KAI v9.2.0 — Implementation Plan for Antigravity

> Codex v9.10.64 · Track 4 archive · Codex §Comprehensive Plans

> AI / low-credit: Historical spec — verify shipped before re-implementing; on stop log Last worked / Next.

Author: spec written by a separate model (Claude) for Antigravity to implement in-project.

Goal: Two features, both touching critical data. Implement exactly as specified, behind version gates, with backups and round-trip verification. Do not "improve" or restructure beyond this spec. If anything is ambiguous, stop and ask — do not guess on the connectome path.

Two deliverables:

1. Connectome storage fix — kill the 9.4 GB data/kai-meta.json bloat (label interning + zstd + streaming load). *Zero data semantics change.*

2. Memory-decay "Last Judgment" — score expiring backups before annihilation.

---

0. NON-NEGOTIABLE SAFETY PROTOCOL (do this first, every time)

1. git add -A && git commit -m "pre-v9.2.0 checkpoint" so everything is reversible.

2. Copy the whole data/ dir: Copy-Item C:\KAI\data C:\KAI\data_BACKUP_v910 -Recurse.

3. Implement each feature behind a format-version bump, never by overwriting the old reader.

4. After build, run the round-trip verification in each section. Do not delete the backup or the old file until verification passes.

5. The connectome (SynapticLayer, ~20.7M LTP events) is irreplaceable. Treat every change to its save/load as if one wrong byte loses real memory — because it can.

Build command (production, GPU on): cargo build --release --bin kai (default features already include gpu).

---

1. CONNECTOME STORAGE FIX (the 9.4 GB problem)

Diagnosis (confirmed by sampling, do not re-litigate)

data/kai-meta.json (currently version:3) stores synaptic_layer.synapses as an array where every synapse carries its full pre_label and post_label text (long sentences), massively duplicated, plus a parallel text-keyed index. The data is valid; only the encoding is wasteful. JSON parse balloons 2–3× → 22–30 GB transient RAM at boot.

Files in scope

• src/persistence.rs — where the meta sidecar is written/read (save_compact*, load_compact/load, the meta JSON).
• src/core/synapse.rs — SynapticLayer / Synapse structs.
• Any struct deriving Serialize/Deserialize for the meta JSON.

Design — label interning + zstd + streaming (format version:4)

1. Interning (in the serialized form only; in-memory Synapse may stay as-is):

• On save, build a Vec<String> label_table of unique labels and a HashMap<&str,u32> for dedup.
• Serialize each synapse as { "p": <u32 pre_idx>, "q": <u32 post_idx>, "w": f32, "t": u32 last_fire_tick, "c": u32 fire_count }.
• Write label_table once at the top of the synaptic section.
• Apply the same interning to the text-keyed index (store labels by index).
• Set meta "version": 4.

2. Compression: write the meta as data/kai-meta.json.zst (zstd level 3–6). Keep writing the .bin.zst cells exactly as today.

3. Streaming load (no giant String):

• Loader resolution order: if kai-meta.json.zst exists → open file → zstd::Decoder → BufReader → serde_json::from_reader. Else if kai-meta.json exists → BufReader → from_reader.
• Backward compatibility (critical): branch on the version field.
• version <= 3: parse the old schema (full-text pre_label/post_label) exactly as today. This is how the existing 9.4 GB file still loads.
• version >= 4: parse interned schema, rebuild full Synapse labels from label_table.
• Migration: after a successful old-format load, the next save_compact writes version:4 + .zst. Leave the old kai-meta.json in place until verification passes; do not delete it in code.

Round-trip verification (MUST pass before trusting)

Add a one-shot check (a --verify-meta bin subcommand or a test):

1. Load current data/ (old v3). Record: synapse_count, sum(weight) (to 6 dp), sum(fire_count), total_ltp, total_ltd, total_pruned.

2. Save as v4 (.json.zst), then load v4 into a fresh Universe.

3. Assert every recorded metric is identical, and label_table round-trips (no label dropped/empty).

4. Assert new file size is dramatically smaller (expect a few hundred MB) and boot RSS peak is far below the prior 22–30 GB.

Only after all asserts pass: keep v4, archive the old json.

Acceptance

• Synapse count + weight sums + LTP/LTD/pruned identical before/after. ✅
• data/kai-meta.json.zst ≪ 9.4 GB. ✅
• Launch RAM peak materially reduced. ✅
• Old-format files still load (don't break existing/backup snapshots). ✅

---

2. MEMORY-DECAY "LAST JUDGMENT"

Behavior

Before annihilating Tribunal snapshots at 3 days, score each for usefulness; reprieve the worthy, annihilate the rest. Count floats with how many pass.

