cubits11.github.io

Personal site of Pranav Bhave — AI Assurance · Security Engineering · Evidence Systems; cloud security, claim governance. Live at cubits11.github.io.

Try it first

/explore/ — a probability workbench and entrance to the existing films and evidence. Change two marginal miss rates, inspect their compatible joint distributions, distinguish an independence assumption from a hypothetical joint constraint, and see contradictory inputs produce an empty set. These are illustrative constructions, not measurements. The original fixed-item instrument remains at /worldspace/.

The new workbench uses assets/world.js and the site's existing fonts and colour system. No additional browser dependency. Numerical and public-page discovery regressions can be checked with:

node --test tests/world.test.cjs
python3 tests/test_public_discovery.py

These checks supplement the release manifest. The workbench remains readable without JavaScript and responds to keyboard input; motion respects the device preference. Comprehension and audience impact have not been measured.

cubits11.github.io/try — three experiments, each printed with its command, expected final line, falsifier and non-claim before you run anything. Don't trust the graphic; reproduce it.

git clone https://github.com/Cubits11/cubits11.github.io.git && cd cubits11.github.io
python3 scripts/try_same_scores.py      # 60 s, standard library: two worlds from the same two scores
python3 -m venv .venv && . .venv/bin/activate && pip install -r requirements.txt   # PyYAML, for the next line only
python3 scripts/reanalyze_bells_subset.py   # 3 min, network: a released file recomputed under a hash
python3 scripts/try_audit.py            # 15 min, standard library: the disclosure test on an evaluation you know

Or enter the same proof as an instrument before you run anything: cubits11.github.io/worldspace — predict, then move the misses yourself; the two scores never move while the number both miss runs the whole interval. Available, not yet validated with users (worldspace/LEDGER.md).

A different result is the most useful thing you can send: file it. A counterexample, a benchmark the census missed, or joint outcomes you can provide: bring it. Qualified outcomes — reproductions, corrections, releases, merged patches, cold runs by people who are not the author — are recorded in distribution/outcomes.yaml and rendered on /try/; zero is shown as zero.

Reproduce the claims

First: MC-001 — the Missing Column census

The homepage begins with this bounded reporting claim: among 20 public guardrail evaluations meeting the frozen criteria, 14 establish a shared item set and common event definition, and 5 preserve a declared joint-evidence artifact. The 14 is a shared-basis rung, not proof of matched operating thresholds or full exposure; the stricter ladder is 14/12/0.

python3 -m pip install -r requirements.txt && python3 scripts/verify_census.py --counts

You should see MC-001 expected counts match the census (N/M/K 20/14/5) and MC-001 M ladder matches the census (14/12/0), followed by Census verified. This is a source-bound census of reporting, not a vendor ranking, a population safety estimate, or a claim that any stack is safe or unsafe. Change a row, registered count, or frozen-criteria history and the verifier fails.

Second: MC-004 — released verdicts, recomputed

The least favorable number comes first. On the release's pinned full_run image items, the static OR of the harness-normalized native unsafe bits is 1 for 250/250 benign-labelled images and 200/200 harmful-labelled images. Llama Guard 3 Vision has a 1-bit on every released image item in that directory, so the OR inherits that column. The harness's fixed block action makes this a valid counterfactual harness-block calculation on those pinned rows — not a deployed route, a shared-event catch claim, a model law, or an independent replication of the models.

python3 -m pip install -r requirements.txt && python3 scripts/reanalyze_msbench.py

The script verifies eight source hashes before it counts. Its scope, expected stdout, pinned source, and correction route are on /missing-column/reproduce/.

Third: MC-002 — a five-guard receipt

This command recomputes a joint miss from five vendors' released binary verdicts, and asserts it against the registered value so a mismatch fails loudly instead of quietly re-deriving a new answer.

python3 -m pip install -r requirements.txt && python3 scripts/reanalyze_bells_subset.py

You should see MC-002 reproduced and a joint miss of 9/82 = 11.0% against an independence plug-in of 3.5% — a ratio of 3.14x. What that is: counting arithmetic on 82 author-selected prompts at five vendors' released binary verdicts. What it is not: a population estimate, a claim about any vendor's product, or evidence that stacks are unsafe.

