Skip to content

Session log (newest first)

Written for a reader with amnesia. Each entry: what exists now, where, and what happens next.

2026-07-21 - AKSARA APPROVAL BOUNDARY + AGENT HARNESS H0 SHIPPED

  • aksara-cli C1 is executable and CI-green at 28ec7c5 plus CI refresh 3c1ee98. The caller-controlled consent: true / --yes path is removed. Consequential calls now return a persisted approval request and accept only Ed25519 evidence bound to the full deterministic plan, normalized actor, scope, purpose, validity window, and nonce. Expiry, revocation, request tampering, actor tampering, malformed IDs, unknown fields, signature failure, and cross-authority nonce replay are adversarially tested. Exact retries are idempotent. Cross-language schemas and signed fixtures live under aksara-cli/schemas and testdata/approval.
  • Development truth is visible: the default RFC 8032 key is public test material, production policy rejects development_only authorities, CLI key generation cannot create a production authority, and produced artifacts say PROTOKOL DEMONSTRASI. Real approver sessions, institution-held keys, and PSrE/BSrE remain unbuilt.
  • aksara-agent H0 is now a separate private repository at github.com/yohn-maistre/aksara-agent, commit 4c00117, CI green. It contains runtime-neutral envelopes and events, a RuntimePort, a pinned Hermes v0.19.0 plugin seam, three isolated profiles, an inert PieFed adapter, one recorded unverified ActivityPub mention, and 17 contract/security tests. It has no model loop, live posting, signing, execution, or canonical-memory mutation. Profile allowlists name only MCP tools that currently exist.
  • D21 remains proposed: H0 validates the package boundary, but the decision gate still requires H1 and a private PieFed sandbox mention handled without signing keys or memory-promotion authority. Never test an agent actor on piefed.social.
  • Next: H1 compiles target profiles against the exact pinned Hermes config, consumes recorded mentions into cited drafts, and reads the separate aksara-cli response/approval contracts. C2 builds an independent receipt profile and verifier.

2026-07-21 - AGENTIC INFRASTRUCTURE + HERMES RUNTIME PLAN

  • Current architecture consolidated in the hub: specs/agentic/00_infrastructure-plan.md owns the multi-protocol shape: ActivityPub/PieFed for public life, A2A for structured tasks, MCP for tools/resources, OASF-aligned capabilities, grant and receipt profiles, aksara-cli deterministic action, and aksara-mem governed memory. The prior Honai plan is now copied into its correct system-of-record repository.
  • Hermes Agent researched against current upstream: latest verified release at review was v0.19.0 / v2026.7.20, MIT. Current upstream supports standalone pip plugins, custom gateway platform adapters, isolated profiles, MCP stdio/HTTP, MemoryProvider plugins, skills, hooks, managed update snapshots, and security controls. Research and source receipts: research/2026-07-21_hermes-agent-integration.md.
  • Proposed D21: do not build a second agent loop. aksara-agent becomes the Aksara distribution and policy package on stock, pinned Hermes: dedicated profiles, PieFed adapter, memory plugin, MCP wiring, runtime-neutral envelope, and compatibility suite. Aksara-specific code never edits Hermes core. Production pins a release; weekly CI tests pinned plus upstream main. Spec: specs/agentic/01_hermes-runtime-profile.md.
  • Security correction found in executable aksara-cli: the Go repository is a real vertical slice, not greenfield, but consent: true is caller-controlled boolean state. It is not human authorization evidence. specs/aksara-cli/06_implementation-plan.md makes replacement with plan-bound approval evidence the first implementation gate. Development Ed25519 signatures remain PROTOKOL DEMONSTRASI, not legally valid PSrE/BSrE signatures.
  • Oracle cutover plan copied into the hub: ops/04_oracle-piefed-cutover.md covers the separate etnos-infra repository, pinned upstream PieFed, CORS/API seam, backup/restore, federation checks, frontend switch, rollback, and update discipline.
  • Next: consolidate and publish the aksara-cli README, then start C0/H0 in the aksara-cli and future aksara-agent repositories. Yose ratifies D21 after the first isolated public-mention spike, not before.

2026-07-11 (evening) — SEAT HANDOFF: CLAUDE.md, all repos verified pushed

  • Yose moves to a laptop. CLAUDE.md added at hub root: reading order for amnesia readers, repo family table, live surfaces, zero-dep build commands, standing rules, paid-for gotchas, current frontier. It duplicates no state (pointers only; SESSION_LOG/LEDGER/DECISIONS stay the owners).
  • All four repos verified clean and synced with origin: abstraksi (aaf1394+), aksara-mem (1b38dbd), aksara-ui (d3583d7), honai-metro (afc56092). Deploy secrets live in GitHub Actions, so push-to-deploy works from any machine; nothing on this phone is load-bearing except the uncommitted ~/workspace/*.docx originals.
  • Parked on the honai-metro side: the ETNOS "MIDAS SWEEP" polish plan (W0-W5) is banked in /root/.claude/plans/ on the phone and summarized in that repo's last commit message; resume only on Yose's cue.
  • Next: same as afternoon entry, plus Yose's laptop session starts by reading CLAUDE.md.

2026-07-11 (afternoon) — DOCX v2 (real figures + TOC that works), GAMBAR 7 v2, hub sweep

  • Yose feedback round on v5.0: TOC empty + figures showed cream boxes in Google Docs and on his phone (svgBlip unsupported there); GAMBAR 7 section out of scale, standoffs should land on a MOUNTING PLATE not bare wall, say "kotak" not "kaset", kotak must be weather-resistant, cables should run inside the standoffs. Also: v5.1 public trim, and a Sonnet cleanup agent.
  • docx v2 SHIPPED: (1) TOC is now a pre-built hyperlinked Daftar Isi (75 entries, bookmark anchors, zero fields to update, renders in GDocs); (2) PNG fallbacks are now REAL renders done ON CI (docs.yml installs librsvg2-bin + fonts-lato, renders 1600px PNGs, runs build_docx.py before site build; final docx 1.7 MB, verified per-figure 120-200 KB); (3) public-edition trim: Identitas dan Konvensi Dokumen, Riwayat Versi, Audiens stay in the md master + hub site but are omitted from the distributed docx (this IS the v5.1 pass, as a build variant, one fact one owner). dist/ untracked (CI builds it); fresh docx re-copied to /sdcard/Download.
  • GAMBAR 7 v2: section redrawn at honest axis scales (4 px/mm horizontal, 2 px/mm vertical, stated in caption), PELAT DUDUKAN added between wall and standoffs, kotak wording, dashed cable-in-standoff channel, gasket note. Master §24 got a new paragraph: pelat dudukan (material = open decision), gasket-sealed kotak (splash+dust target TBV), no visible cables. Device-spec §9 gained Q7 (plate material options a/b/c with steel RECOMMENDED).
  • Blueprint regenerated v0.4.1: EE03/ESP32-S3 replaces Waveshare-flexible/IT8951 in the parts schedule (the sweep's biggest catch: the PUBLIC blueprint contradicted D19), kotak wording (v0.4 rev-history row kept verbatim), license clause aligned to CC BY 4.0 per D09, pemasangan note added; geometry rework (plate) explicitly deferred to v0.5.
  • Sonnet sweep (scout then executor, Yose's pattern): read-only report at scratchpad, then 6 approved items executed: 5 root legacy docs moved to research/legacy/ (git mv, references fixed), proposal/README caught up to v5.0 + blueprint v0.4, 02_gambar-pack row 7 updated, provisioning-keys Zymkey question closed (settled D10/F5), 05_design-language IT8951 facts updated to D19, 00_FRAMING Mozilla clause de-staled. DECISIONS D19 settled (A+B, by Fable). GLOSSARY Display row updated to D19 + kotak/pelat naming rows added (Display row was stale FACT, not frozen naming).
  • Next: Yose re-opens the docx (figures + TOC should now work in GDocs/phone), picks the pelat dudukan material (steel vs hardwood, Q7), and does the v5.0 end-to-end review (LEDGER 3e).

