A visual simulation. Inspect exactly how each input changes its modeled state.
Score rules v3 match whole Unicode words and flag a direct not-word mismatch against earlier memory. The score is a scripted simulation; v1 and v2 receipts remain verifiable from this repository.
μ is a scripted score, not a measure of a person or of consciousness. The breakdown uses prior memory and messages only; this input cannot match itself.
No input has changed the model yet.
Memory ablation lab: compare a draft against the same model with earlier memory and messages withheld. Both start at the current μ. This isolates the next step only; it does not replay a history without memory.
Test whether the scripted response with prior history is more useful to a reader. Enter 1–8 context lines and a final prompt. A dedicated reader view presents A and B in random order without branch labels or scores until a choice is made. For a stronger test, ask someone who did not write the context to rate them.
No plan loaded. You can run a custom trial below.
No text or rating is saved unless you download the report. A single subjective choice is not evidence of a general benefit; downloaded reports include your context and prompt.
Enter one input per line. Both branches start at μ = 0.6400; the control discards prior memory and messages at every step. A fixed one-second clock makes the result reproducible. This experiment does not change saved state.
Run the same scripted model from a fresh start for repeated, unrelated, contradictory, and directly negated phrases. These fixed controls probe positive, zero, and negative history effects. Your saved field stays untouched.
Submit an input to create a portable proof of its score change.
Your entries stay in this browser unless you download them. Export includes text you entered.
This replaces this browser's score, memory, messages, and evidence records after review. Download the current state first if you want to keep it. Prior proof and last-change displays are cleared on restore.
Choose a current-model transition proof JSON. This browser replays the score change and the no-history control from the file's input and prior state. Your own field and ledger stay untouched; the file is not uploaded by this app.
No score proof checked yet.
Older v1 and v2 files require the repository's command-line verifier. A pass checks this scripted model's arithmetic, not who made the file or whether μ measures anything in the world.
Select one reference and 1–20 original receipts in their intended order. A separately kept checkpoint is optional. These files stay on this device.
Choose a reference and ordered receipts.
No audit completed on these files.
Audit first, then compare a saved capture if you have one.
After auditing, withhold one reference record at a time from a temporary copy. See whether each receipt's verdict or computed answer and citations change. The original files and saved capture are untouched.
Run the independent command on these same original files, in the same receipt order: node tools/review-source-probe.cjs reference.json receipt-1.json ... --json > independent-review.json. Add --checkpoint checkpoint.json if selected above. Choose the saved JSON here. This page checks the file hashes and every source-removal result against its current audit. Step-by-step guide
A matching map means this browser agrees with the supplied report. JSON can be edited: this page cannot prove the Python checker ran or who supplied the files. Run the command yourself before relying on an independent comparison. No selected file is uploaded by this app.
A verdict-only flip means the receipt's exact snapshot changed while its computed answer and citations stayed the same. These are hypothetical removals from the supplied reference, not evidence that any real source was omitted. An altered copy would no longer agree with a supplied checkpoint. No single-source result establishes completeness, independence, or truth.
Outside this review: who authored or independently held the files, whether relevant evidence is missing, and whether the underlying claims are true. A capture can replay as MATCH while a receipt is DIFFER or FAIL.
See where a self-consistent answer can still mislead. These fixed examples run through the same receipt, reference, and checkpoint checks as uploaded files. They do not use or change your personal ledger.
An earlier checkpoint matters only when kept independently. This lab tests software consistency under synthetic inputs, not whether any real-world statement is true.
Read five synthetic audit summaries and choose the conclusion each supports. The answers stay hidden until the end. This exercise never reads your evidence files or browser memory and does not save your choices unless you explicitly download them.
A score measures answers to these five examples, not expertise, real-world claim accuracy, or an independent study. The optional file contains scenario IDs, choices, a score, ordering method, and the page script fingerprint; it can be edited by its holder.
Choose a question, brief, correction, correction-chain, or compact correction packet JSON. This browser replays the supplied change; a compact packet is checked against this browser's own ledger. The file is read on this device and is not uploaded by this app.
No receipt checked yet.
Verify a question, excerpt, or brief before comparing it with this browser's evidence.
A full receipt or brief must match this browser's entire ordered ledger. An excerpt can match its answer and cited records while unrelated local entries remain private. A matching local copy still cannot establish who recorded the assertions or whether they are true.
