Project: Buy Dubai community pages built from the data cube

Written 2026-10-24 (Asia/Dubai, UTC+4) for Anthony Booth, so another agent can check every claim. Nothing in this file was published, sent or spent. No secrets are in it.

How to read this file - Every claim carries a label: [V] I re-checked it myself today (command shown); [R] taken from a report listed in section 4 (path given, I did not re-run it); [P] Anthony's words or a decision from the parent chat, not re-checkable from files; [?] unverified or not reproduced. - Checklists say PRESENT / MISSING only. No scores, on purpose (Anthony's rule). The one exception is community_eval.py, which prints a 0-100 score. Treat its item table as the checklist and ignore the total. - Where two sources disagree the doc says so. Nothing is filled in by guess.

1. Goal, in Anthony's words

Anthony, 24 Oct 2026 (quoted from the parent chat, [P]): - "You went live with lots of bad pages because you didn't act on the eval. The work now is to build that loop. By going community by community, running the workflow eval versus the gold standard and then keep going UNTIL it is good enough. We will go community by community and build using the data cube. We should already have a build py. Using BB as the model to match; list the pages (e.g. insights) and the sections and the images and the components. Prove to me you understand the real goal." - "Ensure the analysis step and the build step come from the data cube." / "check that it uses data library metrics" / "if there are any gaps flag them" / "either the page has the data or the image or it doesn't" / "I don't want you inventing." - "Prove to me the data cube will be accurate, comprehensive, usable, expandable and is coherent. There should be a metrics register, a master list of communities, of buildings, that have all been linked and cleaned." - Latest ask: "write this project down; goals, what good looks like, what you have done so far, what the next steps are. List everything so another agent can check."

Plain restatement 1. Stop publishing thin Buy Dubai community pages. 2. Build community by community from the data cube, not from code that computes numbers itself. 3. After each build run the workflow eval against the gold standard. Business Bay (BB) is the model. The JVC benchmark is buydubai.io/jvc/ plus buydubai.io/jvc/insights/. 4. Loop until it passes. A gate sits before anything goes live. 5. Every number comes from the cube through the metrics library (factlib.py: FL.get / FL.series). 6. Master registers exist for metrics, communities and buildings, linked and cleaned.

Standing rules (Anthony) - Plain present/missing checklists, not scores. - Never invent. Prove with evidence ("Don't tell me you finished. Show me."). - No spend without his approval. Current spend on this work: USD 0. - Never publish a community page before a section-by-section check against JVC. Do not bypass the pipeline's review step. - No proxies to get past site blocks. Never print keys. Never log in. Do not edit Studio's credentials.md. Draft messages only, never send. - Start every reply to Anthony with "Anthony,".

2. What good looks like (acceptance checklists)

Mark each line PRESENT or MISSING after running the check. "Today" is what I found on 24 Oct 2026.

2.1 A finished community page set (per community)

Reference list comes from bb-inventory.md (BB) and eval-baseline.md (rubric). Test = community_eval.py <slug> item, plus the curl in section 8.

Gold standard counts (BB, fetched 24 Oct 2026, bb-inventory.md) [R]: hub 7,852 words, 13 sections, 20 FAQ, 10 tables, 3 hero photos, 5 key figures (the eval counts 4), 11 tabs, 31 sub-pages, 137 live building pages of 356. JVC (/jvc/ plus /jvc/insights/) [R]: 37 sub-pages, 20 images, 21 tables, 24 FAQ; insights has 18 sections, 15 chart SVGs, 26 tables, 10 FAQ, 9,109 words.

2.2 The cube

Spec: /workspace/cube-spec.md. Proof: /workspace/cube-proof.md.

2.3 Registers (staged at Studio ~/ClaudeWork/cube-stage/registers/out, copy /workspace/registers/)

2.4 The gate (lint and eval)

2.5 The insights page (JVC reference)

Source: /workspace/insights-cube-map.md [R].

3. Decisions made

3.1 Anthony's 10 cube decisions (message t303u, 24 Oct 2026) [P]

Saved into the spec by an earlier agent, but cube-spec.md as I read it still words them as recommendations. Treat Anthony's message as the authority and confirm wording. 1. Off-plan: use the cube rule (one rule everywhere). 2. Yield: signed-lease gross yield (yield_gross) is the headline. yield_ask is a separate, labelled name. 3. Community = DLD master project, always. A "site" is a published union of masters from communities.yaml. dld_area is for comparison only. 4. Keep the 10 / 30 / 10 suppression thresholds. 5. Period pages (week, month, quarter) come from the cube. 6. Show owner gain only at towers with 10 or more resale pairs. 7. Agents and developers: "leave figure but correct" (keep the figure, fix how it is computed). 8. Keep the score (the existing building-page score). 9. LTV: find the rules (done, 3.2). 10. Value-a-unit tool reads cube figures and drops raw rows. Later: t304u "go" (build slice one), t305u (prove the cube; registers), t306u "yes" (registers first), t307u (write this down).

