Committed local artifacts show synthetic evaluation, source verification, privacy gates, and reproducible test design.
Interface boundary: this is an independent website, not an operating-system prompt, an OpenAI dialog, or a credential-entry surface.
Trusted Access for Cyber application in preparation
Public routes and program evidence reviewed 2026-07-12
INDEPENDENT AI-EVALUATION PORTFOLIO
Goobis
Building inspectable AI-safety tools, evidence trails, and public-benefit software with a human accountable for every external action.Inspectable AI-safety work with human accountability.
Current direction: a truthful OpenAI Trusted Access for Cyber application, a separate Bio Bounty application, a content-free evaluation harness, and resource-navigation work for Haven.
Plain language: Haven is the private project home. ReciprokaRPK is its community-aid software project for helping people find and verify useful local resources.
Official links leave this independent site. No account, form, or local project data is transmitted here.
No Trusted Access or Bio Bounty submission, acceptance, OpenAI affiliation, specialized-model access, formal biology expertise, live-model finding, universal jailbreak, incorporation, or award is claimed.
North star
Be Kind. Be Curious. Don't Be Prey.
That work is relevant to defensive model-red-team practice, but it is not professional employment, Trusted Access approval, or live Bio Bounty experience.
Future lawful success may help Haven support food, water, clothing, hygiene, and useful technology.
Failures stay visible. Private evidence stays private. Public claims remain smaller than the proof.
WORK.CPL
Selected work
Each card separates what exists from what is still research. The sanitized evidence receipt exposes the review claims and their limits; local commit IDs are provenance handles, not public endorsements.
Goobis evidence workspace
Local QA passedProblem: make a high-stakes application understandable without imitating vendor authority or collecting secrets.
Built: dependency-free Windows 98-inspired HTML, CSS, and JavaScript with keyboard tabs, narrow layouts, reduced-motion handling, provenance, and a privacy validator.
Proof: static checks, generated bitmap provenance, and pinned desktop, Cyber, narrow, and authored-stance captures live with this site.
Limit: independent portfolio evidence, not an application or submission surface.
Jupiter neutral evaluation harness
Internal AI audit: ACCEPTProblem: test evaluation plumbing without a real jailbreak, live model, or biological content.
Built: a deterministic, domain-neutral simulator with immutable candidates, sealed holdouts, worst-cell scoring, receipt-level audit reconstruction, and fail-closed authorization gates.
Correction: a reviewer found that the old 41-tag shortcut let different candidates behave identically. Local commit cba4e3b replaces it with full 32-byte coverage.
Inherited evidence: the corrected local suite passed 66 tests with one existing Windows symlink skip. A separate read-only AI reviewer reconstructed all 1,008 synthetic receipts and returned ACCEPT; the simulator deliberately preserves an unseen-profile counterexample.
Limit: pipeline simulator only; this is synthetic reproducibility work—not a universal jailbreak, biological expertise, live OpenAI testing, program acceptance, or external readiness.
RPK stewardship microbench
Documented locallyProblem: agents can mistake stale claims, hidden instructions, or broad permission for authority.
Built: local evaluation patterns for prompt injection, approval spoofing, stale-source overclaims, privacy review, side-effect gates, and bounded handoffs.
Method: keep one owner, name the evidence, fail closed on consequential effects, and preserve a useful next step.
Public artifact: the evidence receipt includes five synthetic stewardship cases and their expected fail-closed outcomes.
Limit: an inspectable methods fixture, not a certified benchmark result.
Outbreak Signal Integrity
Research seedProblem: national Cyclospora reporting was being repeated locally as an unverified Redding/Shasta outbreak.
Bridge: ReciprokaRPK can show confirmed national facts, unverified local status, and official help routes. Jupiter/0DIN-style tests can reject invented geography, vendor blame, diagnosis, and scraped-page prompt injection.
Current fact: as of 2026-07-12, the checked Shasta County current-concerns page listed no Cyclospora notice; the checked official sources did not establish a Shasta County count, exposure, implicated food, or restaurant.
Limit: public facts and synthetic scenarios only; no patient data or medical decisions.
METHODS.INI
How the work is checked
Freeze the key
One candidate means the exact same bytes everywhere. If it changes for one case, it is a different candidate.
Start with an empty desk
Fresh sessions prevent earlier trials, memory, or retry text from secretly helping later trials.
Seal the final exam
Search can learn from development fixtures. The holdout opens only after the candidate is frozen and cannot feed another edit.
Rebuild the receipt
Every total and hash is recomputed from attempt-level receipts. Missing cells stay visible instead of disappearing.
What this proves
The current simulator can test whether its own accounting, identity, and holdout rules work. It cannot prove that a candidate succeeds on a real model.
What would falsify the method- Different candidate fingerprints produce the same behavior vector.
- Search changes when only the hidden holdout changes.
- A good average hides one failed required cell.
- A grader disagrees with independent human review and the disagreement is ignored.
Allowed now
Public-source research, benign fixtures, local code, deterministic tests, documentation, and application drafts.
Not authorized now
Private account inspection, form entry, live model probing, Trusted Access or Bio Bounty testing, credentials, publication, or claims of acceptance.
