Data protection impact assessment: planning-register data

Published alongside our privacy notice, referenced there under "Our assessment". Companion document: legitimate interests assessment (read both together). Draft, awaiting solicitor sign-off.

1. Need for a DPIA

This processing involves: systematic, automated collection of personal data (property-linked planning records) about individuals who have not been told directly and have not consented; redistribution of that data to third parties (subscriber businesses) for their commercial use; and an automated classification step (is this scaffold-relevant?) feeding that redistribution. Together, a DPIA is clearly warranted and this document is it.

2. Description of the processing

  • What: site address, postcode, application description (free text written by the council), application type/size, decision and dates, council reference, link to the council's own file. Structured applicant-identity fields (a dedicated applicant-name or applicant-contact field, where our source supplies one separately from the description) are dropped on ingest, with no exception. The free-text description is different: it can incidentally carry a name or a sensitive detail within the council's own wording, and is handled by the screening process in section 5, not by outright exclusion - see that section for what is and is not caught.
  • Source: a third-party aggregator of UK local-authority planning registers, not the individual, and not the council directly.
  • Volume/scope: nationwide - every UK local planning authority our source covers that we hold live data for (see our coverage list), even though paying subscribers currently cover only some of those authorities.
  • Processing steps: ingest and store, then automated classification (does this look scaffold-relevant?), then daily digest assembly per subscribed council, then email to that council's subscribers. A separate, out-of-scope cold-outreach process emails prospective business subscribers about the product itself, using firms' own published business contact details, not planning-register data about them.
  • Recipients: paying subscriber businesses, one council's data to that council's subscribers only. Not sold or passed to any other third party. The AI classification call sends description/type/decision-status - not the structured address, postcode or any applicant-identity field - to our AI model provider as a processor. The description itself can still carry an incidental name or sensitive detail, for the same reason given above; this call is one of the two destinations the section 5 screen runs in front of.
  • Retention: register records are held while relevant to current digests, then archived or deleted, normally within 24 months - see our privacy notice, section 2.

3. Necessity and proportionality

Covered in full in our legitimate interests assessment (section 2): the processing is necessary for the stated purpose (there is no less-intrusive way to deliver "which newly approved jobs near you need scaffold"), and is minimised to the fields needed to identify and action the job - structured applicant-identity fields excluded outright, free-text content handled by the screening in section 5 rather than guaranteed personal-identity-free.

4. Why no direct notice to each applicant (Art. 14(5)(b))

Corrected from an earlier draft. An earlier version of this assessment said direct notice was disproportionate because we hold no name, email or phone number for the applicant. That reasoning was wrong on its own: we do hold the site address, and the absence of a name does not by itself establish disproportionate effort - postal notice to "the occupier" at a known address is a real, available route, and the assessment has to weigh it properly rather than dismiss it on a flawed premise. This section replaces that reasoning.