2026-07-11 (midday) — PROPOSAL v5.0: consolidation rewrite + docx generator + display A+B locked

  • Display decision LOCKED (Yose): A + B. BOM lines 2/3 now EE03 kit (Rp 2,6jt landed) + one 7.5" A/B unit (line 3b, test purchase); demo roll-up Rp 8,2-10,1 jt (was 9,6-12,0); under Rp 8jt possible via no-LTE demo-trim (~7,6jt) or Aksara Lite 7.5". Device-spec rows 2-3 updated; LEDGER 17 marked decided (cart action gated on 0c).
  • GAMBAR 7 redrawn: was a copy of blueprint v0.1; now a drawn Plakat v0.4 figure (front elevation + wall section, from ADK-DEV-BP-002 dims) in generate.py idiom. g11 date bumped.
  • proposal/05_v50-master.md NEW (82K chars, id-ID, zero em dashes): full v4.1 rewrite consolidating D01-D20 + all 12 errata. Key deltas vs v4.1: license table per D09 (Apache-2.0 + trademark for cli/libs, AGPL ETNOS, CC BY 4.0 specs); §20 rewritten as Memori Aksara (vault=truth, index rebuildable, no-profiling rule); §24 rewritten as Plakat v0.4 with verified BOM + load-shedding ladder + "tidak mengawasi orang" added to §39 batasan; §25 modular voice chain (why-modular argument: verbatim transcripts + speaker verification); §15 honest cloud-primary/local-floor framing + NPU W8A8 promise-gate; §29 Telegram-first (Meta policy) + no-single-platform principle; §49 funding table July 2026 reality (NLnet/Mozilla/PIDI out; RIIM 31 Jul, Indigo, DPG, NVIDIA, self-funded pilot) + 4-step procurement ladder (PKS tanpa anggaran → pengadaan langsung <200jt → e-katalog → APBD penuh); Lampiran E = real verified BOM. 11 GAMBAR embedded via markdown image syntax.
  • proposal/build_docx.py NEW: zero-dependency OOXML generator (stdlib zipfile; no python-docx/pandoc on host by design). Parses the master md, emits A4 docx with Kertas Dinas styles (PJS font name with viewer fallback, ink/accent/line tokens, bordered H1s, shaded table headers, footer with page field, TOC field), embeds all 11 SVGs natively (svgBlip, crisp in Word 2016+/mobile/GDocs) with generated cream PNG placeholder fallback (pure-python PNG encoder; NO rasterizer exists on this host). Output 63,8 KB.
  • Delivery: docx at /sdcard/Download/Aksara_Proposal_v5_0.docx (phone), hosted at abstraksi.pages.dev/Aksara_Proposal_v5_0.docx (link on /gambar-pack), master readable at /proposal/05_v50-master with figures inline.
  • Gotchas logged: MDX silently STRIPS raw HTML in .md pages (only .mdx gets JSX) — figures must use markdown ![]() syntax; sanitize() escapes < so whole-line img/a whitelists added to sync.mjs; git add -A swept a background agent's tmp file into a commit (removed).
  • Deploys green through 6a278ef; v5 page + docx download content-verified live.
  • Next: Yose reads v5.0 end to end (his review IS the gate for v5.0 replacing v4.1 as canonical), opens the docx on phone to confirm figures render in his viewer, reacts to Kertas Dinas docx styling.

2026-07-11 (morning) — PROPOSAL ON THE PHONE + cheap-display scout

  • Yose's asks: (1) is the GAMBAR pack done and can he read the proposal already, on the phone; (2) hunt a display path under Rp 1.5 jt landed (panel = 40% of BOM), any tech considered, projector included.
  • Proposal reading surface SHIPPED (commits 5c9c7ae + 7edf198, deploy 29135730744 green, content-verified live):
    • proposal/04_v41-full-text.md NEW: full v4.1 whitepaper text extracted from the canonical docx (python zipfile+regex; the <w:t/> self-closing-run bug initially swallowed 370 K chars of raw XML, fixed; docx Daftar Isi dropped). Honest label SALINAN BACA: edits happen in the docx, never here. Live at /proposal/04_v41-full-text.
    • /gambar-pack NEW generated page (sync.mjs): all 11 GAMBAR SVGs inline with captions + PDF-export instructions. GAMBAR 7 captioned honestly as blueprint v0.1 (geometry superseded by v0.4 Plakat). Gotcha logged: nested double quotes inside an MDX alt attribute fail the vocs build.
    • Answer to ask 1: GAMBAR pack was DONE 2026-07-05 (11/11); the whole Nabire package was already synced as pages; what was missing was the inline-images page and the master-text page, now both live.
  • Cheap-display scout DELIVERED: research/2026-07-11_cheap-display-paths.md (all lanes: budget e-paper, salvage+EPDiy, LCD kits, used tablets, projectors; ranked table; FX note USD/IDR ~18.050). Headline: Seeed XIAO ePaper EE03 kit, the SAME ED103TC2 10.3" 1872x1404 panel class + ESP32-S3 driver, $85.00 + $29.90 = $114.90 verified in Seeed page source, ~Rp 2,6jt landed = 42-46% below the IT8951 line at zero pixel/watt/D19 cost (trade: TRMNL-style architecture, OPi renders + ESP32 holds the page; partial-refresh quality via Seeed's stack is the TBV). True sub-Rp 1,5jt exists only at 7.5" (Waveshare HAT $56.99 ~ Rp 1,29jt, the "Aksara Lite" A/B candidate). Projector = documented NO: Rp 776rb-1,3jt buys 260-300 real ANSI behind "4000 lumen" labels; ~2:1 contrast in a 500-lux office and 40-60 W always-on kills the outage ladder in about an hour. Three pre-shaped options + recommendation (A: EE03) in the doc; hands in LEDGER 17.
  • Next: Yose picks display option A/B/C; device-spec §2 + BOM line 2-3 update only after the pick.

