#!/usr/bin/env python3 """Wire provenance chips + nav/footer updates into the existing leadership pages. Every substitution asserts its expected match count — nothing silently misses.""" import json import re from pathlib import Path SITE = Path(__file__).resolve().parent.parent / "site" DATA = json.load(open(SITE / "assets" / "claims.json")) CLAIMS = {c["id"]: c for c in DATA["claims"]} def chip(cid, rel=""): c = CLAIMS[cid] du = c.get("decision_use") or {} g = du.get("grade", "?") title = (f'[Decision-use {g} — {du.get("basis", "")}] ' + c["claim"][:190]).replace('"', """) return f'{cid} · {g}' def gapchip(gid, label, rel=""): return f'{label}' def sub_once(text, old, new, where, count=1): n = text.count(old) assert n == count, f"[{where}] expected {count} of pattern, found {n}: {old[:90]!r}" return text.replace(old, new) # ---------------------------------------------------------------- nav + footer, all pages OLD_NAV_ROOT = ''' Thesis Research System Pilot Library''' NEW_NAV_ROOT = ''' Thesis Research System Pilot Decisions Evidence Library''' OLD_NAV_VIEW = OLD_NAV_ROOT.replace('href="', 'href="../../') NEW_NAV_VIEW = NEW_NAV_ROOT.replace('href="', 'href="../../') OLD_FOOT_ROOT = 'Primary artifacts

Research atlas
' NEW_FOOT_ROOT = 'Primary artifacts

Evidence ledger
Decision register
Research atlas
' OLD_FOOT_VIEW = 'Primary artifacts

Research atlas
' NEW_FOOT_VIEW = 'Primary artifacts

Evidence ledger
Decision register
Research atlas
' nav_hits = foot_hits = 0 for p in list(SITE.glob("*.html")) + list((SITE / "documents" / "view").glob("*.html")): if p.name in ("evidence.html", "decisions.html"): continue t = p.read_text(encoding="utf-8") t0 = t for old, new in ((OLD_NAV_ROOT, NEW_NAV_ROOT), (OLD_NAV_VIEW, NEW_NAV_VIEW)): if old in t: t = t.replace(old, new) nav_hits += 1 for old, new in ((OLD_FOOT_ROOT, NEW_FOOT_ROOT), (OLD_FOOT_VIEW, NEW_FOOT_VIEW)): if old in t: t = t.replace(old, new) foot_hits += 1 if t != t0: p.write_text(t, encoding="utf-8") print(f"nav updated in {nav_hits} files; footer updated in {foot_hits} files") # ---------------------------------------------------------------- research.html p = SITE / "research.html" t = p.read_text(encoding="utf-8") CARD_CHIPS = [ ("https://arxiv.org/abs/2606.19544", ["AJ-03", "CT-01"]), ("https://arxiv.org/abs/2410.03775", ["AQ-03"]), ("https://arxiv.org/abs/2603.25133", ["AJ-04", "F26-06"]), ("https://arxiv.org/abs/2603.01562", ["RR-04", "RR-07"]), ("https://arxiv.org/abs/2512.10791", ["GF-07"]), ("https://aclanthology.org/2025.naacl-long.320/", ["GF-03"]), ("https://arxiv.org/abs/2510.26518", ["CM-03"]), ("https://arxiv.org/abs/2407.00215", ["CM-02", "CM-01"]), ("https://arxiv.org/abs/2504.09737", ["CM-05"]), ("https://arxiv.org/abs/2407.18370", ["HS-02"]), ("[link withheld]", ["F26-02", "IP-02"]), ("https://www.pnas.org/doi/10.1073/pnas.2518075122", ["AQ-02"]), ("https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024L2831", ["CG-07", "CG-08", "CG-09", "CG-10"]), ] for url, ids in CARD_CHIPS: anchor = f'

Primary source ↗' chips = "".join(chip(i) for i in ids) t = sub_once(t, anchor, chips.join(['
', f'
{anchor}']), f"research card {ids[0]}") # open-problem card gets a gap chip t = sub_once( t, '
Primary source ↗', f'
{gapchip("gap-01", "GAP-01 · unmeasured")}
' '
Primary source ↗', "research open-problem card") t = sub_once( t, '

Each card gives the finding, the design consequence, and the main caveat. Open the local domain report for the full evidence trail.

', '

Each card gives the finding, the design consequence, and the main caveat. Every tag opens the claim’s entry in the evidence ledger — source, quotes, both verification layers, and its decision-use grade (A durable · B current-gen · C superseded magnitude). Open the local domain report for the full evidence trail.