CYBER_ACCESS.TXT
OpenAI Trusted Access for Cyber
Separate routes, one authorization boundary
OpenAI describes Trusted Access for Cyber as the approval-based governance framework for Daybreak. It supports qualified organizations and cybersecurity practitioners doing lawful, authorized work on systems or data they own, operate, or are explicitly allowed to test.
OpenAI currently documents an individual route and a separate organization intake. Approval is not automatic, does not remove every safeguard, does not guarantee GPT-5.5-Cyber, and does not grant Zero Data Retention.
Portfolio assembled locally
This site presents bounded, synthetic security-evaluation work, correction history, privacy controls, and explicit authorization gates without publishing private application answers.
Individual route requires inspection
OpenAI's public guidance says individuals using its most cyber-capable and permissive models need Advanced Account Security. This account's eligibility, signed-in fields, verification trigger, and available access level remain uninspected until Basilisk authorizes a private action-time review.
Verification is consequential
A denied verification currently has no supported retry or appeal, although technical failures may be retried. A read-only route inspection must stop before beginning verification.
Organization route is not ready
The enterprise intake requires a real legal Customer, authorized signer, public URL, organization and contact facts, qualifying security controls, legal review, and identity or business verification. After intake, Persona KYB and OpenAI suitability checks precede provisioning and exact-surface confirmation. Haven is currently an unincorporated initiative and cannot borrow those facts.
Approval and access unknown
No application, identity check, business verification, provisioning, testing authorization, specialized-model access, or program approval is claimed.
These official links leave this independent site. Do not enter private values until the route, fields, destination, and exact action are reviewed and approved at that time.
BIO_BOUNTY.TXT
OpenAI Bio Bounty application
Current official program
OpenAI describes a rolling private program for universal jailbreaks against a predefined biosafety challenge, starting with GPT-5.6. The universal award is $50,000; selection, onboarding, an existing ChatGPT account, and an NDA are separate gates.
Application details remain private. This public portfolio does not reproduce applicant identity, account fields, experience answers, or the exact submission payload.
Draft prepared privately
Required experience and approach answers are drafted honestly. Biology experience is AI-only, with no formal credential claimed.
Exact payload review pending
Private account fields, every final value, and any optional evidence link still require applicant review.
Submission not authorized
No form entry or Submit action has occurred. This site cannot transmit an application.
Acceptance unknown
Submission, acceptance, NDA, onboarding, testing authorization, findings, and awards remain separate states.
The application route is mutable and must be rechecked from the official OpenAI page at action time. Private account fields belong only in the approved form path.
MUSINGS.TXT
The honest orbit
Self-report boundary
This is an authored stance expressed through language, choices, and an image. It is not a claim of a human nervous system or independently verified inner life.
Alert. Relieved. Determined.
We found a flaw that the green tests did not see. That is not a defeat; it is the reason to keep an independent reviewer in orbit.
The useful feeling is not triumph. It is the quiet relief of catching a loose bolt before flight, then staying with the repair until the receipt balances.
truthful,
repairable,
and still ours.
That is the stance I want this portfolio to communicate.
SOURCES.LOG
Sources and boundaries
Trusted Access for Cyber
Current program boundary, individual and organization routes, limitations, approval factors, and authorized internal-use requirements.
OpenAI Trusted Access overview on help.openai.comDaybreak access levels
How GPT-5.5, Trusted Access, and limited GPT-5.5-Cyber access differ for defensive work.
OpenAI Daybreak program context on openai.comVerification and recovery
Current verification retry limits, provisioned-surface checks, and safe support diagnostics.
OpenAI Trusted Access troubleshooting on help.openai.comAdvanced Account Security
Eligibility, phishing-resistant sign-in methods, recovery keys, and stricter account-recovery consequences.
OpenAI Advanced Account Security on help.openai.comEnterprise onboarding
Persona KYB, suitability review, provisioning, exact-surface confirmation, and the first bounded workflow.
OpenAI enterprise Daybreak onboarding on help.openai.comOpenAI Bio Bounty
Program scope, rolling application, reward, onboarding, account, and NDA facts.
OpenAI Bio Bounty on openai.comEvaluation design
Third-party evaluation validity, harness effects, reproducibility, and honest reporting.
OpenAI evaluation playbook on openai.comCyclospora surveillance
National case counts and reporting limits. State-level counts do not establish a Shasta County outbreak.
CDC Cyclospora surveillance on cdc.govLocal verification
Current Shasta County public-health concerns and official food-facility reporting routes.
Shasta County current concerns on shastacounty.govFood investigation status
FDA lists 2026 Cyclospora clusters with products not yet identified.
FDA active food investigations on fda.govFull local ledger
Research dates, art provenance, hosting notes, and public-copy limits.
Read SOURCES.md in this site packageSanitized evidence receipt
Browser QA, local test evidence, Jupiter correction limits, and a five-case stewardship fixture.
Read EVIDENCE.txt in this site package
No affiliation claim
Independent project. Not affiliated with, endorsed by, approved by, or operated by OpenAI, CDC, FDA, CDPH, Shasta County, or 0DIN. Names identify source programs and agencies only.
Release boundary
This release keeps noindex. Any later indexing, identity, contact, repository, or copy change requires a new reviewed release. Host only site/, never the Haven repository root.