AI Visibility Blinded Review Packet Generator

The short answerFreeze one answer-evidence CSV before labeling. Keep the evidence reviewers need, conceal only predeclared outcome metadata, give each reviewer every eligible row in a separately derived order, and restrict the reidentification key until both independent passes lock.

Selected masking is not proof of blindnessThe target, prompt, answer, and citations remain visible. Those fields can reveal period, system, intervention, or expected result. This page documents a masking attempt and residual cues; it does not certify that anyone was blind.

Preserve needed evidence, conceal selected metadata, keep the key separate

Same eligible evidenceBoth packets contain every eligible response exactly once, with blank mention, citation, recommendation, confidence, abstention, and note fields.

Different derived ordersA versioned deterministic PRNG derives separate reviewer-A and reviewer-B shuffles and opaque packet IDs from one declared seed.

Residual cues remainVisible answers and citations may reveal what metadata masking cannot. Possible cues are counted; known unblinding excludes a row from both packets.

Browser-local pre-labeling control

Generate the two packets and restricted key

Nothing is uploaded. The source is held only in this tab, and every displayed output hash covers the exact bytes offered for download.

01 · Protocol identity

Name the packet, people, codebook, and intended use

These declarations enter the protocol memo and hash records. They do not verify a person’s identity or an external codebook file.

Packet set name

Protocol version

Organization

Coordinator

Restricted-key custodian

Reviewer A label

Reviewer B label

Deterministic text seedReusing the exact source, declarations, and seed reproduces the packet bytes.

Codebook name

Codebook version

Frozen endpoint-rule declaration

Purpose and decision boundary

02 · Frozen masking plan

Declare what reviewers will see before packet generation

The target, prompt, answer, and citation evidence are always visible. These controls govern only engine/surface and run timestamp.

Engine/surface treatment

Run timestamp treatment

Masking-plan owner

Masking-plan lock date

Independent first passesReviewers will work separately and will not see each other’s fields until both passes lock. Key access restrictedNeither reviewer can access the mapping key during the first-pass review. Outcome metadata withheldPrior labels, expected outcomes, period, arm, collector note, and evidence reference stay out of reviewer packets. Plan locked before releaseThe masking choices, codebook declaration, and source file will not be changed after packet release.

03 · Frozen answer evidence

Import the exact 17-column source CSV

Each response ID must be unique. Excluded and unblinded rows remain in the coordinator key and counts but never enter either reviewer packet.

Answer-evidence CSVValid UTF-8, exact header and order, no more than 5 MB or 5,000 data rows.

No file selected.

Local processing. Parsing, exact source hashing, deterministic shuffling, masking, preview, and downloads happen in this browser. Reloading clears the source and generated outputs.

Packet set prepared

Two blank reviewer packets are ready

Source rows—

Packet rows each—

Excluded/key only—

Possible residual cues—

Exact source and output hashes

Exact source file SHA-256—

Reviewer A CSV SHA-256—

Reviewer B CSV SHA-256—

Restricted key CSV SHA-256—

Algorithm record SHA-256—

Codebook declaration SHA-256—

Masking-plan record SHA-256—

Protocol memo SHA-256—

Residual-cue and process warnings

Packet manifest

Reviewer A packet preview

First eight reviewer A rows; classification fields are blank
Packet ID	Target	Exact prompt	Answer text	Citation evidence	Engine/surface	Run timestamp	Mentioned	Cited	Recommended

Reviewer B packet preview

First eight reviewer B rows; classification fields are blank
Packet ID	Target	Exact prompt	Answer text	Citation evidence	Engine/surface	Run timestamp	Mentioned	Cited	Recommended

Coordinator protocol and manifest

Exact source CSV schema

The header is case-sensitive and must appear in this order. The generator does not guess renamed columns or silently choose among duplicate IDs.