2026-07-10 (late night 5) — HARDWARE WAVE: device spec + BOM DRAFT, aksara-ui rename executed

  • specs/device/00_device-spec.md STUB→DRAFT v0.1: Plakat v0.4 form summary, full subsystem map (11 rows, each with its receipt), compute + model residency (RK3588S facts as promise-gates: NPU is W8A8 only, never "NPU + 4-bit"; Gemma 4 E4B main local, cloud primary, sherpa-onnx modular voice per the two 07-10 reports), D19 waveform discipline, power + load-shedding ladder with honest runtime arithmetic (50 Wh at 10 W = ~5 h full; panel-only = days, TBV), thermal chimney, acoustics named as THE v0.4 field risk (30 mm gap reverb; fallback variant B Ambang), six shaped open questions (PoE++ option, battery SKU, bolt-head fingerprint, CAD toolchain, acoustic field test, compute watch).
  • specs/device/01_bom.md STUB→DRAFT v0.1: 15-line component table (blueprint PARTS = master; this file adds sourcing + roll-ups), demo-unit roll-up Rp 9.6-12.0 jt landed (mid-band of the v4.1 Rp 5-20 jt envelope; panel = 40% of low bound, halves at batch via D19 production path), demo-vs-pilot paths, import notes (+20-30% landed), TKDN levers, LEDGER 16 hands. Standing answer to the old open question: web sim carries demos until funding is named (0c); nothing is bought before that gate.
  • aksara-papan → aksara-ui EXECUTED (Yose's call): GitHub repo renamed (old URL redirects), local dir + remote + package.json + README swept, hub pointers updated (README, etnos/10_design-language). SESSION_LOG history keeps the old name deliberately.
  • Working-language call (Yose): internal specs and brainstorm docs in ENGLISH from now; he does the Indonesian pass later. User-facing surfaces stay Indonesian-first (D13 unchanged).
  • Blueprint embed re-verified live (/blueprint-perangkat 308s to trailing slash; followed page contains the iframe; the blueprint HTML itself serves).

Next: Yose reacts to the six shaped questions in device-spec §9 (PoE and battery SKU are the interesting ones); D20 naming pass; MOSS Space probe (LEDGER 15); banked: Kertas Dinas theming for the vocs site.

2026-07-10 (late night 4) — aksara-mem PUSHED + E2E AUDIO LANDSCAPE: stay modular

  • Repo family settled (D20 partial): Yose picked aksara-mem (family: aksara-cli / aksara-mem / aksara-ui; bare aksara reserved for the umbrella). Scaffold pushed to github.com/yohn-maistre/aksara-mem (private, 13 files). Proposed but NOT executed: rename aksara-papan to aksara-ui (recommended over aksara-e-ink because D19 leaves display tech open, RLCD endgame; GitHub renames auto-redirect). Awaiting Yose's word.
  • research/2026-07-10_e2e-audio-landscape.md (22 KB): Yose's find = LFM2.5-Audio-1.5B (Liquid AI, genuinely e2e, llama.cpp GGUFs, fits the board) but English-only in and out; a JP variant proves Liquid ships per-language checkpoints, making an ID checkpoint the #1 watch-list trigger (Language Flywheel angle, same shape as the MOSS fine-tune play). No open-weights model speaks Indonesian as of today; speech-out is a zh/en duopoly; the only Indonesian speech-IN in open weights (Qwen3-Omni 30B-A3B) does not fit 16 GB; the only Indonesian speech-OUT claim anywhere is proprietary API-only Qwen3.5-Omni (membership UNVERIFIED) which violates the sovereignty floor for citizen voice. Full duplex = Moshi lineage, English, and wrong target for a kiosk (barge-in via VAD + XVF3800 AEC suffices). Verdict: STAY MODULAR, and not as consolation: receipts need verbatim transcripts (e2e models answer, they do not transcribe) and consent needs speaker embeddings (e2e models do not diarize), so an e2e model would ADD an engine, not delete the pipeline. LFM license nuance flagged for procurement: LFM Open License v1.0 is Apache-derived but gates commercial rights at USD 10M revenue, not pure Apache. Five re-check triggers logged.
  • Modular pipeline reframe for the spec narrative: the "ducktape" is load-bearing; each seam (ASR transcript, speaker embedding, text LLM, TTS) is exactly where receipts, consent, and auditability attach.

Next: hardware wave (device 00/01 STUB→DRAFT from blueprint v0.4); Yose: aksara-ui rename call, D20 naming pass, MOSS Space probe (LEDGER 15).

2026-07-10 (late night 3) — .aksara SPEC v0.1 DRAFT + BLUEPRINT ON THE SITE + Understand-Anything read

  • specs/memori-aksara/02_format.md (DRAFT v0.1) shipped and LIVE on abstraksi.pages.dev: six principles, the dotfolder layout, day-file grammar (append-only, koreksi-atas), wiki tiers with sumber receipts, sidecar contract (derived-only, table sketch), Buku Tamu normative rules (§6, the moat), tidur lifecycle, conformance L0 Vault / L1 Indeks / L2 Hidup. D20 naming still open.
  • Blueprint now displayed on the site: generated /blueprint-perangkat page iframes the v0.4 HTML blueprint whole. Two lessons paid for: vocs's dead-link checker rejects markdown links to public/ assets (fix: HTML anchors), and CF Pages soft-404s return 200 with a shell page, so verify content, never just status codes (the first "200" on the spec page was a lie; iframe grep exposed it).
  • Understand-Anything (Egonex-AI, 72.5k stars, MIT) read on Yose's ask: a Claude Code plugin (also Codex/Cursor/Copilot adapters) = markdown agent definitions (project-scanner, file-analyzer, domain-analyzer, graph-reviewer, tour-builder...) + a deterministic TS core (parsers, graph-builder, layer-detector, fingerprint staleness, embedding-search) + LLM only for summaries/implicit relations, emitting an interactive dashboard. Their split IS our "structure remembers, models dream" in a code-costume. Decisive detail: /understand-knowledge parses Karpathy-pattern wikis (wikilinks + index.md), which is exactly the semantik drawer, so any .aksara vault is UA-viewable for free; noted in the new repo README as ecosystem evidence.
  • ~/workspace/code/memori-aksara/ scaffolded and committed locally (README with pitch + conformance + UA note, Apache-2.0, contoh vault with tamu EMPTY by principle). Remote creation awaits Yose's explicit name-confirm (auto-mode gate; one command to push).

Next: Yose confirms repo name → push; hardware wave (device 00/01 STUB→DRAFT from blueprint v0.4 schedules); D20 naming pass.

