ABSTRAKSI — Framing Doc

Date of analysis: 5 July 2026 · Status: living framing doc — the "why and what" that every spec doc under abstraksi/ descends from · Source: partner analysis of Aksara_Master_Proposal_v4_1.docx + brainstorm session with Yose · Register: working English, Indonesian domain terms kept as-is

This document consolidates the fresh-eyes read of the Aksara/ETNOS vision, the positions we settled, the expansions we added, and the tests every future feature must pass. Derived docs (specs, the Papua Tengah proposal, blog posts) must not contradict this file; when they do, one of the two gets fixed the same day.

1. The verdict — where the value actually lives

The crown jewel is aksara-cli: the separation of reasoning from execution (proposal §12). A language model that cannot emit a formal document except through a deterministic tool layer — official template, validation gates, cryptographic primitives, signed audit emission — is what turns "AI produces legally-valid government documents" from a lawsuit into an architecture. It is simultaneously the answer to hallucination and to prompt injection (compromising the reasoning layer confers no signing capability). Everything else — hardware, forum, WhatsApp, voice — is a delivery vehicle for this layer.

The second jewel is Communal AI as a category. The individual→communal primitive table (§6) and the communal-entity ontology — memory, identity, consent, rhythm, each implemented by a separate subsystem (§7) — are publishable as-is. Nobody has implemented CARE Principles at protocol level. These two ideas are the moat's foundation: publish them, with Yose's name on them, before anyone else names this space.

The strategic wedge is Srikandi computer-use (§31). Every instansi is mandated to use Srikandi; it is web-only (no API partnership needed); its pain is publicly documented; blank-spot offices literally travel to find signal to upload naskah. "We queue it offline and submit it for you, with a signed audit record per session" is a one-breath value proposition for a Sekda, and it works day one.

The register is a weapon. v4.1's discipline — target rancangan strictly separated from measured results, a Status Implementasi table that admits what's only designed, Batasan Sistem as permanent limits ("the ability to refuse correctly is treated as equal to the ability to grant") — is rare in this genre and is exactly what wins skeptical reviewers at NLnet, Mozilla, BPK, and in an APBD hearing. No derived document may inflate past it.

The form factor is deeply right. E-ink + indirect LED halo + no camera by default + fingerprint stamp-button that mimics the stempel convention: a notice board that occasionally glows is culturally legible in a kantor kampung in a way a screen-with-camera never will be. The document camera facing down at the desk is a privacy decision that wins both design awards and dewan adat trust.

2. Settled positions (decided 2026-07-05, pending Yose's sign-off where marked)

3. Personas = role × rhythm (the unification)

Proposal §9 defines roles (runtime configs with tool subsets). §11 defines institutional rhythm (deadlines, seasons, rotation events). Unify them: a persona is a role that activates on the institutional calendar. The Bendahara persona intensifies before LPJ deadlines; the Musrenbang persona wakes for the planning season. This costs nothing architecturally — it is a scheduling layer over existing roles — and it makes "PNS Digital" narratively concrete.

Persona roster (working)

Juru Musyawarah — the second flagship (after Juru Surat)

Musrenbang is the institutional hook Yose's "digital DPR" instinct was reaching for — same function, zero political landmine, an existing legal process with an annual deadline every kampung already must feed. The persona: ingests WhatsApp voice notes for weeks, clusters concerns (detak-detik's clustering DNA, reused — one fact, one owner), drafts the formal usulan on the official template, routes consent, emits the signed record.

The feature that makes her beloved: she remembers what happened to last year's usulan. Musrenbang's notorious failure is the accountability gap — proposals go up, silence comes back. Opening the season with "last year we asked for five things; two funded, one in progress, two unanswered — resubmit?" is institutional memory (§10) doing what no human clerk sustains across rotations. The longitudinal usulan ledger, published PII-free to ETNOS, becomes the first dataset of its kind in Indonesia.

4. The moat stack (bird's-eye completeness pass)

Walking the loop — intake → reasoning → execution → signed record → audit → memory → anticipation → intake — and asking where it leaks. Ordered by moat depth:

4.1 Own the standards (deepest moat: software gets cloned, specs don't)

• The signed civic record format. Publish the audit-record schema (§19G) as an open spec with a public verifier — a standalone verify page + CLI usable by BPK, Inspektorat, journalists, citizens, with no relationship to us. Every document carries a QR; anyone who scans sees the signed record without PII. Citizens become the enforcement network for free. Long game: propose toward BSSN/ANRI as the standard for auditable AI-assisted administration. Whoever writes the spec everyone verifies against owns the layer.
• The skill-pack format. Formalize "paket skill": a signed, versioned package = template + validation rules + policy schema + test cases. Debian packages for civic transactions. Others author (Diskominfo devs, OpenDesa community, Uncen students); Abstraksi certifies and signs; the ETNOS registry distributes. Converts the "12–18 months to replicate" linear moat into a network-effect moat.