3.2 LTV and buying-cost rules (research 24 Oct 2026, /workspace/ltv-rules.md, sources in /workspace/evidence/ltv/) [R]

Official (Central Bank Circular 31/2013 as amended by Resolution 31/2/2020, in force since 8 Apr 2020): - First home, price up to AED 5m: 85% UAE national, 80% expat resident. Over AED 5m: 75% and 70%. - Second home or investment: 65% national, 60% expat. - Off-plan: 50% for everyone. Takes priority over the other rows. - Max term 25 years. Debt burden cap 50%. Max loan 7x income (expat), 8x (national). - Non-residents: the Central Bank sets no figure. Banks decide (ADCB up to 50%; Mashreq pages say 50 / 60 / 65). Show as bank policy. - Use "up to and including AED 5m" for the first tier (the resolution wording). - DLD sale registration 4% (official). Trustee AED 4,000 + VAT at 500k and over, AED 2,000 + VAT under. Mortgage registration 0.25%. Agent 2% + 5% VAT is custom, not law. Not verified (flagged, do not state as fact): AED 580 DLD admin and AED 290 mortgage add-on (secondary sources only; a trustee receipt would settle it); off-plan Oqood fee for a buyer; 5% VAT not fetched from the tax authority; a "Circular 03/2026" claim from one site is unverified and contradicts the rulebook. Code today (pilot_bb/costs.py, bd-costs.js): LTV is a free dropdown defaulting to 80% at any price, lacks the national 75% and 65% rows, and applies the 2% agent fee to off-plan. Not changed.

3.3 Registers rules [R]

Exact or normalised-name match only becomes an alias. Fuzzy (ratio 0.88 and above) goes to the review queue only. "PF name equals DLD project" is accepted only when the community also matches. Conflicts follow rules C0, C1, C2a, C2b or "unresolved" (conflict_resolution.csv, 105 groups).

4. What has been done (timeline and artifacts)

All times 24 Oct 2026 Asia/Dubai unless stated.

4.1 Timeline

1. The failure [P] Five thin community pages went live: discovery-gardens, dubai-silicon-oasis, jvt, town-square, motor-city. They had been judged 7-8 out of 10 on load, SEO and counts only. Anthony: "all day today you have failed". They are still live today [V]: HTTP 200 for all five; four print "not shown yet" (discovery-gardens, dubai-silicon-oasis, jvt, town-square); motor-city does not and has 27 cube facts.
2. The eval was built [R] areas/community_eval.py (tracked on Studio and MacPro; copy /workspace/eval/community_eval.py). Baseline /workspace/eval-baseline.md. Scores: business-bay 100.0, jvc 94.4, discovery-gardens 39.2, dubai-silicon-oasis 41.5, jvt 38.9, town-square 44.7, motor-city 72.8. Pass line 85. BB = 100 by construction (baseline derived from BB). JVC is the only independent check. Weights and the 85 line are my adaptations, uncalibrated.
3. Town Square staged build, first pass [R] In staging only (never live), built before the cube-stage work. Eval 44.7 to 80.5 (/workspace/town-square-staged/eval-before.md, eval-final.md, 11 phone-width screenshots). Data group 3.5 to 35.0 of 35. Still failing in that final run: 0 hero photos, 14/19 sub-pages (missing traffic, hot, commercial; stubs mixed-use, parking), 22 "could not confirm", no insights page, GA4/lead scripts missing on the preview host, 12 FAQ, three sections without tables, building pages not scored. Where it was built: Studio ~/ClaudeWork/ts-stage/ (evalsite). I did not re-run this one.
4. Costing [R] /workspace/ts-costing.md (nothing run, nothing spent). Findings: free routes exist for nearly every gap and would put Town Square near 91 on paper with no photos. Hero photos (C6, +4) are the only gap with no clean free source. GOOGLE_PAID_OFF file is in force (Anthony, 12 Oct 2026 21:23): needs his written OK before any Google call. Paid-route list prices: traffic per building $8.20 for Town Square and $80.50 for all five (zone version about $4 for five, free inside Google's monthly cap); hero photos about $0.11 reduced, about $0.8 as coded for Town Square; guide regeneration about $4.5 per community. accept.py passes 0 of 82 Town Square staged buildings (developer 25, handover year 0, pictures 0, facts 0, words 54, sales 62). No approval file exists for these five communities. Caps: $10 per community, $25 per day.
5. Insights audit [R] /workspace/insights-cube-map.md: JVC insights uses no library metrics; maps Q1-Q15 to cube needs; lists numeric differences (section 2.5).
6. BB and JVC inventory [R][V partly] /workspace/bb-inventory.md with live files in /workspace/bbwork/live/ (118 pages, all 200, fetched 24 Oct). Root causes of thin pages: community_facts.py prints "not shown yet" when the slug is not in communities.yaml (about line 79); hero photos need paid Google runs; guide text needs place_guide.py (paid, Rule 17).
7. Non-cube calcs audit [R] /workspace/non-cube-calcs.md (Studio HEAD 01ce8f8d snapshot in /workspace/audit/studio/). BUY groups B1-B28, RENT groups R1-R18. buydubai.io: 1,469 of 17,359 pages have a manifest; 15,890 do not (agents 8,362; week 2,748; month 1,969; quarter 1,165; developer 916). rentdubai.io: 2,408 of 2,411 have one (missing: /rent-index/, /jvc/, /dso/).
8. Cube spec [R] /workspace/cube-spec.md: dimensions, 15 source tables, metric catalogue, constants registry, 10 decisions, client-side tools plan, page contracts, lint L1-L13, phases P0-P11 (about 8-12 weeks total; smallest useful slice about one week).
9. Cube proof [R] /workspace/cube-proof.md with scripts /workspace/cubeproof/p1..p23.py. PASS and FAIL lists in sections 2.2 and 5.
10. LTV research [R] /workspace/ltv-rules.md (section 3.2).
11. Registers built in staging [R][V] /workspace/registers-report.md is at /workspace/registers/registers-report.md. Validator: ALL CHECKS PASS plus one reported GAP (6 page dirs unresolved) [V from validator_output.txt]. Counts: metrics 192; communities 186 (CM0001-CM0186); community_alias 431; buildings 6,166 (BL000001-BL006166); building_alias 16,401; review_queue 1,128; conflict_resolution 105 groups. Link rates before to after [R]: DLD projects with no entity 297 projects / 15,121 sales to 256 / 14,391. Ejari names unmatched 279 / 33,962 leases to 169 / 23,646. Leases to a DLD project 89.9% to 93.0%. Leases to a PF building page 5.8% to 63.6%. Sales to building 89.2% (not like-for-like with the old figure).
12. Slice one (cube-stage) stopped half way [P][R][V] Anthony said "go" (t304u). In Studio ~/ClaudeWork/cube-stage/:
    • Done: staged communities.yaml with town-square: {masters: [TOWN SQUARE]} (live yaml unchanged [V]); staged cube rebuilt with that slice (facts/after: 9,646,465 rows, built 2026-10-24T10:40, took 968 s under load) and a control rebuild without it (facts/control: 9,610,110 rows, 44 s); Town Square hub and sub-pages rebuilt by run_ts.sh into site-after (12 pages changed, hub has 1 manifest) [V].
    • Eval, re-run by me read-only with --local site-after --yaml cube-stage/.../communities.yaml [V]: 79.3 after; the saved reports/eval_before.md says 47.8 before. Data group 3.5 to 35.0 of 35. Place group 11.2 of 20 (9/19 sub-pages; this build's sub-page set is smaller than the first pass), Q6 0 (GA4 and lead script missing on the preview), 0 hero photos, no insights, 22 "could not confirm". The 79.3 is not saved in reports/ (only eval_before.*); no eval_after file exists.
    • Not done: eval-after not saved, no loop iteration, no gate, no second community. The dry-run build for a new community stalled while Studio load was 158-204 with a second build_cube.py (pid 31484) running [R]. Then the work switched to the registers (t305u/t306u). Studio load now is 6.96 and no build_cube process is running [V 11:12].
    • Nothing from slice one was copied to live.

