Classified: suspiciously public mathematics
Seeded Gaussian z tests. A twenty-agent campus. A traceable case file.
Local records must verify before calculations can resume.
Verification replays every saved family and can take time for large experiments. Stopping leaves saved records untouched and keeps calculation controls locked. Retry verifies the complete chain again; an unverified recovery export preserves readable records.
Cancel discards an unfinished family without a ledger block or advancing its seeded family index. Already completed results still finish their durable commit. Pause stops after the current family.
01 / Live research campus
Gaussian synthesis, multiple-testing analysis, and a local vault, seen from above.
Local time
No batch yet.
No batch yet.
No batch yet.
No batch yet.
No batch yet.
No batch yet.
No batch yet.
No batch yet.
No batch yet.
No new mail. Restored history does not replay deliveries.
Two crew members collect working input dossiers from room mailboxes and carry them along stone paths as actual work progresses. Completed-stage and durable-save receipts appear immediately. Fast intermediate poses can be coalesced into the latest receipt; animation never delays calculations or creates a completion backlog. These are animated characters, not independent AI agents. Pausing freezes poses; reduced motion keeps static poses.
Historical cameos recall ideas later revised or rejected. Honest errors and useful models are part of scientific progress. These stories are not allegations of misconduct, and they are distinct from the statistical false discoveries counted here.
A fictional literary nod to the writer in Kurt Vonnegut's work, drawn here as an original pixel analyst.
Book reference: Breakfast of ChampionsA fictional literary nod to Vonnegut's extraterrestrials, drawn here as an original pixel analyst.
Book reference: Slaughterhouse-FiveReported N-rays, later debunked. This historical mistaken finding differs from a defined statistical false discovery.
Source: American Physical SocietyAdvocated Martian canals, an interpretation later rejected. This historical mistaken finding differs from a defined statistical false discovery.
Source: NASAA Flying Spaghetti Monster liaison with a pasta badge. This is a separate satire cameo, not an empirical discovery or statistical false discovery.
Source: Henderson's official FSM siteProposed an incorrect three-stranded model of DNA before the double helix was established.
Source: Science Museum GroupProposed the hypothetical planet Vulcan to explain an anomaly in Mercury's orbit.
Source: Washington UniversitySpeculated that even the Sun could be inhabited.
Source: University of OxfordDeveloped an early atom model with distributed positive charge, later superseded by nuclear models.
Source: Science History InstituteUsed a cooling-Earth model that substantially underestimated Earth's age.
Source: Cambridge repository (PDF)Co-developed steady-state cosmology, later displaced by evidence supporting a hot early universe.
Source: American Institute of PhysicsIsolated oxygen but interpreted it through the subsequently abandoned phlogiston theory.
Source: Science History InstituteDeveloped phlogiston as an explanation of combustion, before oxygen-based chemistry replaced it.
Source: University of IllinoisInterpreted his willow-tree experiment as showing that its added substance came from water alone.
Source: Le Moyne CollegeProposed invisible passages through the heart's septum, an anatomical explanation later rejected.
Source: University at BuffaloConstructed an influential Earth-centered model of planetary motion, later superseded.
Source: NASAProposed a hybrid system in which the planets orbited the Sun while the Sun orbited a stationary Earth.
Source: NASAInterpreted microbial growth in heated broth as evidence of spontaneous generation.
Source: Northern Arizona UniversityExplained colors through combinations of lightness and darkness, an account later displaced by modern optics.
Source: Smithsonian LibrariesProposed that material things were composed of four basic elements: earth, air, fire, and water.
Source: Getty Research Institute (PDF)Not started02 / Mission setup
Optional / Crew conversation
Fictional personalities learning to avoid false discoveries.
Test sends your key directly to Groq, then Enable permits synthetic experiment summaries, current crew positions and short dialogue memory to be sent there. The key stays in this tab's memory and is cleared by Clear or reload. It is never saved to the ledger, exports or browser storage. Browser scripts and developer tools can access a browser key; Groq recommends a trusted backend for production secrets. This optional direct connection uses your account's limits and never enables billing or switches models automatically.
One conversation request contains five exchanges: ten alternating messages. Every request, including Test, has a thirty-second minimum gap and a fifteen-second timeout. Rate limits pause chatter for manual resume. Model: openai/gpt-oss-20b. Calculation and saving continue independently.
The next provider request waits this long after a complete conversation. All ten messages arrive together in one response; this setting changes request frequency, not individual message display speed. Test and the first enabled conversation have no extra two-minute delay; the thirty-second request gap still applies. This preference stays in this tab only.
Chatter is off. No conversation request is scheduled.
Disconnected. No provider requests.
No conversation pair yet.
The last five complete conversations stay in this tab until Clear or reload. Commentary can be mistaken. The numerical results and ledger remain the source of truth. Names are fictional cameos, with no historical endorsement or proof-validation claim. After five exchanges, the next partner is chosen from current positions; recent pairs are avoided temporarily.
03 / Intelligence
Current run · completed batches only
FDR control relies on independence or suitable positive dependence assumptions.
Harmonic correction supports arbitrary dependence with valid marginal p-values.