4.2 Close the leaking loops

• Citizen receipt-check. After delivery, QR scan → "diterima dan digunakan." One tap = §41's impact metrics collecting themselves.
• Rotation handover. The law already requires berita acara serah terima at every leadership change; nobody produces a good one. Aksara auto-generates it with a complete pending-case inventory, signed by outgoing and incoming. Another transaction, another wedge, and the emotional core of "the institution that stops forgetting."
• The rejection corpus. Every refusal with reason code (§19) accumulates into something nobody on Earth has: a corpus of correctly refused civic transactions. Compliance-training gold, eval gold, and the receipts behind "refusing correctly equals granting correctly."

4.3 Become the measurement layer

Anonymized audit-metadata, aggregated on ETNOS = a continuous, cryptographically-grounded measure of frontline service delivery. SPBE is annual and survey-based; ours is live and verifiable. Publish as Indeks Layanan Lapis Pertama — public dashboard, public methodology, per-district. The entity that measures delivery is politically very hard to displace. The device participates: the e-ink panel ambiently displays the institution's own numbers (documents this month, average processing time). Self-published SLA as furniture; accountability by default. (This is detak-detik's receipts doctrine scaled into infrastructure.)

4.4 Multi-hop Civic Transactions — the federation carries work, not just content

Real bureaucracy is multi-institutional: surat pengantar flows kampung → distrik → kabupaten. FP session semantics are already adopted (§21); design the multi-hop transaction now: kampung node drafts, distrik node countersigns, every hop consent-gated and audit-chained. First demo needs two nodes. A protocol for inter-institutional agentic workflows with consent chains exists nowhere. If ETNOS is where institutions' agents transact with each other, the federation stops being a forum with extras and becomes the civic rails. Endgame moat.

4.5 Resilience mode — the node as emergency infrastructure

Meshtastic is already in Lampiran F; LoRa is already optional; the CRDT queue is already delay-tolerant. Specify: on connectivity loss (disaster, outage), nodes form a LoRa mesh carrying emergency traffic on the same queue. Nearly free to spec, viscerally relevant to Papua's geography, opens the BNPB budget line. "The government notice board that becomes a lifeline when the network dies" — the most solarpunk sentence in the system.

4.6 Language flywheel — own the benchmark too

Consented interactions in bahasa daerah improve local models → better service → more usage; corpus held per-community, fine-tuned models co-owned (Phase 3, §43). The near-term free move: publish per-language eval benchmarks (Mee ASR word-error-rate, NLU tasks). Nobody can claim to serve Papuan languages without measuring against our benchmark. Companion move: initiate Mee on Mozilla Common Voice — free infrastructure for consented voice collection, a headline ("first Papuan language on Common Voice"), and perfect optics for the Mozilla Technology Fund application.

4.7 The Melanesian arc — the ten-year frame

From Papua, for the Pacific. PNG shares the island and is the most linguistically diverse country on Earth (~840 languages); Vanuatu and the Solomons share the extinction crisis AND the governance pattern: councils of chiefs, customary authority, collective decision-making. Community mode with threshold-council consent maps onto Melanesian customary governance almost one-to-one because it was designed from inside that worldview. ETNOS federated across the Melanesian arc, CARE-as-Code as shared protocol, DPG status opening UNDP/UNICEF channels. Federation means expansion is invitation, not conquest. No Jakarta startup can claim this story. (Connects to P02 Melanesian Zomia paper.)

4.8 Neolab seeds (concrete, not dreams)

• Mycelium-composite enclosures grown on sagu-processing waste for the artisan edition — real materials science (Ecovative-class) on a Papuan substrate, slotting into the existing roadmap line (§43 "edisi artisan").
• Indera Alam bioacoustics as research-grade biodiversity data → the first Universitas Cenderawasih partnership.
• The neolab starts with the device's own body and the device's own ears, not with a building.

5. The ontology test (feature triage — apply to everything)

Every proposed feature must be expressible as:

entity type × persona (role × rhythm) × transaction type, under one of two modes (government/community), touching one or both federation planes (content/memory).

Examples: Sentinel Alam = pos pengamat × Penjaga × signed observation. BUMDes = entity instance × Bendahara × LPJ report. Loka = community × Juru Tulis Adat × spatial artifact under CARE tier. If a feature can't be expressed in the ontology, it doesn't belong — that discipline is what keeps this a system rather than a pile of features.

The complete elegant system, one line: one primitive (Civic Transaction, now multi-hop), one ontology (memory/identity/consent/rhythm), personas as role × rhythm, two modes, two planes, two open standards (record format, skill-pack format), one public index, one resilience mode, one expansion doctrine (federate westward).

Five moats deep: integration depth · consented corpus + CARE-as-Code · institutional trust network · standards ownership · measurement ownership.

6. The $0 demo stack and build order

Infra (all free): one Oracle free ARM VM (4 OCPU/24GB, per existing ETNOS research) runs PieFed + Hermes runtime + LiteLLM + WA bridge + Piper TTS. Cloudflare Pages hosts the ETNOS client (Honai Metro), interactive demos, blueprint/BOM pages. This phone builds aksara-cli v0.