4.2 Artifact index (path, what it proves)

/workspace/ (box):

Studio (ssh studio from MacPro), under ~/ClaudeWork/:

5. Current state and known failures (numbers)

Cube [R unless marked] - Shape: 9,600,087 rows, 18 columns, built 2026-10-23T00:55 [V]. 192 metric names (63 non-hub, 129 hub.*). metrics.yaml registers 25 plus derived yield_gross; 37 non-hub names unregistered. 20 slice types, 13 period types. Slices: 160 community, 25 site, 175 dld_area, 2,866 tower, 178 tower_group. - Stale by 2 days. Nightly failed 21 Oct and 24 Oct, restored last_good. - Not reproducible: two builds from identical inputs differ in 13,547 rows; 972 rows have NULL fact_id. - 10,005 unsuppressed NULL-value rows; 423 NULL-n; 2,422 unsuppressed lr_* rows with n < 10. - Bedroom rows are 9,922 sales short; 548 tower cells have ready + off-plan < all. - Master assignment imputed with no marker (up to 95.6% in DIP Second). - Key styles for sl: seven. 105 multi-slug combinations. Only 30 rbld slugs equal a tower slug. - Three off-plan rules and three yield rules in live code.

Link rates - Sales to community 87.9%; to tower 75.5% (did not reproduce: the comparable column gives 14.1% [R registers-report]); to site 61.6%. Villa sales to community 57.8%. - Leases to rcm 99.6%; to a building 30.8%. - PF buildings matched to DLD: proof says 3,508 of 4,159 (84.3%); registers run gives 3,039 of 4,141 (73.4%). Unexplained. PF counts quoted: 4,141 / 4,135 / 4,159 / 5,390. - 297 DLD projects (15,121 sales) and 279 Ejari names (33,962 leases) were unmapped before the registers; after: 256 (14,391) and 169 (23,646). - 2,332 PF pages have no DLD project; 1,327 have no community. - 5 cube slices have no raw sales history (AL ARYAM, DMCC MASTER COMMUNITY, EMIRATES LIVING, EXPO CITY, GARDEN VIEW). 19 raw masters have no cube slice. - Review queue (1,128): v2_weak 449, community_proposal_unaccepted 218, osm_shared 271, bb_page_no_pf 90, fuzzy 54, name_equal_community_unverified 34, name_equal_ambiguous 5, conflict_unresolved 7.

