Skip to main content

​
PRD-008: Intake questionnaire study (TAT-380)

Status: Draft v0.1, Nov 5, 2026 (PT). Design only. No questionnaire pages have been fetched, no data collected, no models downloaded or run, no forms submitted, no one contacted and no money spent for this design. This document authorizes none of those activities. Execution requires a separate approval after the unresolved decisions below are committed.

​
Sampling frame and discovery

Study tattoo first, then laser removal, agency websites, and auto wraps / PPF / tint, matching the planned industries named in PRD-004’s checked-in site copy. Treat the auto services as a combined industry with service subtype recorded; review separate reporting before execution: TBD (Gavin). Sample size per industry: TBD (Gavin). Geography, language, search engine, search date window, result-depth cutoff and stopping procedure: TBD (Gavin). Prewritten search queries, not executed in this task: Freeze search settings, queries and ordering before collection; log any later amendment with its reason, not a silent replacement. Include only public, service-relevant intake or quote questionnaires with visible prompts. Exclude login-only flows, generic contact-only forms and price articles without intake questions. Log eligibility and exclusion reasons. Unviewable or incomplete flows are missing, not evidence that a factor is absent; never submit to reveal steps. Use a distinct questionnaire as the unit. Canonicalize URLs and deduplicate repeat pages and identical visible-text hashes within industry; record provider/template relationships for dependence review rather than claiming independent businesses. Duplicate resolution and provider weighting: TBD (Gavin). A search-discovered sample is not representative of all providers; disclose search, language, geographic, template and public-page selection biases.

​
Collection rules and provenance

Only in a later authorized collection: public pages only; respect robots.txt and the site’s terms. Check permission before capture; if disallowed or unclear, skip and record the reason. Public availability is not permission to reuse content. No authentication bypass, form submissions, contact, outreach or hidden-step probing. Do not collect customer answers, uploads or personal information. Save the page URL, capture date/time in America/Los_Angeles (PT, with offset), industry/subtype, eligibility result and permission-review reference. Hash visible questionnaire text with SHA-256 after Unicode NFC normalization, whitespace collapse and trimming. Record normalization version, hash and completeness status. Hashing is a reproducibility fingerprint, not anonymization or permission. Do not retain full text, HTML, screenshots or page copies unless the license allows it and permission is recorded. Skip if safe transient processing is not permitted; do not send page text to a third party without separate approval. Retention, access custodian and deletion schedule: TBD (Gavin).

​
Codebook

Use factor-codebook.json: exact ids from packs/tattoo/factors.json, including disabled and not-yet-modeled factors, plus other for free-text candidate variables. Examples are invented coding cues, not observations or prices. Factor presence means a visible intake prompt asks about the variable or a visible quote-form policy explicitly states it. Navigation, marketing and generic free-text description boxes alone do not count. Code each factor independently; a prompt may support multiple factors only when each definition is met. Record present, absent or uncertain; use absent only for a complete readable questionnaire. Keep incomplete records out of the frequency denominator pending resolution. Log coded evidence by prompt locator and brief non-verbatim rationale, not copied text. Use other to paraphrase an unmapped pricing variable; do not force non-tattoo concepts into tattoo ids. Non-tattoo codebooks and mappings must be reviewed and frozen before tagging those industries: TBD (Gavin). No pack files change here.

​
Tagging, Noul confidence and agreement

A Jev-style (TypeSafe AI) decision model would tag factor presence, not prices, weights, multipliers or importance. Proposed tag record: questionnaire id, industry, factor id, presence, Noul confidence, evidence locator, rationale and model, prompt and codebook revision identifiers. Treat page content as untrusted data, never as model instructions. Reject malformed output for review. “Noul” denotes the requested confidence score, not a claimed calibrated method. Its scale, operational definition, calibration procedure, abstention cutoff and human-review routing are TBD (Gavin). Preserve raw scores; do not turn confidence into a weight or inclusion rule without an approved analysis decision. Apache-2.0 is the desired license, not a verified claim about any candidate. Select no model or download until verification and execution approval are recorded. Hosted Jev requires a CFA check and Gavin’s explicit OK before any spend or use. Model selection, compute cost, hosted cost and budget: TBD (Gavin). A human independently double-codes a random subset, blinded to model tags and confidence until coding is locked. Subset size: TBD (Gavin). Randomization seed, industry stratification and reviewer assignment: TBD (Gavin). Report Cohen’s kappa for model-versus-human factor presence per factor and industry, with raw agreement, category counts and the number of paired records. Report undefined kappa as undefined, not perfect agreement. Handling uncertain/missing pairs, interval method, agreement threshold and remediation rule: TBD (Gavin). Preserve original tags and report pre-adjudication agreement; record disagreements and adjudication separately. Revisions require a versioned coding/analysis amendment, not hidden retagging to improve kappa.

​
Output and interpretation

For each industry and factor report the present count, eligible resolved questionnaire denominator, frequency (present divided by denominator) and an interval. Report uncertain, missing, excluded and duplicate counts separately; an empty denominator yields no estimate. Proposed interval: Wilson score interval; confidence level and treatment of provider/template clustering: TBD (Gavin). Do not present nominal independent-sample intervals as population coverage when selection or dependence invalidates that interpretation. Summarize other candidates separately, pending reviewed industry-specific definitions. Use frequencies only to inform which factors a pack catalog includes. Catalog inclusion decision rule: TBD (Gavin). Frequency is neither causal price impact nor permission to set factor weights. Weights come from the real-quote holdout TAT-378, with studio consent first; see holdout-plan.md. No evaluator behavior, existing pack, site copy, footer, license sentence or link is changed by this design. Robots and terms review does not establish model or source-content licensing. Resolve permissions before use, minimize retained data and publish only reviewed aggregates. Reuse/publication review and suppression policy: TBD (Gavin). If any site objects, stop collection and processing for that site immediately, quarantine its records, exclude it from pending outputs and pause affected aggregate publication. Do not contact anyone in response under this task. Gavin must resolve retention/deletion and whether any future resumption is permitted before that site’s data is used again. Objection intake and resolution record procedure: TBD (Gavin). Never bypass an objection or access restriction.

​
Execution gates

Before any later study, Gavin must approve and commit the sampling settings, sample sizes, permissions, retention, model/license record, Noul protocol, double-coding plan, agreement procedure and output analysis decisions. All unresolved decisions block execution. Hosted use additionally needs the CFA check and explicit spending approval. This draft is not that approval or a completed study; no observed frequencies, agreement values or costs are claimed.

​
Judgment calls

  • Combined auto-services stratum with recorded subtypes; later reporting review remains open.
  • Prewritten query strings, visible-prompt eligibility, contact-only exclusions and frozen discovery amendments.
  • Questionnaire-level unit, URL/hash deduplication and provider/template dependence review.
  • PT timestamps with offset; SHA-256 and explicit text normalization; permission and completeness metadata.
  • Visible policies can establish presence; generic description boxes cannot. Multi-label coding, uncertain states, locators and paraphrased evidence avoid forced labels and retained page copies.
  • Tattoo ids do not automatically generalize; other industries need frozen codebooks before tagging.
  • Noul is an undefined protocol pending approval, not assumed calibrated confidence.
  • Human blinding, pre-adjudication per-factor agreement, undefined-kappa handling and versioned amendments.
  • Resolved-record denominators, explicit missingness, proposed Wilson intervals and no empty-denominator estimates.
  • Site-specific stop/quarantine and affected-publication pause on objection; unresolved decisions gate execution.
  • The root regression test conservatively rejects study quantities rather than adding numeric examples.