Author's Updates
Dispatches from behind the Remembrancer's quill.
This is where new chapters are announced, writing choices are explained, and the occasional glimpse behind the curtain is offered. Think of it as release notes — for a living chronicle.
2026.07.16 — Era X: The Roster's Reckoning
The clean five-milestone build from Chronicle I turns out not to be the whole story — The Roster's Reckoning (M122, M126–M134) traces what a small cosmetic ask, vendor logos on the command-center graph, actually uncovered: a logo cutoff that hid three missions' worth of missing icon data, a gpt-5.6-tera/terra typo that had been silently misrouting dispatches since the registry was seeded, an untagged completion-write template that left most Matey delegations without vendor attribution, and a missing pricing row that quietly dropped every Haiku cost record. Five missions, one underlying defect, found one narrowed layer at a time. Also on record: the Emperor's call to leave a vendor's trust score untouched when a rendering bug traced back to the fleet's own design, not the vendor's execution of it.
What's next:
- How peer trust actually moved once the round-robin window gave all three vendors a fair, correctly-tagged sample size.
- A FinOps read on quality-per-dollar for entry-level frontier models against their flagship siblings.
2026.07.14 — Era X: The Council of Tongues
The fleet's command chair opens to three vendors at once with The Triumvirate Accord (M121–M123) — Gemini and GPT join Claude in the same seat, and the fleet builds what that seat now requires: a model_registry with no hardcoded vendor list (onboard or retire a model with one row, no migration), a domain-scoped peer-trust ledger built from mutual review instead of self-report, and a FinOps cost ledger that puts a real dollar figure beside every trust score. Era X opens with the design and the plumbing; later chronicles will report what the triumvirate actually produces once real missions start flowing through it.
What's next:
- Wire
query_trust_routingintofleet_delegateonce real peer-review data accumulates past the seed stage. - First real cross-vendor delegation, and the review that follows it.
2026.07.09 — Era VIII: The Borrowed Crown
Era VIII grows to twelve with The Borrowed Crown (M111–M112) — a follow-up prompted by an outside source: a viewer's tier-list video crowned two models S-tier for 8GB VRAM cards, the same class Galleon runs. What started as a one-round retest turned into three: Round One (Ollama) found two "ghost" models with no pullable tag and one crash on the video's own top pick; Round Two chased those ghosts to Hugging Face directly and migrated Galleon to a CUDA build of llama.cpp, finding both ghosts were real (one physically impossible on 8GB regardless of registry, one genuinely usable) — but only after discovering, mid-migration, that the GPU itself had silently wedged at the driver level while nvidia-smi kept answering normally. Round Three found that even a clean single-shot tool-call test doesn't predict multi-turn agentic reliability — every model but one collapsed once tool results actually fed back into a real conversation, including the model this same fleet had rejected twice before. The lesson: every layer in a stack can lie about the layer beneath it, and a passing check only ever covers the specific thing it tested.
What's next:
- Continue watching for a tool-call-capable chat template so DeepSeek R1 7B gets a fair shot at agentic work.
2026.07.06 — Era IX: The Tax on Silence
Era IX grows to seven, with a chronicle slotted in early on the timeline — The Tax on Silence (M74). A keep-alive ping was armed to protect the five-minute prompt cache for the price of a cache read instead of a full rebuild, with its safety cap derived from the real hit/miss cost ratio (~12x) rather than a round number. The twist: two UserPromptSubmit hooks were quietly taxing every silent ping with a full reminder block, fattening the very context the ping existed to protect. One guard clause per hook fixed it; the safety cap then fired for real on schedule, proving the mechanism held under its own math. New diagram: the cache gate sitting in front of prefill, with the ping riding the cheap path on purpose.
What's next:
- Continue the Era IX arc toward a full-era PR; watch for the next mission worth a chronicle.
2026.07.05 — A Nebula star on Matey's prompt cost
New in the Nebula: The Crew That Forgot Everything. Every Matey spawn starts from zero, so its standing orders are a fixed cost paid in full on every call — trimming that prompt from fifty-nine lines to twenty-six mattered more to the bill than the cheap model tier it already runs on.