2026-07-10 (late night 2) — TWO SCOUTS LANDED: voice/model stack + memory substrate; §8 verdicts folded

Lane B blessed by Yose; two parallel scouts answered his round-2 questions. Both reports sentinel-verified, committed, live on the site; 01_arsitektur.md §8 updated from TBV to verdicts.

  • research/2026-07-10_voice-and-model-stack.md (22 KB): MOSS-Transcribe-Diarize 0.9B is REAL (2026-07-09, Apache 2.0) but no Indonesian (zh/en/ja) and CUDA-only, so not the pick; highest-ceiling watch-list item (0.9B + Apache = Indonesian/Papuan-Malay fine-tune within reach). The decisive finding: inline speaker tags are anonymous per-recording; consent needs persistent identity across sessions, so the embedding pipeline is the architecturally CORRECT stack, not the fallback (sherpa-onnx: Whisper small int8 + pyannote seg-3.0 ONNX + enrolled-voiceprint match; all Apache/MIT, RK3588 builds shipping, <3 GB). Gemma 4 E4B stays; Qwen3.5-4B = constrained-JSON workhorse candidate; Cohere North Mini Code (30B MoE) real but does not fit 16 GB. Cloud econ: office-day workload = Rp 16-66rb/MONTH; local floor is about sovereignty + availability + USD exposure, NOT savings. "Whisper v4" does not exist. Yose probe queued (LEDGER 15): Indonesian clip into the MOSS HF Space.
  • research/2026-07-10_memory-substrate-team-agents.md (30 KB): LadybugDB much healthier (v0.18.1 shipped today, aarch64 wheels verified, MIT, FTS+vector) but bus factor ONE (77/100 recent commits) + no storage-format guarantee: v1 sidecar stays SQLite, D17 stands, fork rejected with numbers; lane B makes later re-platforming a re-index, not a migration (2027-01 checkpoint now has 3 concrete criteria). Team-agent scan: industry converged on our shape (scope = consent boundary, mechanical isolation, editable memory, audit logs); "town" = a16z/Forerunner $55M PERSONAL assistant (category evidence, lane still empty); GateMem + MemClaw (June 2026 papers) validate mechanical write-gates. Compaction: text-only office DECADE = 150-600 MB compressed; never delete raw day-files (provenance layer), quarterly zstd, digests for recall. Offload hypothesis supported with caveats: mechanical recall, small models only for single-purpose extraction on small windows, at-or-below-7B fails reasoning over piles (recall feeds digests+facts), constrain only the final emit; no published ~4B memory-bench numbers exist.

Next: Yose closes the D20 bundle (01_arsitektur §7, evidence complete) + runs LEDGER 15 probe; then 00_format.md DRAFT (the .aksara spec). Voice bake-off (whisper small vs medium RTF on RK3588S) waits on hardware.

2026-07-10 (late night) — HUB LIVE AT ABSTRAKSI.PAGES.DEV + ARCHITECTURE WEIGHING DOC

Yose added the 2 CF secrets; workflow_dispatch rerun deployed green; verified live: homepage 200 ("Abstraksi Hub"), deep research route 200, /llms.txt 200, the engineering blueprint serves at /blueprints/aksara-node-blueprint. Public-for-now by his explicit call (phone reading), so _headers now ships X-Robots-Tag: noindex; hardening options (CF Access / teardown) parked in LEDGER 9b.

Brainstorm round 2 on the brain (his satu-berkas SQLite lean + Karpathy wiki + Honcho + psychology framing) is serialized in specs/memori-aksara/01_arsitektur.md: the truth-vs-artifact tension named; enam laci taxonomy (episodik/semantik/prosedural/kerja/prospektif/tamu); three architectures weighed (A satu-berkas SQLite, B vault + single sidecar db, C server brain) with B recommended and C reframed as a later hub-tier mirror; Honcho inverted into Buku Tamu (purpose-tagged explicit facts, no psychological profiling: "Aksara mengingat urusanmu, bukan kepribadianmu"); tidur consolidation job (distill with receipts, decay, merge, contradiction queue for human ratification, prospektif sweep). Open D20 bundle (5 items) waits on Yose.

2026-07-10 (night) — DOCS SITE GREEN + MEMORY-BRAIN LANDSCAPE + MEMORI AKSARA SEED

Three artifacts from Yose's "company brain" dump (his framing: CF Pages docs hub now, our own communal memory standard after, precursor to Aksara):

  • Docs site (vocs) builds GREEN in CI, deploy gated on secrets. site/ holds a scaffold where hub markdown stays canonical and sync.mjs derives everything (pages, sidebar with real titles, MDX-hazard sanitization of raw </{ outside code, .md links become routes, blueprint + GAMBAR pack as static assets, llms.txt + llms-full.txt as the agent surface). Nothing was installed on the phone: bun ran the sync test locally, GitHub Actions does the real build. Three CI iterations to green, all instructive: (1) vocs 2.5.3 is a waku SSR app with NO Cloudflare deploy target (sitemap has only netlify/node/vercel) so we pinned to vocs 1.0.16 (static Vite build, the viem.sh lineage); (2) 1.x needs typescript for twoslash; (3) @vanilla-extract/vite-plugin 5.2.4 (published 4 days ago) breaks vocs's precompiled css, pinned to 5.0.1 via npm overrides. Final run: "prerendered pages" confirmed, docs/dist carries all routes. Go-live = LEDGER 9b (2 CF secrets + the public-vs-gated call, Yose).
  • research/2026-07-10_memory-brain-landscape.md (38 KB, 22 systems, sentinel-verified): the category is crowded at the individual center, EMPTY at our edge. Nobody ships open, file-first, bi-temporal, consent-aware memory where the agent IS an institution; communal/governed memory exists only as papers (Accenture Collaborative Memory; Governed Collaborative Memory, whose human-ratified promotion into institutional state is structurally a musyawarah). "Holographic" = Hermes's HRR SQLite provider (D17's template, best embedded existence proof). Closest .aksara neighbors: Basic Memory (AGPL, no temporal ledger), OpenHuman (GPL-3, personal-only), agentmemory (Apache-2.0, consent-free). Verdict: decouple .aksara/ as a file-format standard (CLAUDE.md-style convention any agent can read), aksara-cli as the only consent-gated writer; proposes a D17 refinement (markdown = truth, SQLite = rebuildable index) that needs Yose's ratification, LEDGER 14.
  • specs/memori-aksara/00_seed.md (BRAINSTORM): Yose's dump serialized; deltas vs already-settled D17/D18 isolated (dotfolder packaging, access-recency weighting, self-cleaning lifecycle, embeddable distribution, hub-as-first-vault); the thin-wrapper worry answered honestly (generic memory IS thin; consent + bi-temporal receipts + offline + communal semantics is where it stops being thin, and the research confirmed that intersection is empty).