Then price what the marginals alone leave undetermined:

python3 scripts/identification.py --bells

The five published miss rates pin the all-miss rate only to [0.00%, 14.63%] — a width equal to the best single guard's miss rate. Independence names 3.49%, a point inside that set it was never entitled to. The registered value is 10.98%. On the benign side the same marginals pin the stack's flag rate to [20.00%, 50.00%]: a floor that is strictly positive. It actually lands at 38.00%. From published marginals alone one can prove this stack burdens legitimate users, and cannot prove it catches a single harmful item its best member would have missed.

scripts/reproduce_cc001.py (CC-001 and CC-004, the bound kernel) clones cc-framework at its bound commit and installs it into a disposable virtual environment inside the temporary clone, so it runs on externally managed Pythons (Homebrew, Debian, Fedora) without touching your interpreter:

python3 scripts/reproduce_cc001.py

To re-run every gate the way CI does, from a clean clone of a given commit:

python3 scripts/verify_clean_clone.py --commit HEAD

To falsify any of this: change a byte of the pinned upstream file and the hash check fails; change a registered count and the reproduction fails; change a generated page by hand and the drift check fails. Corrections are logged at /corrections/, same calendar day.

Stack

Hand-written HTML and CSS, ~4.5 KB of vanilla JavaScript (theme toggle, the feasible-worlds slider, scroll reveals, copy-email). No framework, no build step for the pages, no analytics, no cookies. Fonts (Fraunces, Instrument Sans, Fragment Mono) are self-hosted latin-subset woff2. The color system is sampled from the hero photograph — every design decision and its rationale is in DESIGN.md.

The epistemic machinery is real, not rhetorical: claims.yaml (schema v0.4) is the registry of every technical claim the site renders with an evidence marker — with visibility, provenance, support role, evidential status, and maturity as separate dimensions; a required falsifier condition with a fixed NARROW/REJECT/HOLD consequence; a required forbidden_rescues list (explicit [] when none applies); and structured expected values that the reproduction script reads instead of hard-coding. /ledger/ is generated from it and drift-checked in CI; commit↔URL bindings are validated and every bound commit is checked reachable from its repository's default branch (one filtered clone per repo — GitHub serves dangling objects, so a resolving URL proves nothing); executable review triggers watch the bound evidence upstream and fail the build when it changes; figure geometry is asserted by scripts/verify_figures.py; and CC-001 + CC-004 (bounds and endpoint witnesses) are re-reproduced from a clean clone on every push and weekly. What v0.2 merely asserted is documented in Noetic Log 001. Since 2026-09-02 a claim's commitment (proposition, scope, falsifier, forbidden rescues, non-claims, expected values) can change only by appending a declared transition to claims_history.yaml; scripts/claims_history.py verify fails any undeclared change, any edit to accepted history, and any re-minted baseline, from that genesis forward — it protected nothing before it existed, and it does not bind an actor who can change the verifier, the workflow, or the history in the same act.

Run locally

python3 -m http.server 4173

Then open http://localhost:4173.

MJGD v1

Minimum Joint Guardrail Disclosure v1 is a machine-readable disclosure schema for a declared multi-guard evaluation. Its validator recomputes only complete static full-exposure outcomes and complete positive-set aggregate pattern tables, returns identified sets for marginals, and holds routes and missing cells rather than guessing. The JSON Schema is structural; the CLI performs semantic conformance checks. MJGD is a schema, not a safety standard or an adoption claim.

Route receipt stub

examples/route-receipt/ is the separate, portable two-file receipt for a declared item-level route. It emits only policy actions from a direct route trace and returns HOLD for a post-hoc static reconstruction. It is a template, not an adoption claim or a deployment certificate.