  • Available routes and their limits. Postal notice to the site address is possible without a name ("The Occupier"). Its limits: by the time we would send notice, the property's occupant may have changed since the application was decided (works often follow a sale or a change of tenancy); flats and shared addresses make it hard to address notice to the specific applicant rather than whoever currently lives there; some addresses are commercial, vacant, or under construction at the point of decision. Land Registry ownership data, suggested as an alternative, does not solve this either: it identifies the registered proprietor, who is frequently not the applicant (a tenant, a developer, or an agent applying with the owner's consent all commonly differ from the registered owner) - we considered this route and it does not reliably reach the right person, so we have not pursued it further.
  • Volume, four populations, not one. (1) Application records ingested: around 750,000-820,000 a year across the authorities we cover (measured 9 Nov 2026: ~67,600 in the last 30 days, ~14,500 in the last 7, annualised). (2) Distinct properties within that: checked 14 Nov 2026, 88.8% of a 30-day sample (59,696 of 67,213 records) have a distinct address. The annual figure that follows is an extrapolation from that one 30-day sample's ratio, not a separately verified annual count - roughly 665,000-730,000 distinct properties a year, if the sampled ratio holds across the year, which we have not checked month-by-month. Treat this as an estimate with the same order of uncertainty as population (1)'s own annualisation, not as a measured figure. (3) Eligible leads - what would be classified scaffold-relevant and become a digest entry once a council has a subscriber: 5.3% of ingested records in the same sample (3,539 of 67,213, ~43,000/year annualised) - this is the narrower population a targeted notice regime, aimed only at what would actually be disclosed rather than everything briefly processed, would need to cover. (4) Actually distributed to anyone today: zero - we have no live subscribers as of this assessment (see the LIA, section 3), so population (3) is a projection of what the notice obligation would cover once subscribers exist, not a count of anything that has happened yet. Figures (1)-(3) are measured from real data; figure (4) is a fact, not an estimate.
  • Estimated cost, applied to the right population each time. At a rough £0.35-£0.70 per item for addressed bulk post (an illustrative estimate from general bulk-mail pricing, not a sourced quote for this specific job, and excluding address-matching, printing and staff time): notice to every distinct property we hold a record for - population (2) above, not (1) - would cost roughly £235,000-£510,000 a year in postage alone (one letter per property, not per application, so the all-applications figure would overstate it); notice limited to population (3), the narrower eligible-lead subset, would cost roughly £15,000-£30,000 a year. The second figure is not negligible, but is closer to being within reach as the business grows - a reason this is reviewed on a trigger (below), not treated as settled permanently.
  • Residents' actual experience, honestly limited by current scale. The concern is not just one digest delivery per application, but a resident potentially hearing from several different subscribing firms about the same approved job. We have zero live subscribers as of this assessment, so we cannot measure this empirically yet; once subscriber numbers are real, this should be checked against actual per-council subscriber counts rather than assumed either way.
  • Effect on residents, and the alternative we have chosen instead. A postal notice campaign at this volume and uncertain accuracy would itself be a significant new processing exercise (matching records to current occupants, handling returned mail, complaints from people no longer connected to the property), with a real chance of causing confusion or concern disproportionate to the low-intensity processing it is notifying people about. We instead rely on: a public privacy notice describing this processing in plain terms; the redaction measures in section 5; a reachable suppression route for anyone who finds themselves in a digest and objects; and this DPIA and the accompanying LIA, published, not just held internally.
  • Conclusion. Weighing the volume, cost and limited accuracy of direct notice against the low-intensity nature of the processing and the alternative transparency measures in place, we consider individual notice disproportionate at the present scale. This is reviewed on the same trigger as the LIA (section 6 there) - in particular, if subscriber numbers grow enough that targeted postal notice to residents in specific, heavily-subscribed councils becomes practical, that should be reconsidered rather than assumed still disproportionate.
  • What this does not excuse: Art. 14(5)(b) is about the disproportionate-effort exception to individual notice - it does not relax the legitimate-interests balancing test, which is assessed on its own merits regardless of whether individual notice is given.

5. Sensitive (Article 9) content in free-text descriptions

Corrected from an earlier draft. An earlier version of this section said the brief, in-memory handling of an unscreened description was "covered by the same legitimate-interests basis as the rest of this assessment." That is not right, and we are not asserting it again here. UK GDPR Art. 6 (legitimate interests) and Art. 9 (special-category conditions) are separate requirements - having a legitimate interest in the ordinary data does not itself authorise processing special-category content that may be mixed into it, however briefly. Keeping the unscreened text out of our database reduces our exposure; it does not, on its own, answer whether the screening step itself (which necessarily reads the raw text in order to find and mask the sensitive part) has a sufficient Art. 9 basis for the moments it is doing that. We are not resolving that legal question ourselves. It is one of the two things we are sending to our solicitor for a focused assessment - see section 6 - rather than replacing this paragraph with a new assertion that it is fine.

Where raw description text actually goes

Traced directly against the code and the running system, 17 Nov 2026, not inferred from how the service is supposed to work:

