {e(row['benchmark_status'])}</p>"
+        prompt = (
+            row["investigation_prompt"]
+            + "\n\nUse the existing sampler pipeline and setup catalogue. Preserve the baseline-prepared model and priors. Distinguish cold/warm/resume starts from compilation/cache state. Propose a bounded implementation or benchmark plan before executing compute. Paper: "
+            + row["paper_url"]
+            + "\nCode: "
+            + row["code_url"]
+        )
+        body += '<div data-candidate><button type="button" data-candidate-copy aria-label="Copy investigation prompt">⧉ Copy investigation prompt</button><span role="status" aria-live="polite"></span>'
+        body += (
+            disclosure(
+                "Investigation prompt",
+                f'<label for="candidate-{i}">Editable prompt</label><textarea class="candidate-prompt" id="candidate-{i}">{e(prompt)}</textarea>',
+            )
+            + "</div>"
+        )
+        out.append(disclosure(row["name"], body))
+    return "".join(out)

insight/cli.py

@@ -300,7 +300,7 @@ def common(p, reading=True):
     args = parser.parse_args(argv)
     try:
         return args.func(args)
-    except registry.RegistryError as exc:
+    except (registry.RegistryError, ValueError) as exc:
         print(f"pyauto-insight: {exc}", file=sys.stderr)
         return 1
 

insight/setup_browser.css

@@ -0,0 +1,13 @@
+.setup-choices{display:grid;grid-template-columns:repeat(auto-fit,minmax(min(100%,18rem),1fr));gap:.75rem;margin:1rem 0}
+.setup-choice{display:block;padding:1rem;border:1px solid var(--line);border-radius:.6rem;background:var(--card);text-decoration:none}
+.setup-choice:hover,.setup-choice:focus-visible{border-color:var(--accent);outline:2px solid var(--accent);outline-offset:2px}
+.setup-metrics{display:grid;grid-template-columns:repeat(auto-fit,minmax(min(100%,12rem),1fr));gap:.75rem;margin:1rem 0}
+.setup-metrics>div{padding:1rem;background:var(--card);border:1px solid var(--line);border-radius:.6rem;min-width:0}
+.setup-metrics dt{font-size:.85rem;color:var(--muted)}.setup-metrics dd{margin:.35rem 0 0!important;font-size:1.15rem;font-weight:600;overflow-wrap:anywhere}
+.setup-page>summary{font-weight:600}.setup-page[open]{border:0}.setup-view .setup-page>summary{display:none}
+.setup-image{max-width:100%;height:auto;max-height:28rem;object-fit:contain;border-radius:.6rem}
+.setup-view .board-nav,.setup-view [data-orchestration-panel]{display:none}
+[hidden]{display:none!important}.setup-page pre{max-height:32rem;overflow:auto;font-size:.85rem}
+.setup-page details{padding:.75rem;border:1px solid var(--line);border-radius:.5rem}.setup-page summary{cursor:pointer}
+.candidate-prompt{width:100%;box-sizing:border-box;min-height:8rem;font:inherit;padding:.75rem}
+.sr-only{position:absolute;width:1px;height:1px;padding:0;margin:-1px;overflow:hidden;clip:rect(0,0,0,0);white-space:nowrap;border:0}

insight/setup_browser.js

@@ -0,0 +1,33 @@
+(() => {
+  function route() {
+    const query = new URLSearchParams(location.search);
+    const pages = [...document.querySelectorAll('.setup-page')];
+    const selected = pages.find(p => p.dataset.instance === query.get('instance') && p.dataset.setup === query.get('setup'));
+    const detail = query.get('view') === 'setup' && !!selected;
+    document.body.classList.toggle('setup-view', detail);
+    document.getElementById('inference-home').hidden = detail;
+    document.getElementById('inference-pages').hidden = !detail;
+    for (const page of pages) {
+      page.hidden = page !== selected;
+      page.open = page === selected;
+    }
+    document.getElementById('setup-route-status').textContent = query.get('view') === 'setup' && !selected ? 'This setup is unavailable in the captured evidence. Choose an available setup below.' : '';
+    if (detail) document.title = selected.querySelector('h1').textContent + ' · PyAutoInsight';
+  }
+  route();
+  addEventListener('popstate', route);
+  document.querySelectorAll('[data-candidate-copy]').forEach(button => button.addEventListener('click', async () => {
+    const panel = button.closest('[data-candidate]');
+    const text = panel.querySelector('textarea');
+    const status = panel.querySelector('[role="status"]');
+    try {
+      if (!navigator.clipboard) throw new Error('Clipboard unavailable');
+      await navigator.clipboard.writeText(text.value);
+      status.textContent = 'Prompt copied';
+    } catch (_) {
+      text.closest('details').open = true;
+      text.focus(); text.select();
+      status.textContent = 'Select and copy the prompt below';
+    }
+  }));
+})();