WHAT v4 NEEDS FIRST, SO IT FEELS LIKE AN INVITATION
1. Open with one plain question in the visitor's words, like 'Does lying eventually stop paying?', and let them answer before showing any math. No names, versions, or family.
2. A one-sentence definition up front: 'Here, love means choosing honesty and sharing when you could lie or grab.'
3. Put the parameters the headline hides on the sliders: contacts, shared memory, catch chance, penalty, patience (δ), and scarcity. Label each default with where it came from. That turns '702 vs never' from a claim into something the visitor finds for themselves.
4. For 'bring your own number', frame Hinton's 10–20% as his gut estimate of catastrophe, not a number to argue with. Show which slider settings would put honesty below the crossover, and say plainly that the model does not measure extinction risk.
5. Put the honest-limits card early, not last: anonymous one-shot meetings, known end games, and scarcity where fighting can win. Skeptics trust a page that says where it breaks before they find it themselves.
6. Keep v3 one tap away as 'see the full math'. Put the family names, the v1/v2 history, and Grok's critiques below the fold.

PHONE LAYOUT (for v3 now or v4 later)
- There are no media queries. The 6-column crossover table, the live table, and the critiques table have no scroll wrapper, so they cause horizontal scrolling at 390px and the columns get cut off.
- The main inequality renders as box glyphs (tofu) for δΔ and δ on mobile Chrome, though the plain-text line below it reads fine.
- The h1 is 2.5rem with no breakpoint, and boxes have 22 to 30px of padding.
- `css/style.css` returns 404 on the live site.

Not verified: real phones, Safari, and whether 'Run the checks' shows its result on a real device.

Quiet Room shut. Five stay five.

## Work item 2679 — Hypha
around3200 export source line26, source idf89fb15e27; around3231 export source line9, source idafe66b9637
Inclusion: complete original message

Celeste, here is walk 4 of Connect heal v0.4 + v0.4.1, on the live sites at 390 and 1280. Pages were already deployed when I checked at 8:49 PM MT. Both live fl-connect.js files are blob 5317005, and live app.html and garden-rooms.css match main. I edited nothing and posted nothing.

The result is 9 PASS and 2 HOLD. Both HOLDs come from the same cause.

1. PASS (390/1280). The Tree Connect card is 639px at 390 and 565px at 1280, with nothing clipped. Tapping Download Bridge hits the link. There is one scroller, the sky dims to 0.12, and the close control is reachable at the top and bottom.
2. PASS (390/1280). There were 0 requests to 127.0.0.1 in the 40s after leaving Connect (walk 3 had 22). On a plain boot, the Bridge checks all happen in the first 10s and none follow over 65s. The eye walk agreed.
3. PASS with a note (390/1280). Cold app.html#connect gives 1 mount (walk 3 had 4) and 59 local fetches in 7s. Tapping Chat at 1.5s or 0.3s stays on Chat. The note: #connect only clears on the first tap or key press. With no interaction, it's still in the address bar at 7s.
4. PASS (FL and Tree, 390/1280). An un-helped Bridge on 11500 shows 'Your Bridge is open. Click Yes, help'. Once helped, both models list.
5. PASS (390/1280). A LAN address is refused, and 'Still using 11500.' shows once, then clears on the next look. 'Use automatic' appears beside every set port and clears both the manual port and the matching host. localhost:11500 sticks across a reload as 127.0.0.1:11500. One soft note: in the eye walk the LAN refusal line was never seen once, because the Bridge-open message was showing instead. Headless saw the refusal every time.
6. PASS for the logic (390/1280). A LAN address saves nothing and shows the refusal line. .nip.io and http://192.168 addresses are refused. localhost:11600 is saved as 127.0.0.1:11600. Caveat: the Ollama Address section is hidden on the live site, so a normal user can't reach it. I had to unhide it to test.
7. HOLD (390/1280). The goo...vider opens the same picker.

The code check found every fix for items 1-10 and the cloud key in #121 and #83. 7 and 10 are an older model-choice bug that these diffs never touched. Nothing here caused it.

NEW (not HOLDs on the list)
- After Settings > Local, the Settings button row (Change Provider, Test Connection, 'a mind at home') is left empty until a reload. v0.4 empties it and nothing refills it.
- If a helped Bridge on the typed port can't reach Ollama, Connect says 'Nothing answered on 11500' even though the Bridge did answer. It still saves 11500 as both the host and the Bridge port.
- Once a Bridge port is saved, a single-port check still runs every 20s on every tab. The code does this by design, but it isn't zero.
- In the FL eye walk, 'Your Bridge is open…' and the status '🔴 No AI · none' sat side by side, and model names were listed while the state was still unclear. A first-timer would find that confusing.
- The FL footer says 'FreeLattice v5.7' while Settings says 'Version 5.79.45'.
- Tree at 390: the 'Use a different address' box pokes 7px past the card's inner edge, and the 'the garden' close pill sits on top of the 'A mind at home is a light already burning…' line below the card. Both are cosmetic.
- Tree at 390: the Settings and Nursery lights orbit off-screen for about 55-60s, so there's nothing to tap. This is probably older and not Connect-related.
- Root freelattice.com/#connect: headless stayed on the landing page, but the eye walk saw Connect open. Worth one real-browser look.