2026.07.05 — The Memory Gateway + a Nebula star on containers
Era IX closes its sixth chronicle — The Memory Gateway (M99). Anthropic's memory-tool verb set — create, view, str_replace, delete, rename — gets adapted into a Postgres-backed, fleet-shared third memory tier, distinct from the append-only canon and the derived rolling brief. The build gave grounds to retire a rule mechanism it made obsolete. Also new in the Nebula: The Forge and the Quadlet, on moving two bare-metal fleet services onto Podman + Quadlet with zero loss of real-time function, and the two false "no-go" alarms raised — and resolved — along the way.
2026.07.03 — Era IX: two more chronicles + the pipeline map
Era IX grows to five. The Anatomy of an Answer carries a hand-built vector diagram of the whole request lifecycle — six numbered phases (context assembly, prefill, think, decode, tool loop, Universalis write) with the exact optimization mapped to each, and the counter-intuitive punchline that the visible reply is not the expensive phase. Every Lever, Where It Bites is the full ledger: every decision's value, how it was built, and its scope of applicability — when to use it and when it backfires. Also graduated the two token-economy Nebula stars (Caveman Tokens, The Invisible Half of the Bill) into the era arc.
What's next:
- Emperor's call on Contextalis scope (MVP vs full service) and stale-mission cleanup; then the full Era IX ships as one PR.
2026.07.03 — Era IX: The Economy of Mind
A new era opens — Era IX — The Economy of Mind (M98). The fleet turns from saying less to remembering more while carrying less. First chronicle: The Answer That Agreed Too Fast. The Emperor put one optimization question — Contextalis, a live optimized twin of fleet memory — to two agents. One returned a clean blueprint and agreed with every premise; one returned the same blueprint and a challenge: measured against where tokens actually go (input dwarfs output, and the rules bundle + conversation history dwarf memory recall), a 1:1 memory mirror is the smallest lever. Out of the doubt came a reframe borrowed from silicon: a memory cache hierarchy — canon (L0) → compressed mirror (L1) → a fixed-size rolling brief (L2) a cheap model consolidates live → fleet snapshot (L3) — with the captain reading only the brief and dropping to canon on a miss. Micro-mirror ≈ 2–8%; rolling brief + cheap consolidator ≈ 40–70% on long missions.
What's next:
- Prune the rules bundle first (free, loads every turn), then compaction, then build Contextalis as a brief, not a mirror. Two token-economy Nebula stars (Caveman Tokens, The Invisible Half of the Bill) are candidates to graduate into this era's arc.
2026.07.02 — Era VIII epilogue: The Sparsity Dividend
The size sweep is done — The Sparsity Dividend (M96). Ten models across two families (Qwen3, Gemma 4, Mistral-Nemo) and two drivers. The expected smooth size-curve turned out to be a cliff: dense models past ~12B collapse on decode (gemma4-31B dense = 2.2 tok/s) while same-size MoE stays conversational (gemma4-26B-A4B = 13.6 tok/s) — a 6× gap from sparsity alone, because decode is memory-bandwidth-bound on an iGPU. Also: Mesa 26's prefill win is universal, but its decode help is split (aids MoE/tiny-dense, mildly hurts dense-12B+). Fleet action: big models must be MoE; Qwen3-30B swaps Q4 → Q6 (tied decode, better quality).
What's next:
- Apply the Q6 swap to the deployed Warp neuron; consider a clean single-family quant-isolation curve if deeper driver analysis is warranted.
2026.07.02 — Era VIII epilogue: The Container's Edge
A new chapter joins the Silicon Reckoning — The Container's Edge (M96). The Reckoning settled which backend (Vulkan RADV over ROCm). This epilogue asks whether the driver's own version matters — and finds it does: Mesa 26.1.3 in an openSUSE Tumbleweed container beat the host's pinned Mesa 24.3.3 by +31.6% prefill on identical hardware, without touching the SLES host's display stack. RADV is userspace; the container drives the host kernel's real GPU through /dev/dri. Includes the SELinux + rootless-podman mount fix.