Power is the mean batch true-discovery fraction (N/A with no alternatives). MC standard error is for the mean FDP across independent replicated families; tests within a common-factor family remain dependent. Normal-approximation 95% intervals appear after 30 families. They are not sequentially valid under optional stopping; zero estimated variance does not prove zero uncertainty.
Monte Carlo mean of completed batch FDPs
A finite simulation fluctuates; it does not prove a general FDR guarantee.
−log10(p), smallest p first
Dashed curves: BH rank × q/m and BY rank × q/(m Hm), transformed to −log10(p). Smaller p sits higher; the step-up rule finds the largest qualifying rank. At most 1,200 actual ranked tests are displayed; every test contributes to the saved results. Truth labels are synthetic and known.
04 / Chain of custody
Single local chain. A fork preserves parent history and records its verified parent tip.
No blocks yet.| Block | Run / batch | BH R / V | BH FDP | BY R / V | BY FDP | SHA-256 |
|---|---|---|---|---|---|---|
| ∅The vault is empty. Start the investigation to seal the first case. | ||||||
Previous-hash links expose changes within a chain. They do not establish authenticity if an entire chain is rewritten. No consensus or cryptocurrency is involved. Import requires an empty vault and checkpoint; existing records are never replaced.
05 / Dispatch
06 / Declassified notes
Known truth. Each batch contains synthetic two-sided Gaussian z tests with known variance 1. Null tests have mean 0; alternatives use the configured mean. The null count is round(m π0); the actual fraction may differ slightly from the entered fraction. When the alternative mean is 0, all m hypotheses are truly null, and power is N/A.
Xij = μi + √ρ Cj + √(1 − ρ) εij
Cj and εij are independent standard Gaussian draws, with Cj shared across tests at observation j. Set ρ = 0 for independent tests. Compute zi = √n X̄i and pi = 2Φ(−|zi|). Seeded PRNG replications use new draws across batches, representing independent replicated families.
BH. Order the m p-values. Find the largest rank k satisfying p(k) ≤ kq/m and reject through that rank. The standard guarantee requires independence or suitable positive regression dependence (PRDS) assumptions.
BY. Use the same step-up rule with q/Hm, where Hm = ∑j=1m 1/j. This harmonic correction controls FDR under arbitrary dependence when marginal p-values are valid.
Dependence. Common-factor mode uses nonnegative ρ and preserves valid Gaussian null marginals. Positive correlation of z statistics alone does not establish the required PRDS condition for these two-sided p-values. This dashboard makes no blanket BH guarantee in that mode.
FDP versus FDR. The realized batch false discovery proportion is V/max(R, 1). Its mean over completed batches in the current run is a Monte Carlo estimate of FDR. Power is the mean fraction of actual alternatives correctly rejected per batch. These finite estimates do not prove general guarantees.
Monte Carlo uncertainty. The standard error is the sample standard deviation of batch FDP divided by the square root of the completed batch count. Approximate 95% intervals use a normal approximation after at least 30 batches. They are not sequentially valid under optional stopping; zero observed variance is not proof of zero uncertainty.
Method reference: R stats: Adjust P-values for Multiple Comparisons , including the original BH (1995) and BY (2001) references.
Capture has no fixed count ceiling; test count and sample size have no preset maxima. Actual browser memory, engine allocation limits and storage quota still constrain work. The foreign-file importer separately limits unknown files to 64 MiB, 10,000 blocks and 25 million observations of replay work. A very large full export from this app may therefore require a compatible bulk loader or migration. Saved history is never silently truncated to fit these import limits.
Calculations and IndexedDB storage stay in this browser. A seed and unchanged parameters reproduce the synthetic sequence. Parameter changes create a distinct run; estimates never pool different runs.
Each completed batch is hashed with SHA-256 and linked to the previous saved hash. Export preserves the full ledger. Import validates the file before it can be accepted. Retain an export if you need a durable copy: browser storage can be cleared. Legacy v1 and current v2 ledger files can be validated.
A verified-chain fork preserves all parent block history and adds a provenance block linking to the verified parent tip. Future batches extend this independent local branch. Branches are not automatically merged or synchronized; there is no consensus mechanism.
Download the local edition · source, launchers and tests
For the downloadable edition, unpack the ZIP and follow its README to launch from a local HTTP origin. Browser storage belongs to that origin.
The browser must remain open. Hidden tabs pause explicitly; closed or throttled browsers do not promise continuous execution. A storage failure pauses work and shows a warning. A block is sealed only after its save succeeds.
FICTIONAL PROP / ZERO CLEARANCE REQUIRED
FAKE-NULL-0000-DO-NOT-USEDefinitely don’t come here.
The null hypotheses might confess.
Fictional decoy. No real credentials, wallets, or visitor tracking. The investigation is entirely synthetic mathematics.
Agents Kilgore Trout and Tralfamadorian are fictional literary nods, drawn here as original pixel analysts. They personify calculations; they are not autonomous AI agents.
René Blondlot reported N-rays, later debunked (American Physical Society). Percival Lowell advocated Martian canals (NASA). These historical mistaken findings differ from the precisely defined statistical false discoveries counted here. Neither scientist used BH or BY.
Agent Bobby Henderson, wearing a small pasta badge as the Flying Spaghetti Monster liaison, is a separate satire cameo. Satire is not an empirical discovery or a statistical false discovery.