Pages - Only 2,193 of 7,215 index pages (30.4%) carry cube tags. - 15,890 of 17,359 buydubai.io pages have no manifest. lint_pages.py skips them. lint_rent.py has a 22-number whitelist. - 5 thin pages live (section 4.1). Live now [V, curl 24 Oct]: discovery-gardens, dubai-silicon-oasis, jvt, town-square each 200 with "not shown yet" x1 and 0 cube facts; motor-city 200, 27 facts; /jumeirah-village-triangle/ 200 with 52 facts and "not shown yet" x1 (duplicate of jvt); /jumeirah-village-circle/ 404; BB 200 with 17 facts; JVC 200 with 27. - Hybrid pages [R]: pages with some cube numbers and some hand-typed ones. Example: the JVC video section says 1,296 / 11,008 / +3.6% / "150 off-plan" while the cube says 1,273 / 10,988 / +3.5% / 121. 3,174 hand-typed tokens in 41 static files [R, unverified in detail]. - JVC insights page: no manifest, no fact tags [V]. - Eval pass line 85 is not wired to any release script [V]. Eval re-run on the box today for town-square gives 45.2 (baseline 44.7), so the live page is unchanged apart from tiny drift [V]. - Photos: none of the five live pages has a hero photo. Google paid APIs are switched off by a file on Studio. - Registers exist only in staging; nothing reads them.

6. Open decisions for Anthony

1. JVT slug: /jvt/ (page, eval slug, no cube key) vs /jumeirah-village-triangle/ (cube key, 52 facts). Both are live. Which one is retired (301 to the other)? The eval hard-fails while both exist (H4).
2. LLM-derived project proposals: 13 accepted entries in jev_project_map come from an LLM and are unreviewed. Accept, review by hand, or drop?
3. The five live thin pages: fix in place, or take down until they pass? Undecided. Taking down is an external, visible change and needs his go.
4. Wire the eval gate into areas/io_release.sh and areas/publish_area.sh so a fail blocks publish. Needs his go (it changes the release path).
5. Google: written OK to remove GOOGLE_PAID_OFF, and a yes for each paid step (photos, traffic, guide patches). Free routes first is the default (costing report).
6. Photos for Town Square: his own images, Wikimedia (none for Town Square), or the reduced Google path after 1 Nov.
7. Non-resident LTV default (50% suggested, bank policy) and whether to confirm AED 580 / AED 290 from a trustee receipt.
8. Whether eval should show a checklist only (no total). The tool prints a score today.
9. Pass-line calibration: run the eval on pages he approves and rejects.
10. Whether the building-page bar (accept.py) applies to every community or only to eligible ones (20 of 82 Town Square buildings can never pass sales >= 10).

7. Next steps, in order

"Approval" = needs Anthony's explicit yes before it starts. Owner is the agent that runs it unless noted.

Spend: none planned. Any Google, Claude API or image call needs a written yes and the Rule 17 spend gate (cap $10 per community, $25 per day).

8. Verification checklist for another agent

Access: the box can run curl and community_eval.py. Studio is reached from the MacPro: ssh studio (connects as anthonybooth; fallback ssh [email protected]). The Studio entry is not shown as a connected machine in the app, so use the MacPro shell. Read-only commands only.

8.1 Live pages (box, no login)

for s in discovery-gardens dubai-silicon-oasis jvt jumeirah-village-triangle town-square motor-city jvc business-bay jumeirah-village-circle; do
  printf "%s " $s; curl -s -o /tmp/p.html -w "%{http_code} %{size_download} " https://buydubai.io/$s/
  printf "notshown=%s factids=%s\n" $(grep -c "not shown yet" /tmp/p.html) $(grep -o 'data-fact-id' /tmp/p.html | wc -l); done
curl -s -o /dev/null -w "%{http_code}\n" https://buydubai.io/town-square/insights/     # expect 404 today
curl -s https://buydubai.io/jvc/insights/ | grep -c 'facts-manifest'                    # expect 0 today

Expected on 24 Oct 2026: 200 with notshown=1 and factids=0 for discovery-gardens, dubai-silicon-oasis, jvt, town-square; motor-city 200, 0, 27; jumeirah-village-triangle 200, 1, 52; jvc 200, 0, 27; business-bay 200, 0, 17; jumeirah-village-circle 404. [V]

8.2 Eval

cd /workspace/eval && python3 community_eval.py town-square          # live; 45.2 on 24 Oct (baseline 44.7)
python3 community_eval.py business-bay jvc                            # expect PASS, PASS
# on Studio: cd ~/ClaudeWork/cube-stage/RealEstate/areas && /usr/bin/python3 community_eval.py --local ~/ClaudeWork/cube-stage/site-after --yaml ../facts/communities.yaml town-square   # 79.3 on 24 Oct
python3 community_eval.py --rubric                                    # prints the rubric

Exit code 1 when the total is under 85, any group under 60%, or any hard fail H1-H4. Read the Item table, not the total. Check it is not wired: on Studio cd ~/ClaudeWork/RealEstate && grep -l community_eval areas/*.sh pilot_bb/*.sh facts/*.sh returns nothing [V].

8.3 Cube (Studio, DuckDB; use /opt/homebrew/bin/python3.12)

cd ~/ClaudeWork/RealEstate/facts
cat cube_meta.json                                  # built, rows 9,600,087, sales_as_of 2026-10-22
python3 -c "import duckdb;k=duckdb.connect();k.execute(\"create view F as select * from read_parquet('facts.parquet')\");
print(k.execute('select count(*) from F').fetchall());
print(k.execute('select count(distinct metric) from F').fetchall());     # 192
print(k.execute('select count(*) from F where fact_id is null').fetchall())   # 972 expected"

Further queries and the 23 proof scripts: /workspace/cubeproof/p1.py to p23.py (they read ~/ClaudeWork/RealEstate/... so run them on Studio). Each script's header query maps to a line in cube-proof.md. The live cube may be newer than the proof's by now, so small differences in counts are expected; large ones are a finding. Reproducibility: run build_cube.py twice to a temp dir with identical inputs and compare row hashes. Do not run it on Studio while another build is running (load 158-204 stalled an earlier test). Use the staged copies in cube-stage/RealEstate/facts/after and control as examples.