After a question, excerpt, or brief passes above, choose a reference ledger file. Add a checkpoint you preserved separately to check that the reference has not changed. These files are read on this device, are not imported, and are not uploaded by this app.
Verify a question, excerpt, or brief first.
A match depends on the reference you supplied. A checkpoint adds evidence only if you kept it independently before receiving the later files; it does not prove when it was made, who recorded the facts, or that they are true.
Use the reference and optional checkpoint selected above. Check up to 20 question, excerpt, or brief receipts together. Every file gets its own replay and reference comparison; identical file bytes are flagged as a repeat. A JSON file with repeated object names is ambiguous and fails instead of silently choosing one value.
Run the audit before downloading a capture.
Run the audit, choose a saved capture, then compare it with these original files. CAPTURE MATCH means the files and recorded results reproduced; check the audit counts separately for DIFFER or FAIL rows. If the capture differs, this page lists changed components without showing the private facts. To preserve this version's verifier code for later replay, see the offline verifier kit guide. A repeated file is not a new independent source. The capture contains hashes and outcomes, not record values or filenames. No files are uploaded or imported.
Keep the selected reference and receipts above as B. Choose a second reference as A to replay those same receipts against both. Optional checkpoints are checked against their respective references before any verdict is compared. The view names changed source IDs and fields, answer kinds, and citation changes without displaying fact names or values. For up to six added sources, each receipt gets separate minimum source sets for the verdict and computed answer. A full snapshot can change verdict with no answer change.
Trace the selected files before downloading a capture.
After tracing, prepare a capture with keyed file commitments instead of public file hashes. It still reveals the verifier version, verdicts, counts, and repeated-file positions. Keep the opening key separate from the capture: anyone with both can test guesses about private facts. This is not encryption, authorship, or proof of truth.
Trace the original files to prepare a private pair.
Trace the original files, then choose both saved files to compare.
The answer-change test finds the smallest added-source sets that reproduce B's computed answer when B only adds up to six sources. It reports when edits, removals, order changes, or a larger search prevent that conclusion. A and B are comparison labels, not verified dates. An unchanged verdict can conceal a different reason. A checkpoint checks the ordered records and next source ID under these rules; a trusted copy held separately is needed for provenance. The capture contains hashes and outcomes, not fact values, but predictable private assertions may be guessable from a hash. Neither reference proves the assertions are true or complete.
A correction can be applied only to the exact starting ledger with the same next source ID. Review its current sources and replacement first. Diverged browsers must be reconciled separately; this action never silently merges them.
Select 2–8 correction receipts in any file order. This device checks their exact continuity. Download the full chain for an independently replayable record, or a compact packet for a matching browser. The compact packet includes changed names and values plus ledger hashes, without unrelated names and values; it requires the original ledger to verify.
A pass checks consistency with supplied records or the matching local ledger. It cannot authenticate their author or confirm the facts. A compact packet's hash is not encryption or proof that an event happened.
Record a fact under a name, then recall it with a source. Questions use two narrow templates and exact stored names; this ledger does not interpret the freeform conversation or verify assertions against the world.
Enter 2–8 exact fact names, one per line. The brief gives each a sourced value or an explicit refusal. It leaves your ledger and modeled score alone.
Brief receipts include the full ledger snapshot, including unrelated facts. Review the JSON before sharing it. Verification checks the supplied records, not their completeness or truth.
Ask about a named fact, or record one below.
For a sourced answer, download an excerpt containing only its cited assertions. It cannot show whether other ledger entries were left out. The full question receipt also supports refusals, but includes your entire ledger and may expose unrelated facts.
No fact has been recalled yet.
Use the fact name and new value above. Review the current sources and replacement before applying. The old assertions remain visible as superseded records.
Correction receipts contain the full ledger before and after the change, including unrelated facts. They check the recorded transition, not whether either assertion is true.
Download a reference ledger to compare shared receipts in the command-line verifier. It includes all records and the next source ID, so keep it private. Save a compact checkpoint separately before receiving a later reference: its digest can detect changes to that reference without revealing its records directly.
A checkpoint is not a timestamp, signature, encryption, or proof of a fact. A guessable ledger can be tested against its digest. Keep the checkpoint independently of any later files you check.
Review every entry and its new local source number before applying. Incoming current claims can create conflicts; nothing is silently replaced.