', "research intro") t = sub_once( t, '
Transfer caveat. Nearly all published judge benchmarks grade model outputs. This product grades human evaluative writeups about model outputs. The architecture transfers plausibly; the headline accuracies do not transfer automatically.
', '
Transfer caveat. Nearly all published judge benchmarks grade model outputs. This product grades human evaluative writeups about model outputs. The architecture transfers plausibly; the headline accuracies do not transfer automatically.
\n
Capability-currency caveat. These decisions would be executed with deployment-tier models — Claude Fable / Opus 4.8 class, GPT-5.5 / 5.6 class — and zero published measurements test that tier on this task. Numbers below carry decision-use grades: durable results (law, statistics, human science, adversarial floors) can be decided on now; capability-dependent magnitudes — in either direction — are priors that the pilot bakeoff re-measures on the deployment tier before any threshold is set.
', "research capability callout") p.write_text(t, encoding="utf-8") print("research.html: 14 cards chipped + capability callout") # ---------------------------------------------------------------- index.html p = SITE / "index.html" t = p.read_text(encoding="utf-8") INDEX_CHIPS = [ ('See evidence and caveat →', ["AJ-03"]), ('See the rubric evidence →', ["AJ-04"]), ('See the human-loop evidence →', ["CM-03"]), ('See the grounding evidence →', ["GF-07"]), ('See the feedback evidence →', ["CM-05"]), ] for anchor, ids in INDEX_CHIPS: chips = "".join(chip(i) for i in ids) t = sub_once(t, anchor, f'
{chips}
{anchor}', f"index card {ids[0]}") t = sub_once( t, 'See what remains unknown →', f'
{gapchip("gap-01", "GAP-01 · unmeasured")}
See what remains unknown →', "index open-problem card") t = sub_once( t, '
\n
✓Choose the construct', '
\n
✓Choose the construct', "index decision anchor", count=1) if '
\n
✓Choose the construct' in t else t # robust fallback: add link line after the section head paragraph t = sub_once( t, '

The first goal is to determine where AutoQA is more valid, more consistent, and cheaper than the current review layer—and where it should abstain.

', '

The first goal is to determine where AutoQA is more valid, more consistent, and cheaper than the current review layer—and where it should abstain. Each ask below — with its options, evidence, reversibility, and failure condition — lives in the decision register.

', "index decisions link") t = sub_once( t, '
Open the pilot and launch gates', '
Open the decision registerOpen the pilot and launch gates', "index hero actions") t = sub_once( t, '

Several findings are strong enough to shape the foundation. The exact performance on human evaluative writeups remains an in-house measurement problem.

', '

Several findings are strong enough to shape the foundation. The exact performance on human evaluative writeups remains an in-house measurement problem — and no published study tests the deployment-tier models this program would actually run (Claude Fable / Opus 4.8 class, GPT-5.5 / 5.6 class) on this task. Every number carries a decision-use grade in the evidence ledger; capability-dependent magnitudes are re-measured at the bakeoff before they set any threshold.

', "index capability sentence") p.write_text(t, encoding="utf-8") print("index.html: chipped + decision links + capability sentence") # ---------------------------------------------------------------- brief.html p = SITE / "brief.html" t = p.read_text(encoding="utf-8") t = sub_once( t, '

Raw agreement exceeded chance-corrected agreement by this amount in one 2026 multi-judge study.

', f'

Raw agreement exceeded chance-corrected agreement by this amount in one 2026 multi-judge study.

{chip("AJ-03")}
', "brief slide 4") t = sub_once( t, '

Evidence-assisted selective hybrid performance in one fact-verification study, versus 87.7% AI alone.

', f'

Evidence-assisted selective hybrid performance in one fact-verification study, versus 87.7% AI alone.

{chip("CM-03")}
', "brief slide 6") t = sub_once( t, '
The honest caveat

No direct public benchmark exists for warranted positive praise in human evaluative writeups. That is the first in-house benchmark.

', f'
The honest caveat

No direct public benchmark exists for warranted positive praise in human evaluative writeups. That is the first in-house benchmark.

{gapchip("gap-01", "GAP-01 · the moat is unmeasured")}
', "brief slide 7") t = sub_once( t, '
Shadow

Prove calibrated risk–coverage before granting authority.

', '
Shadow

Prove calibrated risk–coverage before granting authority.

Every decision, its evidence, and its flip condition: the decision register →

', "brief slide 10") p.write_text(t, encoding="utf-8") print("brief.html: chipped") # ---------------------------------------------------------------- pilot.html p = SITE / "pilot.html" t = p.read_text(encoding="utf-8") t = sub_once( t, '
Risk register

The system is designed around its own fallibility.

These are monitored failure modes, not footnotes.

', '
Risk register

The system is designed around its own fallibility.

These are monitored failure modes, not footnotes. The Ledger column ties each risk to the measured evidence that makes it a design constraint rather than an opinion.