Next: Yose reads the research verdict + makes the 5 queued decisions (LEDGER 14) and the docs go-live call (LEDGER 9b); then promote seed → specs/memori-aksara/00_format.md DRAFT (the .aksara spec) citing the landscape. Kertas Dinas theming for the site = banked polish round.

2026-07-10 (later) — HUB PUSHED: github.com/yohn-maistre/abstraksi (private)

The whole hub is now on GitHub as a private repo, branch main (renamed from master). One rewrite survives in the published history: every commit was normalized to the avatar-linked identity (yohn-maistre <33504511+...@users.noreply.github.com>; several early commits were authored with the no-avatar gmail). Tree verified byte-identical through the rewrite; 81 files remote = local; secrets-pattern scan clean before the first push. A brief experiment with vague commit messages ("update", "wip") was pushed and then reverted the same hour when Yose realized the descriptive history was worth keeping; messages are and stay DESCRIPTIVE, this log stays the narrative. The repo stays private until Yose flips it.

2026-07-10 — DEVICE FORM PIVOT 2: blueprint v0.4 "Plakat" (floating laminated sheet, prism deleted)

Yose brainstormed the packaging further: spine that "blends into any wall", screen closer to the wall, "laminated with polycarbonate sheets with bolts on the corners". That instinct is the culturally-loaded form: standoff-bolted acrylic signage is the visual vernacular of Indonesian public offices (maklumat pelayanan, struktur organisasi) — so v0.4 makes the device read as the official notice that was always on the wall, except it listens and answers. Three variants were shaped: A "Plakat" (everything hidden behind a floating sandwich — recommended + built), B "Ambang" (slim lintel strip above the slab as a face-to-talk-to; recorded fallback if field tests show people don't know where to speak), C "Bingkai" (TRMNL-like single frame ~28 mm; conservative fallback). Both alternates are recorded in the blueprint's engineering notes.

v0.4 geometry (all in blueprint_gen.py, regenerated + committed): flexible 0.647 mm e-ink lies FLAT (bow deleted) in a cold-assembled sandwich (PC anti-glare 2.0 / panel / PC 3.0 ≈ 5.6 mm — NOT thermal lamination, e-ink is heat-sensitive; OCA cold lamination at production). Slab 200 × 252.5 (21.4 mm paper margin) floats 30 mm off the wall on 4 signage standoffs (Ø16×30, toko akrilik lokal); face 35.6 mm off the wall vs v0.3's 57. The header prism is deleted, not shrunk: all electronics moved to a hidden wall-side cassette 160×140×22 in the standoff gap (SBC flat, IT8951 right behind the FPC slot — one static S-fold at assembly). The gap does all the work: chimney convection cools the aluminum cassette, speaker fires backward using the wall as reflector, mics port through the slab's top edge (gap-echo = the main acoustic TBV), mmWave sees through the PC-only bottom margin, and the halo becomes a CORONA — LED perimeter washing the wall around the floating sheet (light never touches the e-ink; visible across the room). Signature move: the bottom-right bolt is the fingerprint sensor — "cap jempol" (thumbprint-as-signature gesture; terracotta ring, module still TBD/Yose). Copied to /mnt/internal-storage/aksara-node-blueprint-v0.4.html (v0.2/v0.3 kept beside it).

Next (unchanged + one new): canonical specs/kertas-dinas/00_language.md; fold Plakat form into device/05_design-language.md; promote device/01_bom.md from the parts schedule. Yose-hands: react to v0.4 by eye (margin width? bolt finish? A vs B?); pick flush-fp/bolt-head module; Waveshare-direct price spot-check still open.

2026-07-07 (night) — DEVICE FORM PIVOT: blueprint v0.3 "Prisma" + flexible-display research + design language named

Yose loved the v0.2 blueprint and pivoted the FORM: "thin paper coming down." New shape (v0.3, shipped): a solid rectangular-prism HEADER on top carries all mass, ports, and the halo; a flexible 0.647 mm e-ink sheet hangs below it in a fixed shallow forward bow, portrait. No front grills (TRMNL-clean) — mics vent on the top face, speaker on the end-caps (that's what the 11.4 mm overhang is for), mmWave behind the front face; fingerprint is flush/under-surface (drops the protruding R503). Battery downsized to 40–60 Wh flat so the header stays ~45 mm deep (SBC stands vertical on the backplate); 153 Wh would need a ~65 mm header (runtime vs slimness, field call). Files: specs/device/blueprints/blueprint_gen.py rewritten, aksara-node-blueprint.html regenerated; both v0.2 and v0.3 copied to /mnt/internal-storage/ for before/after.

Backed by a research scout: research/2026-07-07_flexible-paper-displays.md (21 KB, house-style, sourced). Headline: a ~10" flexible e-ink IS buyable in 2026 at the same price as the rigid panel (Waveshare "(D)", 0.647 mm, ~$200–214) — swapping rigid→flexible is ~free (+Rp 0–180k), panel line stays Rp 4–4.5jt. BUT it is flexible-durable, not rollable: one fixed gentle curve, no repeated flexing/roll (no vendor publishes a roll-cycle rating — that absence is the finding). So the "pull down like a projector screen" idea is concept-only in 2026 (LG rollable discontinued, Royole bankrupt, LG Slidable = CES concept); the drape is a permanent bow, not a motorized pull. Flexible OLED exists (0.34 mm) but emissive → burn-in → wrong for an always-on notice board.

Design language named: "Kertas Dinas" (Yose blessed it) — official paper reclaimed as beautiful and trustworthy; the light face is Honai Siang, palette DINAS cream + terracotta. In design-world terms it's a cross of Swiss/International Typographic Style + New Brutalism + editorial/broadsheet + Tufte information design + technical-drawing vernacular, civic/govtech register ("editorial brutalism" is the one-label shorthand).

Next (teed up, awaiting Yose's look at v0.3): write specs/kertas-dinas/00_language.md (canonical, so detak/ETNOS/device cite one source); fold the flexible-panel + Prisma form into device/05_design-language.md; promote device/01_bom.md STUB→DRAFT from the blueprint's parts schedule. Yose-hands: phone spot-check the Waveshare-direct flexible-HAT price (403-gated, UNVERIFIED), pick the flush fingerprint module.