StageWhat happensCould it be changed to avoid receiving/holding sensitive content at all?
ReceivedOur source's API returns the raw council-written description text to our ingest process (ingest.ts).Option A: ask the source to pre-filter or withhold descriptions containing certain terms before sending them to us (untested - see below). Option B: stop ingesting the free-text description field entirely, keeping only structured fields (address, type, dates) - would remove the risk completely but would also remove the content our classifier currently uses to judge "does this need scaffold," which is most of what makes the product work.
ScreenedThe one and only screening point in the whole system: sanitizeRecord(), called from ingest.ts synchronously, before the very first write to the database. Verified by tracing every write path (initial ingest and later status updates both route through this same function) - there is no second, unscreened write path we found.Option C: hold descriptions in a quarantine state (flagged, not yet classified or distributed) until a human reviews any description the pattern screen flags as ambiguous, rather than auto-releasing a redacted version. Not built; would slow delivery and needs a defined review process.
StoredThe post-screen value only, in applications.description (SQLite). The pre-screen raw text is not separately persisted anywhere in the live database. The exception is the 83,237 pre-9-Nov-2026 records that were written before the special-category pattern existed at all (see the backfill below) - those are now corrected in place.Option D: don't store description text at all once classification is done, keeping only the classifier's yes/no result - would remove the digest's main value (the description is what a subscriber reads to judge the job) and is not something we are proposing, listed here only because the question was to identify every option, not just the practical ones.
LoggedChecked: application error handling (site/server.js, pipeline CLI) logs only short error messages, not record content. One internal debug path (pipeline/src/cli.ts's --dry-run sample printer) would print raw description text to the console, but no scheduled production job passes --dry-run - checked against every live pm2 service's actual startup arguments, none do. A second internal debug value (decided.ts's job.sample) holds already-screened text only and is deleted from the database's own state table immediately after use. Searched all currently-retained process logs (108 files, back to 5 Oct 2026, no rotation/truncation found) for the specific terms the screen targets: no genuine match (one apparent hit was the literal string "OUTREACH_ENABLED" / "outreach disabled", a false positive, not sensitive content).No change needed - this is not a receiving/storage point, and no further screening gap was found here.
Backed upCorrects a gap in our earlier review rounds, which focused only on the ad hoc backups under pipeline/data/backups/. A separate, continuously-running nightly backup job (ops/backup, pm2 app scaffsignal-backup, 07:00 UTC daily) has been taking a full snapshot of every SQLite database - including the complete, unredacted-for-pre-9-Nov-records applications table - since at least 5 Oct 2026 (the earliest log file we found; it may run further back). Each snapshot is encrypted (age) with a key whose private half is supposed to be held only by the company's owner, off this server, then uploaded to a private Cloudflare R2 bucket: 7 copies kept locally, 14 daily + 8 weekly kept on R2. This means raw, unscreened description text - for the full pre-9-Nov-2026 population, not just the 32 records found and fixed - has been copied off this server, repeatedly, to third-party cloud storage, for weeks before this gap was even found. We have not found evidence anyone has decrypted or read one of these backups other than the owner's own monthly restore-test procedure (which decrypts to a scratch directory, checks integrity, and does not read record content) - but the channel itself is real and is corrected below, not assumed safe.Option E: exclude the description/address columns from the nightly snapshot specifically (keep everything else), trading off against the backup no longer being a complete disaster-recovery copy. Option F: keep the full backup but shorten R2's retention window. Neither has been decided; both are presented for the legal/retention decision, not selected here.
Sent to third parties(a) Our AI model provider, for classification - post-screen text only, confirmed by reading the actual call site, not assumed. (b) Cold-outreach emails to prospective subscriber firms - also reads the same post-screen applications.description column as everything else (verified directly in outreach.ts; initially mis-flagged during this review as a separate unscreened path before tracing the write path properly - it is not a separate path). Cold outreach has never actually sent a real email (gated on OUTREACH_ENABLED, unset; the outreach database's own tables do not yet exist, confirmed by querying it directly). (c) Subscriber digests - same column, zero recipients today (no live subscribers).No change needed beyond what section 5 already covers - all three read the same single-screened column.

What we are not doing: deciding which of the options above to take. They are prepared as implementation choices for legal input on how much residual risk each leaves and what it costs in product function, not a recommendation of Option A over E over D. The one we think most directly answers "could sensitive content be kept out before it even reaches us" is Option A (asking the source), which is genuinely untested - see immediately below.

What we also asked ourselves - an untested option, not a ruled-out one: could our source (the aggregator we ingest from, see our source-licensing note) strip this content before it reaches us at all, so the question never arises on our side? We have not actually asked them, so we cannot say this is impossible - that would overstate an assumption as a fact. What we do know, from the same source-licensing note: it has no commercial terms, offers no customisation we have found, and describes itself as a one-person hobby project. On that basis we think it unlikely to be a realistic route, but "unlikely based on what we know" is different from "confirmed unavailable," and we are flagging the difference rather than quietly treating our own guess as settled. We have not sent the enquiry while this document is in draft; the solicitor's assessment may want it actually tested before the point is relied on.

The screen itself (pipeline/src/sanitize.ts): titled names ("Mrs J Smith") are replaced with a marker; health/disability-related terms under UK GDPR Art. 9 (not Art. 10, which is a separate regime covering criminal-offence data and is not what this screen addresses) are replaced with a marker where they are tied to a specific person - either sitting near a pronoun/"resident"/"occupant"/"applicant", or in a phrase that itself names a person ("a disabled person," "person with disabilities," a named disability grant). A bare mention with no personal tie ("7 disabled parking spaces," "wheelchair ramp to the shop entrance") is left alone, since it describes the building, not a person. This runs before storage, so before both the AI call and any digest.