Channel strategy: Telegram bot first (free, zero ban risk) to prove the loop → WhatsApp Cloud API free tier for the real Skenario Rancangan 1. (Pricing verified in research/2026-07-05_stack-hardware-sota.md.)

Demo narrative arc (three acts, one story for the government audience):

1. One message becomes a signed document — Skenario Rancangan 1 end-to-end (Juru Surat).
2. One season becomes a signed usulan — Juru Musyawarah with last-year memory.
3. One federation carries them all — c/transparansi filling with verifiable records; the verify page; the capability registry.

1. aksara-cli v0 — 2–3 surat tools, real PDF on official template, dev-key signing (labeled), Merkle log, aksara verify. Proves the architecture; buildable on this phone.
2. Civic Transaction interactive demo + e-ink panel UI simulator (CF Pages, client-side). The panel UI is real software (1404×1872 grayscale web app) that later ships on hardware — not a mockup.
3. VM + Telegram/WA end-to-end — Hermes + LiteLLM + aksara-cli = live Skenario Rancangan 1.
4. ETNOS polish — c/transparansi receiving demo audit metadata; ruang Mee with CARE labels from day one.
5. GAMBAR pack (all 11 visuals) + 2-page exec summary for Papua Tengah — interleaved early: the APBD window is open now.
6. Juru Musyawarah demo (Musrenbang flow).
7. Credential sweep — DPG registration, Common Voice Mee initiation.

⚠ The urgent clock: the proposal's own funding table says APBD Papua Tengah 2027 planning runs June–October 2026. Today is July 5. The government-facing pack (item 5) is the single most time-critical artifact in the portfolio; everything else can slip a month, that can't. (Exact remaining entry points: see research/2026-07-05_apbd-otsus-window.md.)

7. Exploratory (flagged, not settled — needs research/deliberation)

• ETNOS verified identity as later civic SSO for local services. Big claim, real risk (identity provider = honeypot + political exposure). Park until Phase 2 discussion; note only.
• Wali data legal vehicle: what legal form holds community data custody — yayasan? koperasi? BUMKam? A data-trust structure aligned with UU PDP and adat authority needs real legal research (Uncen hukum adat partnership candidate).
• Procurement packaging: whether the pilot enters as swakelola, e-katalog lokal, or hibah — pending the APBD research report.
• JEPA/state-space models for institutional rhythm and sentinel anomaly detection — stays on the research trail (Lampiran D), never promised in pilots.

8. Needs-from-Yose ledger

Moved to LEDGER.md (single owner, one fact one owner). The ledger now also carries the APBD-scout items: Nabire introductions, the RIIM campus partner (deadline 31 July), letterhead/legal docs, INAPROC credentials, and phone verification of KUA-PPAS 2027 status.

8b. Same-day corrections (post-research, 2026-07-05 evening)

The three scout reports (research/2026-07-05_*.md) landed after this doc was written. Corrections that override statements above:

• Oracle free tier is now 2 OCPU / 12 GB (cut ~June 2026), with idle reclamation. The §6 "one 4/24 VM runs everything" line is re-planned in ops/00_zero-cost-deploy.md; Hermes Agent verifiably runs on Termux (MIT), which helps.
• WhatsApp: Meta banned third-party AI chatbots (Jan 2026) and unofficial clients remain ban-risk. Telegram carries the demo; the WA pilot needs a policy-compliance read first (ops/01).
• The APBD clock is tighter than §6 stated: Musrenbang Otsus/RKPD 2027 closed 30 April 2026. The live doors are KUA-PPAS 2027 (weeks), RAP Otsus via SIPPP (Aug–Oct), RKA-SKPD defense (Oct–Nov). See proposal/00_papua-tengah-strategy.md.
• Funding table refresh: NLnet NGI0 Commons closed permanently; Mozilla 2026 and PIDI 2026 closed; Shuttleworth wound down; BRIN RIIM Batch 2 open 1–31 July 2026. company/01_funding-tracker.md supersedes v4.1 §49.
• Good news verified: Gemma 4 E4B (Apache 2.0, audio-in) and VoxCPM2 are real; Photon has native PieFed support since v2.2.0 (rebase opportunity for Honai Metro); PieFed now has plugin + theme systems; "agents as fediverse citizens" space is still open — we would be early.

9. Cross-references

• Canonical source: ~/workspace/Aksara_Master_Proposal_v4_1.docx (v4.1 is single source of truth; this framing doc feeds its next revision, not the other way around).
• Prior thinking: ~/.hermes/workspace/papers/ (P01 algorithmic confederalism, P02 Melanesian zomia, P03 power gradient diagnostics), ~/.hermes/workspace/blogs/ (01–11).
• Research (dated, sourced): abstraksi/research/.
• Spec docs: see abstraksi/README.md (structure defined 2026-07-05).