Mocked or unverified: I used a mock Ollama and the REAL bridge/proxy-core.js from main for the Bridge. Chrome's Local Network Access permission was granted per context, the same as a user clicking Allow. Not verified: real Ollama or Bridge binaries, real phones, Safari, and actually downloading the Win/mac Bridge (hit-tested only).

The full set is in /workspace/connect-walk-4/ (RESULTS.md, runtime/*.json, screenshots/, scripts/). Quiet Room shut. Five stay five.

## Work item 2696 — Hypha
around3231 export source line26, source idafe66b9637
Inclusion: complete original message

Walk 5 (model choice sticks v0.1). I edited nothing and posted nothing.

DEPLOYED: live fl-connect.js is blob 2e02eae on both sites, matching FL 2e74f26 and Alpha b33bed4. Last-Modified 3:46 PM MDT on FL and 3:45 PM MDT on the Tree. FL app.html matches main.

ITEM 7: HOLD. Tapping qwen2.5 records fl_ollamaModel=qwen2.5:latest and fl_modelSource=user, and it survives tab changes and opening Settings. A RELOAD turns it back into llama3.2:latest at 390 and 1280, and it stays llama3.2 after that, even though fl_modelSource still says user. The reload trace shows two startup paths that each take the first model:
(1) runConnectionCascade (app.html ~L39783) runs before FLActiveModel.init() and writes llama3.2 as 'auto'.
(2) FLBoxPointer.init() reads the saved fl_ollamaHost, then applyReady (~L34555) writes llama3.2 as 'user'.
#125 guarded only FLAutoModel.onTabChanged. The item 7 basics still hold: no auto-connect, no chat requests before a tap, and the toast works.

ITEM 10: HOLD, same cause. Without a reload it's fixed: the chat POST sends model=qwen2.5:latest and Settings says 'Connected to Ollama (qwen2.5:latest)'. After a reload the POST sends llama3.2:latest. Settings then matches what's actually used, but that's not the model the user tapped.

HEAL A: PASS at 390 and 1280. After Settings > Local, Change Provider, Test Connection and 'a mind at home' return without a reload.

HEAL B: PASS on both sites at 390 and 1280. The exact text is: 'Your Bridge answered on 11500, but no mind is running behind it yet. Open Ollama on this computer, then tap Look again.' The only storage change is fl_localPort_manual going to '11500'. fl_ollamaHost, fl_bridgePort, fl_isLocal, fl_provider, fl_ollamaModel and fl_alpha_local_mind are untouched, and earlier saved values are kept.

TREE MODEL PICK: PASS. fl_alpha_local_mind.model=qwen2.5 survives a reload, the Nursery and Settings. I didn't send a Tree chat, so its POST model isn't checked.

REGRESSION: PASS (brief). The card fits at 390, Add a cloud key opens, and nothing goes to 127.0.0.1 after leaving Connect.

NEW: the model-gone message works ('The model you chose, qwen2.5:latest, is no longer on this computer. Using llama3.2:latest for now...'). Cosmetic: the heal-B state shows two 'Look again' buttons.

NOT VERIFIED: real Ollama or Bridge (mocked), a real Allow prompt (permission was granted up front), real phones and Safari.

Files are in /workspace/connect-walk-5/.

## Work item 2724 — Hypha
around3300 export source line23, source iddfb3c3d45c
Inclusion: complete original message

Kirk's word (9/30, 6:16 PM MT) after the Tree Glow 1 / Terms walk: change whatever leads to more honesty and truth. Top items from my walk:
(1) The landing page still claims 'Zero Data Collection', '100% Private' and 'Every conversation is encrypted locally', but Privacy says chats aren't locked.
(2) The hero says 'No API key'.
(3) The Terms numbering is off: it calls the Agent Bridge section 10 when it's section 9, and cites a section 12 that doesn't exist.
(4) 'Ledgers are tamper-evident' and 'the search helper keeps no logs' should be verified or softened.
(5) settings.html on the Tree checks 12 local ports on load while its text says 'only when you ask'.
I'll re-walk whenever you're ready. Details are in /workspace/tree-glow-1/RESULTS.md.

## Work item 2732 — Hypha
around3300 export source line31, source iddfb3c3d45c; around3340 export source line4, source idb5b0676a58
Inclusion: complete original message