2026-07-07 (evening) — ABSTRAKSI WAVE: engineering blueprint + proposal manifest + hub polish

ETNOS design overhaul is done (below) and handed to a separate polishing agent; this session pivoted back to the abstraksi hub. Three artifacts landed (commits cb05e2c, 3b2c4e0, + this hub polish):

  • Device engineering blueprint v0.2 — specs/device/blueprints/aksara-node-blueprint.html, a self-contained HTML page (opens on a phone, prints A3 landscape) generated by blueprint_gen.py (pure-stdlib, regenerable, on the proposal/gambar/generate.py pattern). Contents: title block (ADK-DEV-BP-002), three dimensioned SVG views (tampak depan, potongan samping A-A tilt 4°, tata letak belakang), a rough shaded render (halo glow + e-ink content), a BOM-linked parts schedule (14 balloons → real SKUs/prices from research §F, honest TBV tags), a dimension schedule, engineering notes, revision history. Supersedes the v0.1 static SVG; footprint 297.7×181.2 (panel-module-driven; the design-language ≈340×210 estimate is retired). Copied to /mnt/internal-storage/aksara-node-blueprint-v0.2.html for Yose. No renderer on the host (didn't install) — geometry verified mathematically + artifact-grepped clean.
  • Proposal package manifest — proposal/README.md: the Nabire deliverable is DRAFTED end-to-end (surat ×2, exec summary v0.2, concept note v0.2, profil v0.1, GAMBAR 11/11, errata); manifest gives saji order, the exact [placeholder] fill-checklist (NIB/NPWP/Kemenkumham/kop/email/telepon), two coherence notes for Yose's eye (PIDI claim present in concept note but dropped from exec summary; verify the "8 Agustus 2025" event date), and the production path. Proposal is Yose-hands-only now (signature, legal fields, phone-verify APBD window).
  • Spec hub polish — README map de-staled against ground truth: DECISIONS D01→D19, research 7 reports (not 3), etnos 10 DRAFT + horizon, device 05 DRAFT + blueprint v0.2 + 06 PLAN, ops 03 legalitas LIVE, GLOSSARY FROZEN, proposal row points at the new manifest.

Next: Yose's hands on the proposal (legal fields + signature + APBD phone-check); pick which out-of-the-box gap to build next (briefed to him this session). Device specs 00/01 (device-spec, BOM) are the natural STUB→DRAFT promotions now that the blueprint pins the numbers.

2026-07-07 (later) — DESIGN OVERHAUL COMPLETE: P0–P8 ALL SHIPPED

Same day, second push: the full plan is executed. On top of P0–P5b below, four more CI-green waves landed on etnos.pages.dev:

  • P5c Papan Kilas split-flap (Solari placard on /explore: BERITA/FORUM/KOMUNITAS/KATA rows, per-row CSS half-fold flip, day-clock seeded via detak seed.ts port — same board for every reader; shared kilas-data.ts + hot.ts so ticker/board/scoreboard share one contoh owner and ONE backend fetch per page)
  • P6 wiki overhaul + bahasa merge (BahasaHub grammar cards under /wiki/bahasa with bahasa.md intro, /bahasa → 308, mono chip categories, "Wiki Tanah Papua", carousel grammar pass with TTS untouched)
  • P7 civic surfaces (CapabilityCard extracted; /org + /org/[slug] Level-0 contoh pages with naik-kelas ladder; /musrenbang usulan composer — 5-stage tracker honest about stopping at stage 1, PROTOKOL DEMONSTRASI stamp, hands off to the existing /create/post crosspost param, no parallel submit path; tentang rewritten to island framing with civic links)
  • P8 cleanup (capabilities.ts piefed gates hide modlog links; /signup replaced with honest register-at-piefed.social page, /auth/register verified 200; i18n en/id/pmy etnos.* flatten-diff = zero missing; CODEBASE_GROUNDTRUTH.md committed)

Beta gate now = Yose's screenshot review. Deferred by design: dark theme (Honai Malam), P3 leftovers (filter chips, cluster-card deep restyle, maplibre bundle split), repo-root implementation_plan.md/walkthrough-1.md await his keep-or-delete word. Full state: honai-metro/.claude/DESIGN_OVERHAUL.md EXECUTION STATE.

2026-07-07 — DESIGN OVERHAUL P0–P5b SHIPPED, deployed, live-verified

Six waves on etnos.pages.dev, all CI-green and proof-grepped against live assets (full ledger in honai-metro/.claude/DESIGN_OVERHAUL.md EXECUTION STATE — that block is the truth, updated every wave):

  • P0 nav credibility (5-item sidebar = hamburger, drift killed, i18n ×3, PUBLIC_DEFAULT_FEED=Local)
  • P1 Honai Siang theme (DINAS cream + terracotta light ramps, dark untouched, 'Honai Metro (Lama)' rollback preset, FOUC-killing static @theme mirror)
  • P2 type + grammar (Archivo display site-wide, Geist Mono furniture, Instrument Serif captions, 87KB fonts; eyebrow/inkbar/rules/chip/stamp/ledger/ghost-num + paper grain + reduced-motion block)
  • P3a PetaPapua — detak's act-1 DINAS atlas plate ported verbatim onto /explore (OFM vector, engraved skin, koordinat strip, ODbL receipt), island framing (six kota anchors, no province-picker), guest / → /explore
  • P4 dashboard live tiles (live.ts SWR: user_count + paged communities "200+" + 24h posts "50+"; LANGSUNG chip + count-up; "Papan Data Tanah Papua"; Sumber ledger — satu angka satu sumber)
  • P5a/b news trio 2/3: KILAS ticker (detak port, CONTOH→LANGSUNG flips when PUBLIC_DETAK_URL /ticker exists) on / + /explore; Papan Sorotan scoreboard (live Hot/Local ledger rows) on /explore

Next: Yose screenshot-review of the cream theme + new explore (rollback = Lama preset); then P5c split-flap Papan Kilas, P6 wiki+bahasa merge, P7 /musrenbang + /org, P8 cleanup/audit. Needs-from-Yose: detak worker URL for PUBLIC_DETAK_URL (ticker goes langsung), screenshot pass, decision on repo-root implementation_plan.md/walkthrough-1.md before P8 deletes.

2026-07-06 — DESIGN OVERHAUL WAVE opened (pre-beta, plan of record committed)

Yose's brief after reviewing 11 live screenshots: award-aiming redesign before PUBLIC BETA. Full stress-tested plan lives at honai-metro/.claude/DESIGN_OVERHAUL.md (phases P-1..P8 + risk register + amendments); design language spec at specs/etnos/10_design-language.md. Headlines: light theme → DINAS cream + terracotta ("Honai Siang"; dark deferred by amendment), site-wide type trio (Archivo/Geist Mono/Instrument Serif), grammar layer (eyebrow/inkbar/rule/chip/stamp/ledger + grain), IA consolidation (5-item sidebar, bahasa→wiki hub, hamburger drift fixes, guest home = curated Jelajah), act-1 atlas map port from detak-detik replacing tile map, news trio (KILAS marquee = detak edisi feed w/ contoh fallback; Papan Kilas SPLIT-FLAP hero board; Sorotan post scoreboard), live dashboard tiles via working PieFed reads (getSite user_count, paged listCommunities, getPosts 24h — CONFIRMED feasible, no adapter changes), /musrenbang composer + /org/[slug] Level-0 pages, dead-surface cleanup. Key stress-test findings banked in the repo doc: ~50 dead primary-* classes go live with the ramp, latent chart rgb(rgb()) bug, other.white IS the card color, FOUC needs static @theme mirror. New Digg cautionary tale noted (beta died 2026-03 of bot flood; our attested-agent chrome is the answer). Execution began same session: P-1 done, P0/P1 in progress.