Required 17-column answer-evidence source
Column	Reviewer visibility	Rule
`response_id`	Key only	Unique and nonblank.
`target_entity`	Visible	Required for eligible rows.
`exact_prompt`	Visible	Required for eligible rows; never rewritten.
`answer_text`	Visible	Required for eligible rows; markup is rendered only as inert text.
`citation_evidence`	Visible	Required; the literal value `none` is allowed.
`engine_surface`	Plan-controlled	Visible, deterministically aliased, or concealed.
`run_timestamp`	Plan-controlled	Visible or concealed.
`collection_period`	Key only	Concealed from both reviewer packets.
`comparison_arm`	Key only	Concealed from both reviewer packets.
`prior_label`	Key only	Concealed; may be blank if none exists.
`expected_outcome`	Key only	Concealed; may be blank.
`evidence_reference`	Key only	Replaced by the packet ID in reviewer files.
`collector_note`	Key only	Concealed from reviewers.
`review_eligibility`	Key only	Exact value `eligible` or `exclude`.
`exclusion_reason`	Key only	Required when eligibility is `exclude`.
`residual_cue_status`	Key only	Exact value `none known`, `possible`, or `unblinded/exclude`.
`residual_cue_note`	Key only	Required for `possible` and `unblinded/exclude`.

Method, export behavior, and limits

The algorithm uses one versioned, deterministic text-seed function and Fisher–Yates shuffle. Reviewer A and reviewer B use separate derived streams. Opaque packet IDs are also stream-derived and are joined back to the source only in the restricted key. Determinism supports reconstruction; it does not make the order cryptographically unpredictable.

The codebook-declaration hash covers the codebook name, version, and endpoint-rule text entered here. It does not hash or authenticate an external codebook file. The masking-plan hash covers the declared modes, owner, date, purpose, and process confirmations. The source hash alone covers the exact imported file bytes.

Spreadsheet-oriented export defenseEvery CSV field is enclosed in double quotes. If the first non-whitespace character is =, +, -, @, ＝, ＋, －, or ＠, the exported value receives a leading horizontal tab inside the quoted field. This intentionally changes the downloaded cell, is counted in the memo, and is not a universal guarantee across spreadsheet programs or downstream transformations.

The generator does not upload, store, email, or share packets; identify reviewers; inspect hidden file metadata beyond the parsed CSV; redact sensitive content; call an AI model; label an answer; measure agreement; compare to a reference standard; adjudicate; prove independence; prove masking success; certify compliance; or support a causal treatment-effect claim.

Method sources and implementation context

Rayyan blind-mode documentation illustrates concealed independent decisions before conflict resolution in a two-reviewer workflow.
Potato annotation documentation documents configurable, seeded assignment and randomized annotation workflows.
JAMA: Effect of Blinding and Unmasking on the Quality of Peer Review reports that textual and contextual clues can defeat masking even when identifying fields are removed.
ASHA Journals anonymized-review policy inventories identity leakage through text, acknowledgments, comments, revisions, properties, and filenames.
OWASP: CSV Injection explains why ordinary quoting does not by itself prevent spreadsheet formula interpretation.

Frequently asked questions

What source CSV does the packet generator require?

It requires the exact 17-column answer-evidence schema shown above, with one unique response ID per row, declared review eligibility, and a residual-cue status. Imports are limited to valid UTF-8, 5 MB, and 5,000 data rows.

What evidence stays visible to reviewers?

The target entity, exact prompt, answer text, and citation evidence stay visible because reviewers need them to label mentions, citations, and recommendations. Engine and timestamp treatment follow the masking plan.

Does shuffled metadata make a review blind?

No. Shuffling and selected field concealment can reduce some cues, but answer content, citations, phrasing, and subject matter can still reveal period, surface, intervention, or expected outcome.

Why generate two independent packet orders?

Separate deterministic streams give reviewer A and reviewer B different row orders while preserving the same eligible evidence set. If the two derived orders happen to match, the generator warns instead of silently trying seeds until it gets a preferred result.

Where is the restricted mapping key stored?

Nowhere by this page. It is generated locally for the coordinator to download. The page does not upload, retain, share, email, or verify access to the key.

Are formula-like CSV values changed?

Only downloaded CSV cells whose first non-whitespace character is a formula trigger receive a disclosed leading tab inside an always-quoted field. The browser view preserves the supplied value, and no defense is universal across every spreadsheet importer.

After both passes lock

Verify reviewer returns before agreement QA

Reconstruct both blank packets, verify the exact restricted key and mappings, and preserve both original label sets before calculating agreement.

Check reviewer return integrity