', "pilot risks anchor") t = sub_once( t, '\n ', '
RiskControlSensor
Consistent but invalid judge
\n ', "pilot risk table header") RISK_CHIPS = [ ('', ["AJ-03", "AJ-02"]), ('', ["CT-02", "CM-03"]), ('', ["RR-05", "CT-04"]), ('', ["CT-09", "F26-03", "AQ-02"]), ('', ["CT-05", "AJ-08"]), ('', ["CM-08", "CM-07"]), ('', ["AJ-11", "CT-08"]), ('', ["CG-07", "CG-11"]), ] for cell, ids in RISK_CHIPS: chips = "".join(chip(i) for i in ids) t = sub_once(t, cell + "", f'{cell}', f"pilot risk row {ids[0]}") t = sub_once( t, '
  • Robustness to paraphrase, order, and evidence-swap probes.
  • Random audits of confident passes remain active.
    • ', f'
  • Robustness to paraphrase, order, and evidence-swap probes.
  • Random audits of confident passes remain active.
    • {chip("GF-10")}{chip("CM-02")}
    ', "pilot acceptance gate chips") t = sub_once( t, '
  • Meaningful human review where consequences or law require it.
    • ', f'
  • Meaningful human review where consequences or law require it.
    • {chip("F26-04")}{chip("CG-07")}{chip("CG-11")}
    ', "pilot rejection gate chips") t = sub_once( t, '

    Engineering is not the scarce resource at the start. Aligned examples and decision rights are.

    ', '

    Engineering is not the scarce resource at the start. Aligned examples and decision rights are. Each ask — options, evidence, reversibility, failure condition — lives in the decision register.

    ', "pilot decisions link") t = sub_once( t, '

    Run model and architecture bakeoffs

    Compare deterministic checks, specialist verifiers, frontier reasoning models, reference timing, decomposition granularity, cross-family challengers, and confidence signals by criterion family.

    ', '

    Run model and architecture bakeoffs

    Compare deterministic checks, specialist verifiers, frontier reasoning models, reference timing, decomposition granularity, cross-family challengers, and confidence signals by criterion family — on the deployment-tier set (Claude Fable / Opus 4.8 class, GPT-5.5 / 5.6 class, cross-family). No published anchor covers these models on this task; this gate produces the numbers the literature cannot.

    ', "pilot bakeoff deployment tier") p.write_text(t, encoding="utf-8") print("pilot.html: risk table + gates chipped + bakeoff tier") # ---------------------------------------------------------------- architecture.html p = SITE / "architecture.html" t = p.read_text(encoding="utf-8") t = sub_once( t, '
    Claim ontology', '
    Claim ontology', "architecture ontology anchor") ONT_CHIPS = [ ('
    ', ["GF-01", "GF-07"]), ('', ["GF-03"]), ('', ["CM-06"]), ('', ["GF-05"]), ('', ["AJ-02", "F26-08"]), ] for cell, ids in ONT_CHIPS: chips = "".join(chip(i) for i in ids) t = sub_once(t, cell, cell[:-5] + f'
    {chips}
    ', f"arch ontology {ids[0]}") t = sub_once( t, '

    Run deterministic checks, grounding checkers, reasoning judges, context-closed evidence checks, and challenge passes.

    4', f'

    Run deterministic checks, grounding checkers, reasoning judges, context-closed evidence checks, and challenge passes.

    {chip("F26-01")}
    4', "arch verify-lane chip") t = sub_once( t, '
    Package rule: autonomous acceptance or rejection should not be enabled below Level B.
    ', f'
    Package rule: autonomous acceptance or rejection should not be enabled below Level B. Where consequences reach EU workers, the human floor is law, not policy. {chip("CG-07")}{chip("CG-11")}
    ', "arch conformance chips") p.write_text(t, encoding="utf-8") print("architecture.html: ontology + pipeline chipped") # ---------------------------------------------------------------- README.md p = SITE / "README.md" t = p.read_text(encoding="utf-8") t = sub_once( t, "- `pilot.html` — pilot, metrics, launch gates, and risk register", "- `pilot.html` — pilot, metrics, launch gates, and risk register\n- `decisions.html` — decision register: authorizations, open calls, defaults, flip conditions\n- `evidence.html` — evidence ledger: all 132 corpus claims with two verification layers\n- `assets/claims.json` — machine-readable ledger (claims, sources, verdicts, usage)", "README contents") t = t.replace("Generated 2026-07-15.", "Generated 2026-07-15. Evidence ledger and decision register added, all claims re-verified against their primary sources, and the site manifest regenerated the same day.") p.write_text(t, encoding="utf-8") print("README.md updated")
    RiskControlSensorLedger
    Consistent but invalid judgeValidity by criterion and agreement stratum.Anchoring A/B and overturn quality.Disagreement clusters and appeal themes.AutoQA score rises while gold score stays flat.Error-overlap and judge–generator similarity.False-accusation rate and new-defect rate.Performance by difficulty and expert slice.Appeal outcomes, adverse-action audit, legal review.
    {chips}
    Span-closed entailment plus source-quality check.Reasoning judge constrained to the evidence record.Explicit expected-element checklist; “not contradicted” unless seeded recall is proven.Never truth-score the opinion; assess contradiction and warrant.Unknown or escalation—not a forced pass or fail.