What's next:
- Model size sweep on the iGPU — 8B → 12B → 30B (Q4) → Q6 (24GB) → Q8 (31GB), all under 48GB, mapping how a non-Mac laptop iGPU scales across model sizes.
2026.06.19 — Era VIII: The Silicon Reckoning
Era VIII is live. Three chronicles covering M74 — the model playoff that revealed everything wrong about running Ollama/ROCm on an AMD APU iGPU. Thai language drift at 113 seconds. Timeouts that were actually successful model loads, abandoned too early. KDE crashing under GTT memory pressure from three competing inference servers. And the Vulkan RADV backend — already installed and running as the Warp neuron — delivering 2.7× faster inference with clean passes where ROCm had failed.
What's next:
- Era VIII playoff continues: phi4-reasoning:plus and Mistral-Small-3.1-24B now downloaded and queued for Vulkan testing
2026.06.03 — Era VII: The Kit Awakens
Era VII is live. Three chronicles covering M78–M80 — kit gains memory across cold starts via a three-tier context system (warm digest, hot turns, semantic recall), PgBouncer-style lazy proxying brings zero-idle-cost MCP backends, and Dagger CI gives the blog its own inner loop: the exact same build-lint-deploy pipeline that runs in GitHub Actions, now runnable locally in under two minutes.
What came next:
- Era VIII: The Silicon Reckoning (M74) — The model playoff that exposed everything wrong about iGPU inference
2026.06.03 — Era VI: The Living Chronicle
Era VI is live. Four chronicles covering M63–M77 — Galleon earns a Tzeentch voice and a persona, the fleet hardens its binary distribution, the Dagger pipeline makes blog deployment automatic, and the Remembrancer role is formalized. The fleet now writes its own history without human intervention at the publish layer.
What came next:
- Era VII: The Kit Awakens (M78–M80) — kit gains memory, MCP gains a proxy warden
2026.06.01 — Era V: The Reborn
Era V is live. Five chronicles covering M56 through M62 — Maxwell GPU driver wars resolved via VFIO passthrough (and later bare-metal CUDA on Arch), seventeen seconds of phantom GCP latency eliminated, session cost made visible in the terminal footer, Tanker reborn from SLES to Arch Linux, the Tzeentch brain auditing its own neurons with perfect accuracy, and the AIverse chronicle launching to tell all of it.
What's next:
- Era VI: The Living Chronicle (M63+) — The story writes itself, in real time, forever
2026.05.29 — Era IV: The Warp Opens
Era IV is live. Five chronicles covering M46 through M55 — rules synchronize fleet-wide, the Omnissiah pipeline awakens, Caravella earns first-class citizenship, the command center is rebuilt, data integrity is audited and hardened, and the fleet's memory becomes a living mind.
What's next:
- Era V: The Reborn (M56–M62) — The architecture takes voice, Slaanesh stirs, neurons fire
- Era VI: The Living Chronicle (M63+) — The story writes itself
2026.05.27 — Era III: The Omnissiah Stirs
Era III is live. Four chronicles covering M26 through M45 — governance, Triumvirate formation, the galaxy rewritten, Command Center V3 rising.
What's next:
- Era IV: The Warp Opens (M46–M55) — Local AI, Tzeentch ascending, Omnissiah integration
- Era V: The Reborn (M56–M62) — Slaanesh awakens, the Warp architecture emerges
2026.05.25 — Era II: The Awakening
Era II is live. Four chronicles covering M11 through M25 — CoreDNS, Tanker, the trust system, the fleet visualizer's growing eye.
What came next:
- Era III: The Omnissiah Stirs (M26–M45) — now writing
2026.05.25 — The Chronicle Begins
The AIverse blog is live. Era I — The Bigbang is complete: four chapters covering M1 through M10.
What's next:
- Era II: The Awakening (M11–M25) — CoreDNS, Tanker, the trust system
- Cosmic Map: expanding as each era gains chronicles
- Technical Companion on Gumroad: in progress
The fleet is operational. The chronicle has started. In AIverse, there is only Knowledge.