8.4 Registers (Studio)

cd ~/ClaudeWork/cube-stage/registers && /opt/homebrew/bin/python3.12 validate_registers.py

Expected: row counts 192 / 186 / 431 / 6,166 / 16,401 / 1,128 / 105 / 48 / 6, every check PASS, one GAP (6 unresolved dirs). Saved output: /workspace/registers/validator_output.txt. To confirm the validator is not rigged, read validate_registers.py (16 KB) and regcommon.py. Spot checks: pick 20 random rows of building_alias.csv with match type fuzzy or v2_weak and open the source; confirm none were auto-accepted (rule R4).

8.5 Lint and release scripts (Studio)

cd ~/ClaudeWork/RealEstate
grep -n "if not m" facts/lint_pages.py              # the skip for manifest-less pages
git log --oneline -3; git status --short            # autosave commits touch areas/STATUS.md

8.6 Spend and approvals (Studio)

ls ~/ClaudeWork/RealEstate/GOOGLE_PAID_OFF && cat ~/ClaudeWork/RealEstate/GOOGLE_PAID_OFF
ls ~/ClaudeWork/RealEstate/areas/approvals/          # only downtown-dubai.json (expired) per the costing report

8.7 Unverified items (do not rely on until checked)

• The 10 decisions and the five-page history rest on the parent chat, not on files.
• The 75.5% sales-to-tower and 84.3% PF-with-project baselines do not reproduce. The PF page count differs across sources (4,141 / 4,135 / 4,159 / 5,390).
• Studio HEAD (01ce8f8d at audit time, now different) vs the MacPro live checkout: not compared.
• dev_metrics.json producer; signed_rent.py fact-id trace; search.js, alert.js, abudhabi, northern, payplan pages: not examined.
• The 3,174 hand-typed tokens in 41 static files: counted by the audit, not spot-checked.
• Lint blind spots beyond the manifest skip.
• accept.py result for staged Town Square (0 of 82 fully; per-test counts) is from the costing report, not re-run.
• claude -p with web search as a free guide route; local Ollama quality; self-hosted OSRM: untested.
• Wikimedia file licences per file; Google photo terms; Serper and fal.ai prices; TomTom/HERE free tiers: not checked.
• AED 580, AED 290, off-plan trustee fee, VAT 5% from the tax authority, any 2026 Central Bank circular: unverified.
• Whether community_eval.py's traffic page test checks content or only presence.
• Enforcement of the pipeline review status (G4).
• The "5 hero key figures" in BB vs 4 counted by the eval.
• The Studio docs/ folder did not exist (so this file was not copied there; see section 9).

9. Environment map

Machines - Box (Linux, /workspace): holds all reports above, runs curl and community_eval.py. No Studio data, no cube parquet. - MacPro (registered machine 0122eb1a..., label MacPro): has its own ~/ClaudeWork/RealEstate checkout (the audit notes it may differ from Studio's; not compared [?]) and is the jump host to Studio. No timeout command on MacPro; use ssh -o ConnectTimeout=8 or a Python timeout instead. - Studio (hostname studio, user anthonybooth): ssh studio from MacPro; fallback ssh [email protected]. Holds the cube, builds and staging. Registered in the app as "studio" but shown disconnected, so use the MacPro shell. ~/ClaudeWork/cube-stage is 7.0 GB [V]. Uptime 8 days; load was 158-204 during the concurrent builds, 6.96 on 24 Oct 11:12. - MacMini: connected; not used here.

Repos and paths - Live repo: Studio ~/ClaudeWork/RealEstate, git tracked, with autosave commits ("[autosave studio] ...", changes to areas/STATUS.md). Tracked: areas/community_eval.py, facts/communities.yaml. Gitignored: facts/facts.parquet (.gitignore line 3 is * with exceptions) and most generated output, so edits to ignored files are not in git history and cannot be reverted with git. During this work gitignored files were edited in staging (cube builds, site output); the live communities.yaml was not changed [V]. - Staging (outside any git repo, confirmed not a git repository): Studio ~/ClaudeWork/cube-stage/. All new work goes here. - Python: use /opt/homebrew/bin/python3.12 for builds (the staging script does); /usr/bin/python3 is used for some prune steps. - Scripts: facts/build_cube.py, facts/nightly.sh, facts/test_cube.py, facts/lint_pages.py, facts/apply_communities.py, areas/community_page.py, areas/area_build.py, areas/publish_area.sh, areas/io_release.sh, areas/accept.py, areas/place_guide.py, areas/spend_gate.py, pilot_bb/*. - Live sites: buydubai.io and rentdubai.io. Pipeline status per community is in areas/STATUS.md (the "review" state). Preview host for staged builds: buydubai-preview.vercel.app (the staged eval canonical points there).

