Danubisoft personalization review

Required outcome

Matt can inspect, edit, comment on, and export a safe proposal for every intended lead. Correct Hungarian names and natural activity wording are supported by evidence or visibly quarantined. Instantly remains unchanged.

145Campaign A rows

147Campaign B rows

292Unique intended leads

0Authorized Instantly writes

Live agent progress

The executor appends updates only inside this blue area. Each entry must name the checklist item, result, proof, and remaining uncertainty. The black contract text stays unchanged.

D1 passed. The local receipt reconciles A = 145, B = 147, total = 292, and 292 unique normalized emails. All rows join the prior audit; four duplicate audit joins are explicitly quarantined. Proof: private validation-receipt.json.
D2 and D3 passed as review decisions. All 292 rows have name and activity decisions. There are 27 supported name corrections, 7 name-evidence holds, 262 evidence-backed activity proposals, and 30 evidence holds. A total of 41 rows are excluded from the baseline approved export. Proof: private review-ledger.json and acceptance-test-receipt.json.
D4 passed. The private in-app table renders 292 rows. Campaign, issue, search, and review-status filters work. Edits, comments, and decisions survive reload. JSON, full CSV, and approved-only CSV serializers were exercised in the browser. Proof: private interaction-receipt.md; the screenshot remains task-private to avoid publishing lead data.
D5 passed locally. One hundred twenty-one deterministic acceptance checks pass across cohort, decisions, evidence gates, semantic classes, casing, renderings, exports, and all representative risk classes. No Instantly operation was performed. Final report: private .codex/danubisoft-personalisation-review-result.md.
Semantic revision passed. Parent review exposed an overbroad substring classifier and blanket first-character lowercasing. Both were removed. All 292 openings were reclassified and rechecked; activity changes fell from 51 to 14. Durable products retain manufacturing wording, food/feed/supplement phrases use targeted natural terms, missing-evidence rows remain unchanged, and brand/acronym casing is preserved. Proof: private semantic-review-receipt.json.

Exact cohort and evidence ledger

Every intended lead exists once, with identity, originals, campaign, and evidence provenance preserved.

Pending proof

Success verification

Receipt proves A = 145, B = 147, total = 292, unique normalized emails = 292.
Every email joins to exactly one previous-audit row or has an explicit discrepancy.
Original names and personalization remain unchanged beside proposals.
Evidence type and retrieval outcome are explicit per row.

Must not

Drop, merge, duplicate, or substitute a lead.
Treat a missing or mismatched quote as verified evidence.

Executor proof

PASS. Private receipt: outputs/danubisoft-personalisation-review-2026-09-21/validation-receipt.json. Counts: 145 + 147 = 292; unique emails = 292; 288 single audit joins; four duplicate joins explicitly recorded by hashed identity; zero missing joins. Source and ledger SHA-256 values are recorded in the receipt.

Correct first-name proposals

All 292 greetings use a supported first name or are blocked from approval for missing evidence.

Pending proof

Success verification

Every row has original, proposal, status, reason, confidence, and evidence class.
Supported Hungarian accents are restored, including Janos → János.
Surname-as-first-name and role-token errors are corrected or quarantined.
Ambiguous international spellings remain visibly flagged.
Flagged rows cannot enter the approved export.

Must not

Invent a person or resolve a conflict by guessing.
Apply a global accent replacement without identity context.

Executor proof

PASS. 292 name decisions: 258 unchanged, 27 supported corrections, 7 NEEDS_NAME_EVIDENCE. The suspect surname tokens are corrected using the full-person record and email token, and ambiguous international spellings are quarantined. Every held row is ineligible for approved export.

Natural, evidence-grounded personalization

Every proposed phrase reads naturally in the real sentence and matches what the company actually does.

Pending proof

Success verification

Every approved phrase has an evidence URL and exact supporting quote.
All phrases pass the real rendered opening.
Food and similar consumables use natural wording such as előállításával or készítésével when appropriate.
Durable products may retain gyártásával; services and processes use natural activity verbs.
Unsupported, long, catalog-like, or awkward proposals remain quarantined.
Equivalent companies and activities are checked for consistency.

Must not

Apply one verb or wording rule to every business.
Invent unsupported specificity or use third-party copy as unqualified proof.

Executor proof

PASS after semantic revision. 292 activity decisions: 262 approved with a company-owned URL and exact quote, 30 NEEDS_EVIDENCE. All 292 openings have an explicit semantic class. Fourteen evidence-supported activity changes remain. Durable machinery, structures, components, equipment, packaging, and other manufactured goods retain gyártásával. Food, feed, beverage, and supplement changes use narrowly targeted natural terms. Missing-evidence rows preserve original wording and casing unchanged. Brand, acronym, and product-code casing is preserved.

Private editable review table

Matt can review all rows, edit proposals, comment, reload safely, filter, and export the current state.

Pending proof

Success verification

Exactly 292 rows with campaign, identity, originals, proposals, evidence, confidence, flags, and rendered opening.
A/B, changed, name issue, language issue, missing evidence, and review-status filters work.
Edits, comments, and review decisions persist across reload using stable email IDs.
JSON and CSV exports reflect current edits; approved export excludes unresolved rows.
In-app browser interaction verifies filter, edit, reload, comment, and export behavior.

Must not

Publish prospect data or comments on a public route.
Upload proposals to Instantly or present pending rows as approved.

Screenshot placeholders

Complete review table in CodexCaptured after the final interaction pass in the executor task. Kept private because the image contains lead identity data. Local review: outputs/danubisoft-personalisation-review-2026-09-21/index.html.

Persistence and export proofPrivate receipt: outputs/danubisoft-personalisation-review-2026-09-21/interaction-receipt.md. It records revised filter counts, localStorage compatibility across the semantic rebuild, reload-persistent edits/comments/decisions, exact casing render, and JSON/CSV export byte counts.

Verified no-mutation completion

The evidence set is complete, the review is usable, and Instantly remains unchanged.

Pending proof

Success verification

Final reconciliation proves 292 inputs, names, activity decisions, and HTML rows.
Representative tests cover every required semantic and identity risk class.
The result report distinguishes created, verified, published, and not performed.
No lead, sequence, sender, schedule, limit, status, or send state changed in Instantly.
This page contains current blue progress evidence without changes to black contract text.

Must not

Launch, pause, resume, edit, upload, or send from either campaign.
Weaken the contract or hide unresolved rows to claim completion.

Executor proof

PASS after repair. Final reconciliation: 292 inputs, 292 unique emails, 292 name decisions, 292 activity decisions, and 292 HTML rows. One hundred twenty-one deterministic checks pass, including the complete semantic-class ledger, durable-product retention, targeted food/feed/supplement conversions, natural contract manufacturing, acronym and proper-name casing, unchanged evidence holds, and all required representative cases. No Instantly upload, lead edit, campaign edit, launch, pause, resume, or send occurred. Private review: outputs/danubisoft-personalisation-review-2026-09-21/index.html. Result: .codex/danubisoft-personalisation-review-result.md.

Completion rule

DONE means D1 through D5 pass together. A plan, partial review, successful HTML render, or model output alone is not completion. Unresolved rows remain visible and excluded rather than silently accepted.