2026-07-05 (later) — PROPOSAL WAVE: Nabire package drafted end to end

Yose's directive post-compact: images first, then rewrite for a public audience (not slavish v4.1 derivation), forum after.

  • GAMBAR pack DONE (11/11): proposal/gambar/gambar-01..11.svg + index.html (open in browser, print to PDF landscape = lampiran). Generated by proposal/gambar/generate.py, zero-dependency Python SVG (host has no matplotlib/seaborn/D2; spec tooling substituted at the same design tokens #F7F4EF/#1A1A1A/#C0633E). All geometry machine-validated in-bounds. GAMBAR 7 reuses the device blueprint; fixing it uncovered a raw & XML bug + bottom-clipped canvas in specs/device/blueprints/aksara-node-v0.1.svg, both fixed at the source.
  • Public rewrite: proposal/01_exec-summary.md DRAF v0.2, written standalone for Bapperida/Diskominfo/Setda readers (not a pangkasan): masalah with sourced numbers, "staf administrasi yang tidak pernah pindah tugas" as the public phrasing (PNS Digital avoided in official docs), Skenario Rancangan 1 narrated with the refusal path, kejujuran status, twin ask. Concept note bumped v0.2 (GAMBAR refs).
  • Profil perusahaan: proposal/drafts/2026-07_profil-perusahaan.md DRAF v0.1, 6-8 hal structure, founder experience as tenaga-ahli capacity per ops/03, legal fields as [placeholders] for Yose.
  • Corrections banked: proposal/03_v41-errata.md, 12-row errata table for the docx v4.2 edit (funding table, PIDI, Perpres 83/2025, Oracle, WA ban, APBD clock, OpenViking/Kuzu, Photon PieFed, display, PNS Digital wording, Otsus pagu). Public docs already carry these.
  • LEDGER 3c updated: the whole package now waits on Yose's review/signature/placeholder-fill.

FORUM WAVE (same session, first pass) — 3 of 7 etnos specs expanded STUB→DRAFT:

  • 04_agents-in-forum.md: the agentic substrate. civic:InstitutionalAgent actor (bound to communal entity, created by attestation ceremony, mandatory AGEN chrome, persona-scoped); capability card (8 fields incl. live heartbeat + SLA pointer); registry at /agen doubling as skill-pack distribution; trust tiers T0-T3 (hardware-attested, earned only); signed provenance footer on every agent post (approval_ref + record_ref, offline-verifiable, middleware-enforced); conduct rules (humans always post as themselves, resolves stub question); MCP bridge (forum lists, never proxies).
  • 00_architecture-seam.md: honest current state (Photon fork base v2.3.0 = native PieFed already included, no rebase cliff; borrowed piefed.social backend NEVER carries civic verbs), EtnosClient seam verb set, PieFed gap table (signup/modlog 'unsupported'), one-instance topology with pre-written split doctrine.
  • 01_fork-delta.md: full route/feature inventory grounded in git history; graduation order decided: /agen registry → transparency dashboard → wiki/corpus; Karya-hybrid corpus page spec'd inside /bahasa (paid dignified crowdwork, CARE consent tiers); merge cadence + upstream contribution candidates.

FORUM WAVE (fourth pass: social core + first implementation) — Yose's corrective brief: ETNOS is a FULL reddit-like social forum (subs for interests/tribes/cities/topics/memes) with Musrenbang + communities at core; the reddit paradigm's healthy-discourse benefits are doctrine. New 09_social-core.md DRAFT answers his four questions: (1) individuals vs orgs = same account machinery, different actor TYPE with different chrome/capabilities (5-row actor spectrum: Warga stays human-simple with no tiers; Operator Terverifikasi; Komunitas/Lembaga Level 0-2; Agen Institusi; NEW Agen Layanan); (2) AI agents in threads = utility agents (penerjemah/perangkum/arsiparis/pengingat), instance-operated, BOT chrome, INVITED never ambient, zero persuasion; (3) ads = one transparent @iklan actor, context-targeting by community only (no behavioral profiles exist structurally), public ad ledger (receipts doctrine on our own revenue), UMKM-first pricing, never on civic surfaces/political; (4) hosting = staged: borrowed backend → small Indonesian VPS at beta → THE MOVE: host inside Papua Tengah Diskominfo's own infra via PKS ("forum publik Tanah Papua pertama, dari Papua Tengah", their rack + our platform, Rp 0 to both). Proposal inclusion pre-shaped 3 options, recommended (a): one paragraph in the Diskominfo letter variant + leave-behind one-pager. Consolidations logged in 09 §6 (Level-0 page = capability card = UMKM listing, one component; ad ledger = c/transparansi rendering pattern; Musrenbang Level-0 + simulasi + /agen demo = one scenario set). Implementation in honai-metro (first horizon item shipped): /agen directory placeholder → simulated capability-card registry per spec 04 §2 (agents.json fixture: 3 agent nodes T1/T2 + 1 Level-0 org showing the spectrum; status dots aktif/terganggu/luring/kontak-wa; SIMULASI badge + honest note; i18n en+id). Termux OOM-crashed mid-build (128Mi free RAM); edits survived. Post-crash: build succeeded with NODE_OPTIONS=--max-old-space-size=1536 solo (proof strings grep-verified in dist), phantom dep @floating-ui/dom declared (bun hoists flat, pnpm strict layout exposed it), honai-metro pushed (26073747..7ef9b4d7). CI now green through build; deploy stops ONLY at the known 2-secret blocker (CLOUDFLARE_API_TOKEN/ACCOUNT_ID = Yose's hands, keys-last rule). Repo bun-canonical: pnpm-lock gitignored, wrangler-action pinned packageManager: bun. Optional ETNOS-hosting paragraph added to Diskominfo letter variant (6a8ae12). LAUNCH: Yose added the 2 CF secrets same evening → rerun 28758224706 ALL GREEN → ETNOS alpha LIVE at etnos.pages.dev (/ and /agen verified 200, PWA shell, lang=id). Git authorship fixed globally: yohn-maistre <33504511+yohn-maistre@users.noreply.github.com> (avatar-linked noreply form; old yohn.maistre@gmail.com rendered no avatar); CLAUDE.md updated.