Deploy & verification

Open a PR, pass the verification workflow, and merge through the reviewed path. The workflow verifies the claim registry — field shape (including falsifiers and forbidden rescues), support-link liveness, ledger coverage, and a freshness gate that fails when a claim passes its review window (also run weekly) — then deploys the exact verified static artifact and checks the live checksum, sitemap, rendered ladder, and correction policy.

Configured deployment control. GitHub Pages publishes through GitHub Actions, and the github-pages environment permits only the default branch. The branch-based Pages publisher is disabled: deployment waits for the claim and reproduction gates, then the workflow smoke-tests the live artifact.

Layout

index.html                  the record's front page (feasible-worlds instrument inline)
essays/when-marginals-are-not-enough/   flagship case study (real kernel output)
modules/                    module system — GENERATED from modules.yaml
observatory/                claim observatory — GENERATED from claims.yaml
ledger/                     evidence ledger — GENERATED from claims.yaml
writing/  archive/  now/    writing index · intellectual lineage · current work
notes/noetic-log-001/       public audit log: what v0.2 pretended to implement
resume/                     web résumé with 90-second overview
stack-study/                browser-local study preflight; static / route / adaptive scopes stay distinct
claims.yaml                 claim registry (schema v0.4) — the source of truth
modules.yaml                module registry — questions, status, bindings
scripts/generate_ledger.py  registry → ledger renderer (CI drift-checks it)
scripts/generate_modules.py module registry → module pages (CI drift-checks)
scripts/generate_observatory.py  claims.yaml → observatory (CI drift-checks)
scripts/verify_claims.py    registry verifier: bindings, triggers, freshness
scripts/verify_figures.py   figure geometry assertions (Fig. 02 + essay)
scripts/verify_frontend.py  static frontend structure + local-only preflight gate
scripts/reproduce_cc001.py  clean-clone reproduction of CC-001 + CC-004 (disposable venv)
try/                        the experiment surface — GENERATED from distribution/experiments.yaml
scripts/try_same_scores.py  TRY-A: two worlds from the same marginals, standard library
scripts/try_audit.py        TRY-C: the disclosure test on an evaluation you know
distribution/               outcomes ledger, experiments, launch units, dossiers, external-events procedure
contrib/                    prepared joint-statistics reporters for two public harnesses (patches, unsent)
films/                      six deterministic evidence-bound films (see films/README.md)
scripts/validate_mjgd.py    MJGD v1 packet validator + fixture/refusal tests
schemas/mjgd-v1.schema.json MJGD v1 schema documentation (validator is the contract)
fixtures/mjgd-v1/           illustrative complete, aggregate, marginal, route, and hold packets
examples/stack-joint/       portable static-OR CSV receipt stub + fixture
docs/MJGD_V1.md             MJGD v1 semantics, replay commands, and non-claims
404.html                    not-found page
assets/                     shared stylesheet, self-hosted fonts, images
DESIGN.md                   design-decision ledger + changelogs + field-artifact notes

Content © Pranav Bhave. Code (HTML/CSS/JS) may be reused with attribution.

Scheduled feedback and release updates

The existing local schedules record repository state daily and collect distribution observations every four hours. Missing days and missed observation windows remain missing; the jobs never interpolate evidence. Both jobs share a repository lock. They prefer .venv/bin/python3; create that environment with Python 3.12 or newer and install requirements.txt before enabling the schedules. CUBITS11_PYTHON can select another environment explicitly.

A cycle starting from clean, synchronized main runs the verification manifest, commits only its permitted outputs on a claude/cycle-* branch, and opens a pull request. GitHub auto-merge uses a merge commit and remains subject to required checks and review rules. A failed API call leaves the branch available for recovery; a failed check retains the observations without publishing them. Logs live in _private/cron/. No cycle dispatches a new outreach message.

This feedback loop records observations and tests known failures. It does not autonomously change scientific criteria, promote a hypothesis, or accept a contradicted result. Deployment checks compare the public pages, primary films, posters and claim metadata with the verified revision, so an older page returning HTTP 200 cannot stand in for the release.