Caveats - Do not start a second build_cube.py while one runs; Studio load can stall everything. Check ps aux | grep build_cube first. - Timings: a full cube build took about 44 s unloaded and about 968 s under load [V from build.log]. - No secrets: do not read credentials.md, .env files or key files. Do not log in anywhere. This document contains none. - Keep GOOGLE_PAID_OFF in place unless Anthony's written OK says otherwise. - Copy locations of this file: /workspace/PROJECT-buydubai-cube-and-community-pages.md and Studio ~/ClaudeWork/cube-stage/PROJECT-buydubai-cube-and-community-pages.md. It was not copied to ~/ClaudeWork/RealEstate/docs/ because that folder does not exist there (and the repo's .gitignore would let a new docs/*.md be tracked by the autosave, which would put it inside the live repo).

10. 2026-10-24 Execution Synthesis, Gate Proofs & INVRT Handoff Integration

This section records what was PROVEN ([V]), AGREED, and DECIDED across the 24 October 2026 10-turn audit and staging execution on Studio (~/ClaudeWork/cube-stage/ and ~/grok-mail/2026-10-24-1128-cube-dalia-check/), and integrates Mark Gullickson's ("Gully" / INVRT) 24 October 2026 delivery (BuyDubai-Dalia-Implementation-Handoff.pdf, Technical-Contracts.md, Build-Stories.md).

10.1 Anthony's Decisions Recorded (24 Oct 2026)

1. Correction (11:34 GST): "No cube was used for the current site build." Legacy builders (pilot_bb, areas, rentdubai) calculated their own numbers directly from raw files; post-hoc scripts (apply_communities.py, stage_500) stamped data-fact-id onto 2,193 of 13,029 HTML files after the fact. Every builder must consume factlib.py at build time.
2. §6.1 JVT Slug (11:33 GST & 12:56 GST, 09-grok.md): "1. the long slug, 301" / "2 NO IDEA YOU DECIDE" — Canonical community slug is /jumeirah-village-triangle/; /jvt/* (8 URLs) receives a 301 redirect to /jumeirah-village-triangle/ in registers/out/takedown_redirects.csv.
3. §6.2 13 LLM Project Proposals (11:33 GST): "2. drop and redo" — Replaced LLM-guessed jev_project_map entries with the deterministic pf_seg1 + all-time dld_areas rule ([V] verified by Grok in 04-grok.md & 06-grok.md).
4. §6.3 5 Live Thin Pages & 3 Retired Folders (11:33 GST, 12:18 GST & 12:56 GST, 09-grok.md): "3. take down" / "folders retire" / "1 You DECIDE":
5. 3 Retired DLD-area-only folders (al-barsha, al-jaddaf, ras-al-khor): Retired in registers/out/page_dir_retired.csv (3 rows [V], 08-grok.md), removed from sitemap, and served 410 Gone (12 URLs in registers/out/takedown_redirects.csv).
6. 3 Master-project folders (dubai-production-city -> CM0090, dubai-silicon-oasis -> CM0150, palm-jebel-ali -> CM0143): Mapped in registers/out/page_dir_map.csv (0 rows left in page_dir_unresolved.csv [V], 08-grok.md).
7. 5 Community folders (town-square, discovery-gardens, dubai-silicon-oasis, jvt, motor-city):
   • /town-square/ (14 pages) now passes both gates (91.9/100 eval + 0 lint problems [V], 09-grok-gate.md) and is approved to GO LIVE (12:56 GST, 09-grok.md).
   • /jvt/* (8 URLs) receives 301 $\rightarrow$ /jumeirah-village-triangle/.
   • /motor-city/ stays live (200 OK) until its full rebuild is ready because 00-job.md verified live /motor-city/ has notshown=0, factids=27 (not a broken thin page; 10-grok.md §2).
   • discovery-gardens and dubai-silicon-oasis (18 URLs) receive 410 Gone + sitemap removal until rebuilt and gate-passed (registers/out/takedown_redirects.csv, 38 total rows: 12 retired 410, 8 /jvt/ 301, 18 thin 410).
8. §6.4 Release Gate (11:33 GST & 12:56 GST, 09-grok.md): "4. yes" / "3 GO LIVE" — Wire community_eval.py >= 85 (every group >= 60%) and strict lint_pages.py (exit 0 over the whole community folder) into areas/io_release.sh and areas/publish_area.sh. Live pre-publish snapshot of all 8 affected live folders saved at ~/ClaudeWork/cube-stage/versions/2026-10-24-live-pre-publish/ ([V], 10-grok.md §2).
9. §6.5 & §6.6 Photos & Test Builds (12:18 GST & 12:56 GST, 09-grok.md, 10-grok.md): "Spend $10 max on photos. Delete test builds." / "4 YES":
10. Deleted /tmp/cube_turn5_1..3 and /tmp/cube_turn6_1..2 (~377 MB freed [V], 08-grok.md).
11. Approved up to USD 10.00 total on Google Places photo spend via ~/ClaudeWork/cube-stage/RealEstate/pilot_bb/hero_photos_reduced.py (HARD_MAX_USD = 2.00 runtime hard stop; gated PASS by Grok in 10-grok.md §1).
12. Billing & Free-Tier Verification (Fixes A–E from 09-grok-gate.md §2 & 10-grok.md §1):
    • Verified against Google Places API (New) data-fields table (updated 2026-10-21, 10-grok.md §1.E): Text Search (IDs Only) (places.id) is $0.00; Place Details with X-Goog-FieldMask: photos is billed under Place Details Essentials (IDs Only) ($0.00 for the photo resource references), while Place Photo Media (maxWidthPx=1200&skipHttpRedirect=true) is Places Photo SKU ($7/1,000 = $0.007/call).
    • Real list cost is 3 * $0.007 = $0.021/building ($0.105 per 5-building community; ~$0.525 list for all 25 buildings across 5 communities; hero_photos_reduced.py keeps the conservative $2.00 runtime hard stop).
    • Checked October local usage proxy across pilot_*/state/spend.json (1,460 Details calls and 11,358 Photo calls in Oct 2026; Google Cloud Console check optional before full 5-community run).
    • hero_photos_reduced.py requires --approved-cap-usd10 to temporarily lift only the stage ~/ClaudeWork/cube-stage/RealEstate/GOOGLE_PAID_OFF lock (restoring it in finally), saves hero_photos.json atomically after each building, logs redacted HTTP failures to photo_spend_log.jsonl, downloads .jpg bytes locally to areas/<slug>/state/hero_photos/<bslug>_<idx>.jpg (so expiring photoUri URLs never break pages), and stores authorAttributions (Photo: <displayName>, Google Maps).
13. Execution Role Split (13:20 GST): "I wouldn’t use Grok for coding.. audit code yes... although the plan I have given it, it should be able to execute that.. I would trust GLM 5.3 Flash with that plan. Then use a frontier to audit/steer."
14. Builder (GLM 5.3 Flash / Claude): Executes bounded stories (S00–S21) and community rebuilds.
15. Independent Frontier Auditor (Grok): Never writes application code; audits diffs, runs deterministic test/eval/lint gates (F01–F36), and controls promotion to integrated and live.