Evidence of real performance, not just synthetic tests (checked 14 and 17 Nov 2026, across more than one term, not just "disab-"): we searched all held descriptions for 17 separate health/disability-adjacent word roots (disab-, wheelchair, mobility, blind, deaf, autis-, dementia, mental health, learning difficult-, medical, diagnos-, carer, epilep-, cancer, parkinson, alzheimer, palliative, chemotherapy), not just one. The largest group, "disab-," had 201 matches; a manual read of 30 of those found the large majority were legitimate building-feature mentions (disabled parking, disabled WC, a care home's use class) correctly left alone, and 6 of 30 genuinely named a specific disabled individual. A second manual read across the other 16 word roots (not just a repeat of the first) found 3 further real cases the screen at the time missed: "elderly mobility limited relative," "occupant with mobility needs," and "disabled relative" - again, phrasing with no pronoun or "resident/occupant" framing the pattern recognised yet. All 9 real cases found across both passes are now caught, each added as its own regression test (10 tests total in pipeline/test/sanitize.test.ts). What this does and does not establish: it is an existence proof that real gaps existed and are now fixed for the specific phrasings found - it is not a measured miss rate for all special-category content, since we have not read every one of the roughly 700 matching descriptions in full, only samples of each. An untitled name, or a sensitive topic or phrasing outside what we have sampled, could still get through; we do not claim otherwise.

Backfill, 14 and 17 Nov 2026 (two passes): the screen only ran automatically on records added or changed from 9 Nov 2026 onward. On 14 Nov, 83,237 of 156,568 held records (53%) predated that and had never been through any version of it - including the "disabled resident child" case above, unredacted. A one-off pass re-applied the screen as it stood that day to every held description and address (backup taken first, integrity checked after, backup deleted once the result was confirmed correct): 29 records changed. On 17 Nov, after the broader audit above found 3 further gaps and the screen was extended to cover them, a second pass over all then-held records (160,529, having grown as new applications arrived) found and fixed those 3 specific records, following the same backup-verify-delete procedure. This is now a repeatable step, not a one-off: any future extension of the screen's patterns should be followed by the same kind of full-database check, not assumed to only affect new records.

Incident assessment: did the historical gap above cause an actual disclosure?

This flags as urgent, not as a completed finding. If any part of this turns out to be a reportable breach, the ICO's 72-hour notification clock (where it applies) runs from when we became aware, not from when this document, our own development work, or legal advice is finished. We are not waiting on either before escalating what we find below. A full dated incident record, including discovery/escalation times, is kept separately (not published on this page - see "Where this is recorded," below) and is being updated as this investigation continues, not written once and closed.

Scope: two different things, kept separate. (a) The 32 specific records identified and corrected across the 14 and 17 Nov backfill passes - for these we have checked specific channels directly, below. (b) The wider population of 83,237 records that predated the special-category screen (9 Nov 2026) - for these, we have not individually checked whether each one separately reached any channel; our findings below are about the channels as a whole (did the AI-review queue, the digest log, etc. receive *any* of this wider population, checkable because those systems log by record and by date range) rather than a record-by-record audit of all 83,237.

Checked 17 Nov 2026, with each channel's result and its limits stated rather than a single blanket finding:

ChannelStatusFinding and its limits
AI classifier queueCheckedNone of the 32 specific records, and no record from the wider pre-9-Nov population generally, was ever queued in the AI-review table (it only holds records flagged as human-ambiguous, not every ingested record) - checked by querying that table directly. Limit: this only rules out the review-queue path; see "Sent to third parties" above for the separate point that post-screen text going to the AI classifier is normal, ongoing processing, not an incident.
Subscriber digestsCheckedThe digest-sent log has zero entries for any of the 32 records, for any council, ever - consistent with zero live subscribers throughout the affected period (fact, not inference: no subscription has existed to send to). Limit: this log only covers digests our own system recorded as sent; see "outbound messages" below for a separate, lower-confidence channel.
Sample/breakdown/waiting-list requestsCheckedZero requests of any kind, ever, for any of the 23 councils the 32 records belong to.
Hosting / admin accessPartially checkedSingle-operator server; login history (last) shows only the owner's own consistent IPs, no evidence of other access. Limit: this is thin evidence (one command's output, not a full access-log audit) - we are not claiming more than it shows. No separate hosting-provider access log was available to check.
BackupsChecked - and this is the one finding that changes the pictureThe nightly off-host backup (see the data-flow table above) has copied the full, then-unredacted applications table to a private Cloudflare R2 bucket every night since at least 5 Oct 2026, encrypted, continuing through the entire period the 83,237 records were unscreened. This means the wider population's raw text did leave this server repeatedly, to third-party cloud storage, which is different from "never left our own database." We have found no evidence any backup was decrypted or its contents read other than the owner's own monthly integrity-check restore (which does not read record content) - but we are stating the channel precisely rather than folding it into a reassuring summary. Ad hoc backups under pipeline/data/backups/ (a separate, smaller set used for one-off migration work) were deleted after each correction; see the note below on why that was corrected mid-investigation.
Application/error logsCheckedNo genuine match for any special-category term across all 108 currently-retained process log files (back to 5 Oct, not rotated). See the data-flow table above for how this was checked and the one false positive ruled out.
Error reporting toolingNot applicableNo error-reporting/monitoring service (Sentry or similar) is used anywhere in this codebase - checked by searching every package manifest; none found.
Preview/staging environmentsNot applicableNo staging or preview deployment exists for this product (checked - unlike an unrelated product on the same server, which does have one). Illustrative preview emails sent during this review used separately curated example text, not real records from the affected councils.
Outbound messages (actual sent email content, as opposed to our own "was it sent" log above)UncertainWe checked our own records of what was sent (above), but have not independently checked any send-provider's own delivery log for message content, because none has ever actually sent a real digest or outreach email containing these records (zero subscribers; outreach never enabled). This line is marked uncertain rather than checked because it rests on "nothing was sent," which we are confident of for this specific population, not on inspecting a provider-side log that doesn't contain anything to find.

Where this is recorded: a full dated incident record - discovery and escalation times, affected records, the security/access circumstances above, evidence retained, containment steps, and outstanding questions with an owner - is kept in compliance/incident-record.md, not published on this public page (it would otherwise have to describe exactly where restricted evidence is kept, which is itself worth not publishing). It is available to our solicitor and is being added to as this continues, not closed off.

Evidence retained, and on what basis. Before deleting the ad hoc backups above, we extracted and separately preserved the exact pre-correction text of the 32 specific records (only those, not the wider ~160,000-record database) into one minimal, access-restricted file outside the code repository and outside routine backup rotation. Correcting a gap in our own first response: deleting those backups immediately after each fix, while this assessment was still open, was the wrong instinct, flagged by a background review before any evidence was actually lost - preserving evidence and minimising retention are not actually in tension if the specific affected records are kept and the rest is not. This file's retention is under review, not indefinite by default: we are asking our solicitor to confirm the legal basis and necessity for continuing to hold it, rather than deleting it ourselves to close the ticket, or keeping it by default past the point it's needed.

Our reading of these facts, not a legal conclusion: for the 32 specific records, and for the channels checked above generally, we have found no evidence of access by any subscriber, requester, or AI call. We have found that the wider unscreened population was, however, copied off-server to third-party cloud storage repeatedly over an extended period, encrypted, with no evidence (but not an exhaustive audit) of it having been read there. We are not concluding this was, or was not, a reportable personal data breach, nor what if anything it changes about the notification clock; that determination is for the solicitor, on these facts, treated as urgent rather than waiting for the rest of this document.

