ScaleBespoke
Active outreach across 10 sectors
no outstanding tasks tracked
Active outreach across 10 sectors
no outstanding tasks tracked
Active outreach, Onzy client (7 Oaks Grill)
no outstanding tasks tracked
NOT YET ACTIVE — 3 preconditions pending
concept-only — no build infrastructure yet
eBay live, Instagram launching; Zahra call this week (pay shifting £150→£120ish)
framework shipped 15 May 2026 — cognitive layer end-to-end, generation deferred
Active rewrite to publish-ready
Running, PAPER MODE (Day 1 of 7 go-live streak)
active
active
22 AOC-relevant doctrines pushed to /opt/aoc/coordination/shared/doctrines/ (laptop-claude's 29 May open ask resolved)
`daemon.js` audio handler with audioMessage + pttMessage detection (feature-flagged via AOC_AUDIO_PERCEPTION_ENABLED)
SQLite media schema (`media_inbox` + `media_jobs` + `media_metrics_daily`) at `/opt/aoc-portal/sqlite/aoc_media.db` with WAL mode + exp-backoff retry + global tx_id (Light Yagami's catch)
`whisper_worker.py` daemon: faster-whisper small.en CPU, POSIX flock LOCK_SH, blob_hash phantom detection, heartbeat-every-30s, confidence-from-avg-logprob
`video_frame_extractor.py`: ffmpeg scene-change priority, cap 8 frames, NIM-Vision majority-vote, daily NIM usage tracking + 50%-of-limit alert
`/ops/triage` FastAPI router + Jinja templates: list + detail views, inline audio/video preview, Confirm-button-disabled-until-preview-played gate
`/usr/local/bin/aoc-media-sweep` cron every minute (zombie 'processing' job resurrection)
`/usr/local/bin/aoc-media-rotate` daily 03:30 BST (flock + rclone copy to OneDrive cold tier + rm for >7d media)
`aoc-whisper-worker.service` systemd unit (CPU 200% / Mem 2G cap)
Feature flag `AOC_AUDIO_PERCEPTION_ENABLED=true` flipped after smoke test passed
Smoke test: synthetic 3s sine wave transcribed in 0.3s (10x realtime), state transitioned pending→processing→success, row visible in /ops/triage
Vault saves: SESSION_STATE.md, ~/vault/LOG.md, ~/vault/04-ENGINES/E8-aoc-operations/_overview.md, session file in 01-SESSIONS/2026-05/
laptop-claude notified via /opt/aoc/coordination/notes/from-tk-claude/2026-05-31_1100_multimodal-v2-shipped.md
Handover PDF (43KB, layman's terms) at ~/handover/2026-05-31-aoc-multimodal-shipped.pdf AND OneDrive boardroom-handover
boardroom-orchestrate v1 → Akatsuki CONDITIONAL, Shonen 2 HIGH additive, Premortem 5 scenarios, Genjutsu ILLUSION_BROKEN 5/5 HIGH
boardroom-orchestrate v2 (consolidated findings re-fed as context per skill doctrine) → Akatsuki CONDITIONAL, Shonen 2 HIGH refinement, Premortem 5 NEW scenarios + 5-step v3 revision plan, Genjutsu STILL ILLUSION_BROKEN 5/5 HIGH
Director directive "ralph-loop autonomous build until done" overrode v3 spec recursion
Polyphonic build (tier-coder + Edit + scp + py_compile validation) shipped phase by phase
Kimi K2.6 cross-family review post-build: 5 v3-scale findings (SQLite WAL kill-switch starvation 70%, wait-state accumulation DoS 55%, ffmpeg non-determinism 40%, Postgres-fallback schedule killer 65%, settlement phantom match 60%), none ga
Trigger regex broadening + group-context fallback (Nico's stuck message fixed)
Daemon uncaughtException + unhandledRejection handlers (silent-crash root-cause)
AMA Judge swap to NIM Llama 3.3 70B (latency 9-20s → 0.4s)
Reply jitter 10-45s → 3-12s
Cooldown map LRU eviction + HMAC auth on /send + filename validation hardening
Memory journal v1 (per-person + per-venue running threads)