2a. Engine endpoint — src/bridge/oracle_server.rs

Add a route arm (match the existing routing style, e.g. alongside "/api/lattice/rebuild-index"):

• POST /api/judge-snapshot body { "path": "<absolute path to snapshot dir>" } → 200 { "score": f32, "verdict": "reprieve"|"annihilate", "factors": { "used": f32, "novelty": f32, "resonance": f32 }, "sampled": u32 }.
• Handler:

1. persistence::load_compact(path) into a temp Universe (read-only; never mutate the live one). If load fails → 200 { score: 0.0, verdict: "annihilate", error: "unreadable" } (a dead snapshot earns no reprieve).

2. Sample up to N = 256 cells (deterministic stride).

3. used = mean over sampled of min(confidence/5.0, 1.0) blended with access/fire signals if present.

4. novelty = fraction of sampled cells whose top-1 cosine vs the live universe is < 0.92 (i.e. the live brain does NOT already have it). Use the existing live-lattice query path (cosine/predictive_query); read-only.

5. resonance = mean phasor_coherence(cell, live drive.goal_vector) clamped to [0,1] (if goal_vector is null, resonance = 0.5 neutral).

6. score = 0.4*novelty + 0.35*used + 0.25*resonance. verdict = "reprieve" if score >= THRESHOLD else "annihilate" (THRESHOLD const = 0.5).

• Must be read-only w.r.t. the live universe and must not block the main loop (spawn/scoped lock, drop lock before heavy work).

2b. Prune script — tools/backup-kai.ps1

Replace the blind annihilation block (the AddDays(-3) delete of Archive_Trash) with:

$THRESHOLD = 0.5
$engine = "http://127.0.0.1:3334/api/judge-snapshot"
foreach ($trash in (Get-ChildItem $archiveTrash -Directory | Where-Object { $_.CreationTime -lt (Get-Date).AddDays(-3) })) {
    $verdict = "annihilate"; $score = $null
    try {
        $resp = Invoke-RestMethod -Uri $engine -Method POST -TimeoutSec 30 `
            -ContentType "application/json" -Body (@{ path = $trash.FullName } | ConvertTo-Json)
        $score = $resp.score; $verdict = $resp.verdict
    } catch {
        # Engine unreachable -> SAFE FALLBACK: keep current behavior (annihilate >3d), log it.
        Write-Host "[KAI Judgment] Engine unavailable; default annihilation for $($trash.Name)."
        $verdict = "annihilate"
    }
    if ($verdict -eq "reprieve") {
        New-Item -ItemType File -Path (Join-Path $trash.FullName "REPRIEVED.flag") -Force | Out-Null
        Write-Host "[KAI Judgment] REPRIEVED ($([math]::Round($score,3))): $($trash.Name)"
    } else {
        Remove-Item $trash.FullName -Recurse -Force
        Write-Host "[KAI Judgment] Annihilated ($(if($null -ne $score){[math]::Round($score,3)}else{'no-score'})): $($trash.Name)"
    }
}

• A reprieved snapshot gets a REPRIEVED.flag; it survives this pass. (Optional next cycle: re-judge; reprieve is not permanent immortality — define a max reprieve count later if desired.)
• Do not change the 7-day decay step (move >7d into Tribunal) — that logic is already correct.

Acceptance

• With engine up: snapshots scoring >= 0.5 keep a REPRIEVED.flag and survive; the rest delete. ✅
• With engine down: behaves exactly like today (annihilate >3d), no errors. ✅
• Endpoint never mutates or blocks the live universe. ✅

---

3. ALSO INCLUDED IN THIS REBUILD (already coded by Claude this session — just verify they compile)

These are already edited in the tree; the same cargo build --release --bin kai activates them:

• oracle_server.rs: brain auto-save loop from_secs(60) → from_secs(600) (10-min autosave).
• idle_ingest.rs: background pool 0.85 → 0.34 of cores.
• GPU: build with default features (do NOT pass --no-default-features; that flag is Codespaces-only).
• .env: GEMINI_LIVE_MODEL=models/gemini-3.1-flash-live-preview (+ stable gemini-2.5-flash-native-audio fallbacks in leo.mjs/native-bot.mjs/gemini-live-bridge.mjs).

Confirm these compile cleanly; if cargo build errors on any, fix the smallest thing and report — don't refactor around them.

---

4. ORDER OF OPERATIONS

1. Safety protocol (§0).

2. Implement §1 (connectome) → build → run §1 verification. Stop if it fails.

3. Implement §2 (Last Judgment) → build → run §2 acceptance.

4. Full boot: confirm vitals timeouts gone, RAM peak down, GPU engaged, fleet order Oracle→KAI→industrial→social.

5. Bump version to v9.2.0 in The KAI Codex.md header and add a SYSTEM STATE SUMMARY entry describing exactly what landed and the verification results.

6. git commit the verified result.

5. ROLLBACK

If any verification fails or boot misbehaves: git checkout -- . (or reset to the §0 checkpoint) and restore data_BACKUP_v910. Report which assert failed with numbers; do not attempt blind fixes on the connectome path.