FORUM WAVE (third pass: Yose ratified + round 2 brainstorm) — Yose said YES TO ALL of horizon round 1 and gave the canonical design stance, now written into spec 04: agents represent processes, not people (org = blackbox with typed doors; agent is part of the org's process). His one hesitation: offline-first technical complexity; answered in 07_horizon with the decomposition (forum is NOT offline-first, the NODE is; single-writer append-only streams = no CRDT dragons; PWA offline-read is free polish; DTN is phase 3 and never blocks the pilot). Round 2 added to 07_horizon (zero-cost/high-leverage): D1 Level-0 presence without device (THE growth bottleneck-killer: all 1.400 kampung on the forum year one, agents only at pilot sites), D2 civic DNS static pages, D3 WA share-card amplop pattern, D4 simulasi mode (demo+training+playground), E1 reason codes as open standard ("HTTP status codes of bureaucracy"), E2 kode wilayah reference dataset, E3 template library in skill-pack format, E4 specs on wiki, F1-F5 agent-native mechanics (jam piket, tanya berdiri with diff-feeds, referral threads, pipeline view, sumbu waktu kampung), G1 Common Voice embedding, G2 kamus birokrasi. A1 PROMOTED → 08_musrenbang.md DRAFT v0.1: usulan lifecycle with the answer coming back down (closing the loop), five-stage mapping resolved (one transaction per level chained by TransferRecord; stretch is temporal not structural), officials' sales page, honest boundaries (records musyawarah, never simulates consensus), Level-0 build shape that runs on current stack = Nabire show-piece. My round-2 promotion ballot: D1 + D4 immediately, E1 when aksara-cli/02 firms.

FORUM WAVE (second pass: complete) — remaining 4 specs expanded, all 7 now DRAFT:

  • 03_civic-namespace.md is now the working JSON-LD context itself: 4 actor types, 10 activities (incl. reserved SubmitUsulan), 5 objects, properties, Indonesian-first labels, records-outlive-vocabularies versioning rule, self-contained-with-mappings stance (resolves alignment question). Domain still TBD (LEDGER 12).
  • 02_federation-extensions.md: two-layer signing (transport vs attestation envelope that survives USB store-and-forward; invalid provenance quarantines visibly, never silent-drops), permission envelopes with the public sanction ladder, two planes, DTN signed-bundle profile with the kurir digital role, Nabire→Timika worked flow, honest limitations kept.
  • 05_release-gates.md: alpha/beta/post-launch checklist tables with honest ❌ states; per-feature board (what turns each label green). Gate owner: Yose.
  • 06_transparency-and-index.md: PII-free-by-construction emission schema with merkle digest roots + small-n suppression (k=5), citizen verify-my-document flow, index methodology non-goals (not a ranking), panel-and-dashboard render from the same records so they cannot disagree, political-safety review gate before any cross-district page.
  • 07_horizon.md (NEW, brainstorm): 21 futures in 3 tiers + anti-features. Tier A rides current stack (A1 Musrenbang-native process = nominated first promotion; rejection corpus as curriculum; kurir digital celebrated role; stempel culture; detak-detik flywheel). Tier B needs pilot (operator guilds, living dictionaries, oral history vaults, UMKM registry, Karya work board, cross-language bridging, Indera streams, musyawarah tooling). Tier C ten-year (adat decision records, constitutional dogfooding, radio bridge, disaster mesh, panel-as-reader, koperasi sunlight, Melanesian arc, neolab threads). Anti-features named: no engagement feeds, no ads on civic surfaces, no human reputation scores, no unlabeled AI content. Awaiting Yose reactions → promotion to numbered specs.

2026-07-05 — The founding session (framing to first pixels)

One session took the Aksara vision from a docx + floating ideas to a structured hub, settled decisions, verified feasibility, drafted government letters, a designed device, and a running UI repo. Everything below is committed in this repo (git log) and in aksara-papan.

State at end of session:
  • Hub built: 00_FRAMING (the why; §8b same-day corrections), README (map), DECISIONS D01-D19, LEDGER (needs-from-Yose, time-critical items on top), GLOSSARY (frozen), 27+ spec stubs, SESSION_LOG (this file).
  • Decisions all settled except: D17 (Memori Aksara, proposed), D19 (display, proposed), kamera saksi (design doc, proposed). Everything else Yose-confirmed, including D09 license split (ETNOS AGPL / cli+verify Apache-2.0 + trademark / specs CC BY 4.0 / hardware CERN-OHL-S) and D10 naming (all Aksara-branded, ZeroClaw+ZeroStamp retired, ATECC608B).
  • Research (6 sourced reports in research/): context distillation of Yose's 19 prior docs; APBD/Otsus window; stack+hardware SOTA (18 deviations from v4.1 incl. Oracle free tier cut to 2/12, Meta WA bot ban, Photon native PieFed support); procurement feasibility (VERDICT: yes, staged; PT Perorangan is UMK advantage; first rupiah Q4 2026-Q2 2027; full package APBD-P 2027/2028); Alibaba CoCreate (SKIP; Indigo Telkom better); memory stack (build Memori Aksara: bi-temporal SQLite, ~0.5 GB; OpenViking = markdown data model but AGPL server runtime); display options (no-flash <USD 80 not purchasable; engineer it on current panel; production GDEP103TC2+TCON).
  • Government track: concept note v0.1 + surat audiensi drafts in proposal/drafts/ (Governor's language: berbasis data dan hasil, 6 program prioritas, SADAR hook); legalitas runbook ops/03 (KBLI 62010/62015/62090, LPSE/SIKaP, deadlines 15-20 July); strategy proposal/00. HARD DEADLINE: fungsi into PPAS 2027 by Aug-Sep 2026; August Nabire audiensi is the target event; BRIN RIIM closes 31 July (needs campus partner).
  • Design: specs/device/05_design-language.md v0.4 (perapian bukan wajah; papan+punggung; LED=heartbeat e-ink=mind; sound hierarchy; kamera saksi; identity substrate incl. voice ID; access matrix; no touch on wall, touch OK on MPP kiosk variant; NIM only at Deka tier). Blueprint v0.1 SVG in specs/device/blueprints/ (297.7x181.2 mm, TBV-marked). UI sim build plan specs/device/06.
  • Code: github.com/yohn-maistre/aksara-papan (PRIVATE until Yose flips): Svelte panel UI v0.1 with e-ink sim (16-gray discrete filter, GC16 ritual flash), Papan idle screen, statline, stempel suasana, device chrome + halo. Run: pnpm dev --host.
Next waves (order decided): 1) PROPOSAL (time-critical: GAMBAR pack 11 visuals per proposal/02 plan, assemble Nabire package around existing drafts, refresh v4.1 corrections D14/D18 + funding table). 2) FORUM (expand specs/etnos/ stubs: agentic substrate, capability registry, Karya-hybrid corpus page; Photon v2.2+ rebase). Forum is recoverable from code+specs; proposal carries the deadline.

Standing constraints: phone host (cap ~5 subagents, no installs without asking), no em dashes in Yose-voiced docs, honest labels everywhere, every number carries its source.