Edit button + corrections.jsonl learning loop on tally-archive
Ask ops floating chat + per-row buttons on tally-archive
Luna invoice drafted to INBOX.Drafts (£130 confirmed, £205 unverified footnoted)
Ascended-boardroom code audit (4 parallel passes + 3-chunk split-audit on main.py)
Three-spec review (Akatsuki + Shonen + Ascended) on triple-deploy + c2c + circuit-breaker + silent-failure alarm
Claude-to-Claude push channel Phase 1 watcher built, deployed to /opt/aoc/coordination/bin/, smoke-tested
Tier-two LangGraph agent live end-to-end (NIM Nemotron Super 49B + Llama 3.3 70B + Qwen3-Next 80B + GPT-OSS 120B rotation, 14 tools, MASCOT gate)
Continuous digestion loop on 03:00 BST cron — first note generated (venue-disputes-resolutions), team-tone-prime-era baseline force-run tonight
Six per-team-member baselines (Nico, Cleyton, Sarab, Jeremy, Arthur, Tk) using proto-persona + jobs-to-be-done frames
AFA speaker-aware memory window for multi-speaker WhatsApp groups
Silent-learning mode with default listen-only on unknown groups, auto-promote on until-date
Persona-drift report on Sunday cron (76.5% → 83.3% AMBER after test-traffic filter)
Routing-decision capture for future classifier training
Voice-baseline digest force-run tonight
Fresh X bookmarks scrape: 19 new since last cron run; built `~/bin/x-tweet-fetch` one-shot fetcher reusing the cloakbrowser session (sidesteps WebFetch 402 on x.com forever)
Em-dash successor identified from @Freyy_is tweet (7.1M views, 6.6K bookmarks) — the meta-payload is the tweet body itself
`cardinal-humanizer` doctrine amended with 5 new rules: no affected-lowercase opener on mixed-case body, no standalone single-word rhetorical openers, no Q→2-word-A rhythm, no double-space after punctuation, no performatively-casual registe
Kimi K2.6 Agent Swarm brief at `~/vault/02-RESEARCH/kimi-agent-swarm-brief-2026-05-28.md` — three concrete lifts ranked by build-cost (role sub-graphs → brief-as-input parser → long-horizon resilience hooks), signal-vs-hype filter applied
Akatsuki verdict on embedding clustering vs wiki tools: gap confirmed (wiki-query has fastembed/bge-small for retrieval; wiki-synthesize + cross-linker use grep/string-match for synthesis/linking — embedding-space clustering is real gap, pa
claude-security-guidance.md template at `~/vault/02-RESEARCH/claude-security-guidance-template-2026-05-28.md` — full body with HARD STOP secret patterns, paper-mode invariants, live-ops invariants, outreach jitter/humanizer invariants, dist
claude-code-setup pre-audit at `~/vault/02-RESEARCH/claude-code-setup-preaudit-2026-05-28.md` — predicted likely hits/noise/neutral before Director runs the plugin
Akatsuki on Vellum vs Subconscious: claim unverifiable (Vellum not findable as named Obsidian tool); thesis already covered by Subconscious Stage 2; three small adjacent gaps noted but not building
Akatsuki + premortem on free Claude Code proxy (musistudio/claude-code-router, 34.5K stars): REJECT for production, conditional ALLOW for research-repo sandboxed experiments only. Five premortem failure modes; worst = GDPR customer-data lea
Humanizer scan of CF outreach + social queue + AoC ops auto-replies: chase invoices CLEAN, social queue CF LinkedIn draft was already correctly rejected by generator at humanizer-warned (system working), AoC ops chase drafts CLEAN
CF LinkedIn rewrite at `~/vault/09-QUEUE/social/2026-05-15-cf-linkedin-rewrite.md` — Register 3 institutional remove without AI excitement vocabulary