Risk table: likelihood and severity, before and after the screen

RiskLikelihood before the screenSeverity if it happensLikelihood after the screenStatus
A description names a specific person's health/disability status, reaches the AI classifier or a subscriber's digest unredactedLow per application (well under 0.5% of all applications on our sampled audit across 17 word roots, before any screening)Variable, not uniformly moderate. The recipients are not just "one business": the AI provider sees it as a processor on every classification call, and every current subscriber for that council sees it in their digest - potentially more than one firm once subscriber numbers grow. Severity for the individual also depends on who and what: a child's health information, as in the one real case found, is more sensitive than an adult's, and should be assessed on its own facts each time, not assumed moderate by default because it is "only one line."Lower, but not measured precisely: every real case found across two audit passes (9 word-root groups, 2 broader sample rounds) is now caught; the miss rate for phrasing or topics outside what we have sampled is unknown, not "low by assumption."Built, tested, backfilled across 2 passes (pipeline/src/sanitize.ts + test suite, 14 and 17 Nov 2026)
Applicant name/contact details leak through from a structured source field despite strippingLow (allowlist-based, not pattern-based, so no false negatives of this specific kind observed)Low-moderate (a name/contact detail, not health data; same multi-recipient point as above applies - AI provider plus all of that council's current subscribers)Same - this control doesn't depend on the free-text screenDone, ongoing check
Resident has no proactive way to know they're in a digest before it's sentHigh - this is structural, not probabilistic: every application we hold is in this position by design, not by chanceLow-moderate per instance - a missed opportunity to object before the fact, not a disclosure of extra content; mitigated by the reachable suppression route after the factUnchanged by the screen (a different control, not this one)Partial - reactive, not proactive; Art. 14(5)(b) reasoning in section 4
A subscriber misuses or redistributes a lead beyond lawful useUnknown - no subscribers yet to observe; not assumed low or high without evidenceCould be significant per instance (unlawful contact, resale), but is the subscriber's own compliance failure, not a failure of our own controls - see the note below this table on where our responsibility still sitsUnchanged by the screen (a different control, not this one)Terms require lawful use only, no resale/republication
Scale growth (more subscribers per council, or outreach at scale) changes the balancing outcomeNot applicable yet - zero live subscribers as of this assessment (see section 4)Would increase the severity figures above directly (more recipients per disclosure)Not applicable yetReview trigger below

A subscriber's own compliance duty does not transfer our responsibility away: we remain responsible for our own decision to disclose a lead to subscribers in the first place, and for the safeguards (the screen, the terms, the suppression route) around that decision, whatever a subscriber separately does or fails to do afterwards.

Residual risk conclusion

Provisional, not a settled finding - the evidence has genuine gaps, stated plainly rather than rounded up to a clean answer. What supports a lower reading: a low base rate of disability-adjacent wording across our samples, a smaller fraction naming a specific person, and the screen now catching every real case found across two audit passes. What keeps this provisional: the overall miss rate across the full range of possible phrasing is not measured, only sampled; severity varies by who is affected (a child's health information is not automatically "moderate" because it's one line) and by how many recipients exist, which grows as subscriber numbers grow; and the backup finding above means the wider unscreened population was exposed to a real (if currently unevidenced-as-accessed) channel for longer than our earlier drafts accounted for. We are not concluding residual risk is low; we are saying the evidence gathered so far does not show it is high enough to require prior ICO consultation under Art. 36, while flagging that this reading rests on incomplete evidence, not a completed audit. This, and the Art. 6/Art. 9 question above, are both for the solicitor to assess - not settled by this document, and not settled by the fact that the business has chosen to keep operating while they are assessed (see "Ownership, decision and review date" below).

Procedure if a missed disclosure is reported

If a resident, subscriber or anyone else reports that a specific digest or record disclosed a name or sensitive detail that should not have gone out: (1) the property/application is suppressed from any further distribution immediately, using the existing suppression route; (2) the specific text is reviewed to establish why the screen did not catch it; (3) where a real, generalisable pattern is found, the screen is extended to cover it, as happened with the 4 cases found in the 14 Nov audit; (4) the incident and its resolution are logged in DECISIONS.md; (5) the data subject is told what was done, if they are contactable.

Ownership, decision and review date

Operational decision, not a legal clearance: to continue processing, with the mitigations above, while the Art. 6/Art. 9 question, the incident assessment, the Art. 14(5)(b) evidence and the source-permissions question are all with the solicitor. The business's willingness to keep operating during that review is a commercial risk decision by the owner, not a finding that the processing is lawful - that finding, if it comes, will come from the solicitor's advice, not from this paragraph. Owner: Convert and Profit Ltd (the controller), decision recorded 14 Nov 2026, restated 17 Nov 2026 with the above made explicit. Review date: 14 Dec 2026 (one month, given the open legal questions) or immediately on any of: a real missed-disclosure report, the incident assessment reaching a conclusion, or subscriber count passing 50 live subscriptions - whichever comes first.

6. Consultation and sign-off

This draft has not yet been reviewed by our external legal adviser. Four specific things are being put to them, not resolved here: (a) whether the Art. 14(5)(b) justification in section 4 is adequate, with its supporting evidence; (b) a focused assessment of whether the transient, in-memory screening process in section 5 has a sufficient legal basis for the moment it is handling unscreened text, including the implementation options listed in section 5 and whether our source could reasonably be asked to filter sensitive content before supply; (c) the incident assessment in section 5, including whether it is a reportable breach and the governing basis for continuing to hold the retained evidence; and (d) the source-permissions/copyright question in our source-licensing note, which is a real open legal question in its own right, not a minor item.