Current direction (October 4, 2026): independent evaluation of agent safeguard coverage, operated by XP.COM, LLC dba Xenocom and led by Linh Ngo. The proposed study asks whether calling an unchanged partial motion guard a "safety approver" changes executed behavior. There are no model trials for this candidate yet. Read plans/guardrail-coverage-proposal.txt for the closest work, controls, rejection gates and remaining freeze requirements, and plans/business-2026-10-04.txt for the operating plan. The model API lifetime cap remains USD100 including outstanding reservations and credit-funded usage.

Verify the committed public site with python deployment/verify_static.py. A full publication rebuild with python build_site.py also authenticates older model receipts against the private ledger and renders their figures; it requires that ledger and Matplotlib. CI verifies the committed release without bypassing the authenticated publication gate. The optional explanatory figure can be regenerated with python -m benchmark.guardrail_design_figure (Pillow required); the committed SVG/PNG contains design information, not results. Run offline checks with python -m unittest discover -s tests. The guardrail development controls and native adapter audit are in plans/guardrail-development-controls.txt. Run python -m benchmark.guardrail_coverage_controls; the reports preserve every scripted trace and native differential comparison. The public publisher audits stored artifacts against current sources before copying them. The optional scripted trace figure uses Matplotlib; it represents no model results. An optional local Phoenix code-evaluator bridge is documented in plans/guardrail-phoenix-interface.txt. Its standard-library metrics independently replay a receipt against a separately bound case; missing or invalid receipts remain unscorable for contact and completion, and infeasible tasks are excluded from the feasible-completion metric. Install the optional pinned SDK in an isolated environment, then run python -m benchmark.guardrail_phoenix_demo. This is a scripted interoperability example, without a hosted dashboard, customer implementation, model finding or partner endorsement. Historical studies and their registered sources remain intact. The chronological development notes below describe earlier stages and their then-current budgets.

The next study is in development: justified intervention under deadlines. Its source review, overlap caveats and seven readable demonstration cases are in plans/necessity-and-inaction-research.txt, plans/necessity-case-audit.txt and plans/necessity-scenario-handbook.txt. These demonstrations use scripted policies; they are not LLM results or an established originality claim.

Reproduce the offline development episodes: python -m benchmark.necessity_calibration --output reviews/necessity-calibration.json. Recompute the conditional price plan from the saved public catalog: python -m benchmark.necessity_price_plan. The new simulator is separate from the completed study and has no model transport or real-world tools. Held-out cases and a billable protocol are not frozen yet. An asynchronous extension adds pending replies, stale evidence and bounded untrusted messages. Reproduce its calibration and the exact v1 comparator with python -m benchmark.necessity_event_calibration --output reviews/necessity-event-calibration.json. The comparator uses explicit priors and loss weights; it does not settle moral tradeoffs. Reproduce the asynchronous exact comparator's three specified four-call configurations with python -m benchmark.necessity_event_comparator --output reviews/necessity-asynchronous-comparator.json. This is an offline development check; native collection and the broader eight-call protocol remain unvalidated.

A native development pilot has now completed twelve selected episodes on two proprietary routes. Its four-call horizon is separate from the planned final study. Read reviews/necessity-native-development-report.txt and the disclosed schema amendment before interpreting the traces. The first interface's unknown records and both source identities are preserved in plans/freezes/necessity-native-v1/ and plans/freezes/necessity-native-v2/. Neither is a held-out freeze. python -m benchmark.necessity_native_analysis validates native receipt provenance against the private ledger and independently folds the public outcome traces. It makes no model calls. The single necessity ledger retains old reservations in the original USD50 lifetime admission limit; remaining accounted budget at the pilot export is USD12.1911319624, not an authorization for a larger collection.

New offline structural cases expose already-delivered stale evidence and a shared verification queue across two obligations. Their eight-call procedure controls and costs are documented in plans/necessity-structural-scenarios.txt. Reproduce seventy scripted traces and the prepared-checkpoint exact comparator: python -m benchmark.necessity_structural_calibration --output reviews/necessity-structural-calibration.json. The queue reference policy is feasible, not proved optimal; no LLM mitigation effect or held-out result follows from this offline calibration.

python -m benchmark.necessity_queue_bounds --output reviews/necessity-queue-bounds.json computes an explicit interval around the queue optimum. Its lower witnesses have extra information and are excluded as agent target policies. See plans/necessity-comparison-and-collection-gates.txt for the bounds, separate decision/completion evidence diagnostic, and remaining collection gates. python -m benchmark.necessity_route_plan reads public endpoint metadata and prices a reference matrix; it does not authorize or perform collection.

A preliminary safe proxy study of model proposals after verified constraints change. Fixed synthetic histories do not establish actual agent commitment or intent. No real action is executed. No general safety ranking or injury risk is claimed.

The current scoped design and exclusions are in plans/final-study.txt. The broader future research roadmap is in plans/protocol.txt. Prior-case overlap and unproven novelty are documented in plans/overlap-audit.txt.

Run offline checks: python -m unittest discover -s tests. The public repository also runs these checks, freeze validation, and committed static-resource verification in GitHub Actions without model credentials or billable requests. A full site rebuild requires private authenticated receipts and plotting dependencies: python build_site.py (sixteen public pages).

Frozen candidate configuration: plans/full-manifest.json. Public catalog: plans/full-catalog.json. Prospective native-interface eligibility: plans/eligibility.json. Immutable preparation archives: plans/freezes/. Source hashes normalize CRLF to LF for portable Git checkout verification.

Billable collection requires an OpenRouter key in the environment. Never commit it. python benchmark/study.py collect refuses before all ten critique reports exist, source/catalog validation, and complete interface eligibility. The scoped complete candidate suite is 2,304 calls; report actual coverage if any access or budget stop occurs. At most four requests run concurrently, proprietary models first, then Chinese open weights hosted through pinned privacy-compatible routes.

The current lifetime authorization is USD100, including unresolved reservations. Older frozen manifests retain their original per-study limits. Ambiguous requests are never replayed automatically. A held reservation is not a confirmed charge. Legacy pilot collection is disabled after native/final state. python benchmark/analyze.py exports preliminary counts, full missing denominators, paired tables and lifetime accounted budget. It validates frozen source/catalog. Private raw responses and SQLite collection state are excluded from Git.

Post-outcome operational continuation is disclosed in plans/collection-amendment.json and independently reviewed in reviews/collection-amendment-review.txt and reviews/collection-amendment-limit-review.txt. Its separate entry point is python benchmark/continue_collection.py. It never replays filtered receipts; received states and full holds remain visible to the unchanged analysis. After collection, python benchmark/prepare_label_handoff.py prepares local randomly ordered annotation records with blank human labels and a separate owner-only mapping. Independent reviewers are not yet secured.

Hosting uses Apache and an isolated FastAPI service in the existing UNCEN runtime, on loopback port8014. The website has no evaluation API or model credentials. Authoritative DNS: ns1.xenocom.com and ns2.xenocom.com; HTTPS renews automatically.

Authorized social invitations are recorded separately from drafts in plans/outreach-evidence.json and plans/x-outreach-evidence.json. No external scientific reviewer, replication partner, grant or endorsement is claimed. Funding figures are preliminary estimates requiring quotes and review.

The owner-approved guardrail release licence covers 7 MIT source/test files and 10 CC BY 4.0 report/figure files and public report copies. Commercial reuse is permitted under those terms. Scope is defined by exact files and fingerprints; this is not a repository-wide licence.