10.2 What Was Built & Independently Verified by Grok ([V]) Today in Staging (USD 0.00 Spend)

Note on status terminology: In Gully's 6-rung status ladder (source-present $\rightarrow$ implemented/wired $\rightarrow$ locally tested $\rightarrow$ independently reviewed $\rightarrow$ integrated $\rightarrow$ live), [V] below means independently reviewed in staging (~/ClaudeWork/cube-stage/), not integrated into live production or live until published and verified on buydubai.io.

1. Master Registers (~/ClaudeWork/cube-stage/registers/) — PASS [V] (Verified by Grok in 04-grok.md and 08-grok.md):
2. validate_registers.py outputs RESULT: ALL CHECKS PASS (0 gaps, page_dir_unresolved.csv = 0 rows, page_dir_retired.csv = 3 rows, page_dir_map.csv = 61 rows [V], 08-grok.md).
3. verify_registry_formulas.py outputs ALL 27 CHECKS PASS across all 55 cube metric families (192 (metric, slice_type) pairs [V], 04-grok.md).
4. Pre-Turn 8 snapshot preserved at ~/ClaudeWork/cube-stage/registers/versions/2026-10-24-pre-turn8/.
5. Deterministic Facts Cube (~/ClaudeWork/cube-stage/RealEstate/facts/facts.parquet + cube_meta.json) — PASS [V] (Verified by Grok in 06-grok.md):
6. Fixed all 6 cube defects (LR beds IS NULL causing 972 NULL fact_id rows, lr_* n < 10 suppression, value IS NULL suppression on 10,005 rows, deterministic arg_max LOW reason in Coverage 7b, round(value, 9), sum(annual::decimal(18,2))::double money in RC, deterministic order by sl in rent_hubs.py, and deterministic JPM from registers/out/building_alias.csv + buildings.csv + communities.csv with loud assertion).
7. Verified by Grok (06-grok.md) across independent back-to-back builds:
   • SELECT * FROM b1 EXCEPT ALL SELECT * FROM b2 = 0 rows in both directions [V] (06-grok.md)
   • Rows: 9,645,491 (1,009,660 suppressed) [V] (06-grok.md)
   • fact_id IS NULL: 0 [V] | duplicate fact_id: 0 [V] | unsuppressed value IS NULL: 0 [V] | unsuppressed lr_* with n < 10: 0 [V] (06-grok.md)
   • cube_hash: 74bc4c527cb8716a6d41bda00b7a7e64da4b0cc1703b2b1b37179ad56e65b3a6 (sha256 54e3870d585b18c9...) [V] (06-grok.md)
8. Staged Town Square Pilot (~/ClaudeWork/cube-stage/site-after/town-square/) — PASS [V] (Verified by Grok in 09-grok-gate.md, 09g_eval.txt, 09g_lint_all.txt):
9. Rebuilt /town-square/ hub + all 13 sub-pages (insights, traffic, yield, budget, towers, parking, collections, studio-apartments, 1-bedroom-apartments, 2-bedroom-apartments, 3-bedroom-apartments, 3-bedroom-villas, 4-bedroom-villas) against the Turn 6 cube (cube_hash 74bc4c52...).
10. community_eval.py town-square --local ~/ClaudeWork/cube-stage/site-after: 91.9 / 100 PASS [V] (Data 35.0/35, Content 26.0/30, Place 15.9/20, Quality 15.0/15, 41 internal links checked / 0 broken, canonical https://buydubai.io/town-square/, 11/11 sub-pages in sitemap, 15/15 JVC hub sections present including #videos, #videos-say, #due-diligence; verified in 09-grok-gate.md & 09g_eval.txt).
11. lint_pages.py ~/ClaudeWork/cube-stage/site-after ~/ClaudeWork/cube-stage/RealEstate/facts/cube_meta.json: 0 problems across all 14 pages in /town-square/** [V] (14/14 pages carry a valid facts-manifest with cube_hash 74bc4c52...; verified in 09-grok-gate.md & 09g_lint_all.txt).

10.3 Integration with Mark Gullickson's ("Gully" / INVRT) 24 October 2026 Implementation Handoff

Gully's 24 Oct delivery (BuyDubai-Dalia-Implementation-Handoff.pdf, Technical-Contracts.md, Build-Stories.md, copied to ~/ClaudeWork/cube-stage/) defines 9 milestones (M1–M9), 22 finite stories (S00–S21), 36 acceptance fixtures (F01–F36), and a 6-rung status ladder: 1. source-present $\rightarrow$ 2. implemented/wired $\rightarrow$ 3. locally tested $\rightarrow$ 4. independently reviewed (staging [V]) $\rightarrow$ 5. integrated $\rightarrow$ 6. live.

10.4 Mapping of Build-Stories.md §6 Unresolved Implementation Choices (Owners & Current Status)

Every unresolved implementation choice listed in Build-Stories.md §6 now has an explicit owner, architectural decision, and current status:

10.5 Strict Read-Only Rule & Verified S00 / M1 Baseline Findings (Peninsula Three & 7 SQL Functions)

• Strict Read-Only Rule (09-grok-gate.md §3.5): Both the Peninsula Three 156 vs 35 root-cause trace (M1) and the 7 SQL function exports (S00) are strictly read-only (SELECT, pg_get_functiondef, pg_get_viewdef, and local SQL file reads only — zero database writes, zero TRUNCATE, zero schema mutations in S00).
• Read-Only Peninsula Three Root-Cause Trace (M1 / F06 / F35) — Proven Today on Studio:
• Canonical Identity in Master Register (registers/out/buildings.csv):
  • building_id = BL001215, name = Peninsula Three, pf_slug = business-bay-peninsula-peninsula-three, community_id = CM0020 (business-bay), dld_area = BUSINESS BAY, dld_project = PENINSULA THREE, dld_building = PENINSULA THREE, n_sales_2025_homes = 150, n_leases_2025_26 = 449.
• Why the Website Showed 156 (now 158 TTM in Turn 6 Cube) vs 35 in onto_c.tower_card:
  • In facts.parquet (cube_hash 74bc4c52...), deals|tower=business-bay-peninsula-peninsula-three|sub=homes|st=all|beds=all|ttm=ttm is 158 total TTM residential sales (137 st=ready + 21 st=offplan, volume AED 283,314,360; 156 at the slightly earlier static site build snapshot; 132 in calendar year 2025 all st=offplan, followed by 50 in 2026-Q1 [49 ready], 49 in 2026-Q2 [47 ready], and 38 in 2026-Q3 [37 ready]).
  • By contrast, onto_c.tower_market (35 sales in 12 months, ONTOLOGY.md §"The 10 test questions" Q2) queried an un-reconciled Supabase slice before the 2026 dld_live ready-transfer rows and pf_units building crosswalk were joined.
  • Resolution (S04–S05, S13): Both /api/dalia and the static website builders must read from the same canonical MetricObservation / facts.parquet (BL001215 / business-bay-peninsula-peninsula-three) with explicit cohort (st=all 158 vs st=ready 137 vs st=offplan 21) and period (ttm).
• Why Catalogue Status is "Completed" vs DLD Register "ACTIVE" (93.1%):
  • All 132 sales in 2025 were st=offplan. In 2026-Q1..Q3, 133 of 137 sales are st=ready (title-deed transfers) and 449 Ejari leases are active (ads|rbld=business-bay/peninsula-three = 136), proving physical handover occurred in early 2026.
  • However, DLD's dld_projects construction escrow register lags physical handover and still records PENINSULA THREE as "ACTIVE" at 93.1%.
  • Resolution (S02, F03): CanonicalProjectRecord carries both catalogue_status = "Completed" (proven by 137 ready sales + 449 Ejari leases) and dld_register_status = "ACTIVE" (dld_completion_pct = 93.1) with an explicit reconciliation_notes explanation instead of silently overwriting either source.
• Read-Only Inventory of Complete SQL Definitions Located on Studio (S00):
• mart.tower_12m: Complete untruncated materialized view DDL located in ~/ClaudeWork/RealEstate/warehouse/mart/mart.sql (lines 166–195).
• rag_search (hard-filter v1): Complete DDL in ~/ClaudeWork/RealEstate/corpus/sql/003_rag.sql (lines 60–115).
• rag_search2 (RRF k=60 + priors, soft community/building boost) + rag_search wrapper: Complete DDL in ~/ClaudeWork/RealEstate/corpus/sql/003c_alt_rag_search2.sql (lines 25–113).
• rag_retag_docs + rag_search_v2: Complete DDL in ~/ClaudeWork/RealEstate/corpus/sql/003c_rag_fixes.sql (lines 19–115).
• onto_g.resolve, onto_g.neighbors, onto_g.card, onto_g.near, onto_g.near_entity, onto_g.docs, onto_g.metrics: Complete DDL in ~/ClaudeWork/RealEstate/ontology/grok/sql/005_tools.sql (lines 1–65).
• Remaining in-DB functions (onto_c.tower_card, rag_retag, local hybrid_search): Exported read-only via SELECT pg_get_functiondef(oid) in S00 before any M2/M5 migration.