CineNexus Pro · AI media studio
Upscaling, video and audio restoration, and English–Thai dubbing with consented real voices and watermarks.
Proves: high-risk AI under governanceRead the case study report (PDF) ↗See the app in development ↗A home for developers, researchers and organizations to build Application Profiles on heain-core, propose real use cases, and share papers and related work.
Every tier above a Worker is the same Master, differing only in scope. When a Zone fails, its most stable Worker acts for it, and hands the slot back when the owner returns.
A seven-procedure protocol that carries work from Cloud to Edge over mTLS. Raw data stays where it was generated, and every significant decision leaves verifiable evidence. The core holds no application code; applications live in Application Profiles.
A fractal hierarchy that deepens with the organization; it decides where work runs and how authority is structured.
Swarm knowledge exchange (P7), self-healing role succession (P6) and policy-boundary reasoning (P5).
Image, audio, video and business models run inside P3 through heain-sdk, without touching the core.
An administrator-set ceiling removes ineligible nodes before selection.
Grants only the duties a node can carry, instead of letting it fail in a cascade.
Bounds upward flow and catches silent widening of approval thresholds.
In this phase we test heain-core against very different case studies, refining it for broad use before the license terms are set and the binary is released.
Upscaling, video and audio restoration, and English–Thai dubbing with consented real voices and watermarks.
Proves: high-risk AI under governanceRead the case study report (PDF) ↗See the app in development ↗Raw data stops at the district; only distilled totals go up. Every decision affecting an individual needs an independent Approver, and complainants cannot be linked to their complaints.
Proves: data sovereignty and independent approvalRead the case study report (PDF) ↗Warn locally first, report afterwards. When the network breaks during a disaster, warnings must continue and history must be complete on reconnection. Simulated across 4 provinces, 8 districts, 29 nodes.
Proves: STANDALONE mode and journalRead the case study report ↗Hat Yai flood replay (2025) ↗Bangkok flood replay (2026) ↗Earthquake replay (2025) ↗PM2.5 season replay (2026) ↗Adds a district tier without changing the protocol. Ballots are sealed at once and opened only under an approved grant, and cross-group vote coordination is detected from anonymous ballots. Tested on synthetic data only.
Proves: adding a tier and independent oversightRead the case study report (PDF) ↗New idea: blind tokens (PDF) ↗400 constituencies, 99,487 polling units: count offline with no loss, no duplicates, and secret ballots.
Proves: unlinkability and scaleRead the case study report (PDF) ↗Simulates the effects of five real attacks where the attacker holds only a leaked certificate or key. A taken machine reaches nothing above it and is cut off 1.2 seconds after revocation, every false update is refused, and work goes on with the cable cut or the Master down.
Proves: contained blast radius and recoveryRead the case study report (PDF) ↗A 5-machine lab replays a worldwide central outage caused by an operator's own change. No job is sent twice and sites work on for 30 minutes without a centre, but one wrong setting can still stop the whole service, so core needs further changes.
Proves: sites independent of the centre, and what still needs fixingRead the case study report (PDF) ↗Have work that needs AI at the edge, data sovereignty, or operation through network cuts? Propose it as the next case study.
Join and proposeheain-core will be released as a binary with a developer license, for building and testing Application Profiles. Register to be notified at release.
go get github.com/heainframework/heain-sdk@latest
The SDK enrolls, registers, sends heartbeats, writes audit records and enforces AI reasoning records for you. Other languages use the heain-agent sidecar.
All are built on one contract: manifest, App API and heain-sdk. None modifies the core, and each counts as done when it passes conformance C1–C14.
The HEAIN research series, from theory to protocol to running system. Members can submit papers and related work for the community.
The 5-machine lab passes all 38 checks: no job is sent twice even when a machine stops mid-job, the Master holds when 1,000 clients reconnect at once, and sites work on for 30 minutes without a centre. It also records 8 findings that need core changes, such as one wrong setting stopping the whole service in about 6 seconds.
A 3-machine lab simulates the effects of five real attacks and passes all 93 checks. A taken admin account can only propose once approver keys are off the Master, all 5 false updates are refused, and a Master restored from encrypted backup leads the lab again in 9.2 seconds.
An observer with no certificate found 50 votes added during aggregation in 3.07 seconds, a standby took over a failed reporting point in 14.7 seconds, and at national scale 478 nodes received all 99,487 unit forms in 34 seconds. Ballots open only with committee approval plus 3 of 5 outside organizations. Synthetic data only.
Volume 1 separates identity, votes and evidence into three lanes and passes 152 checks on 32 nodes. Volume 2 proposes blind-signature tokens per RFC 9474, so the roll and the voting machines together still cannot link voters to ballots; 155 checks passed. Synthetic data only; the work judges no allegation.
Bangkok flood: the first warning went out 0.03 hours after TMD's, instead of 105; 26 of 26 checks passed. Earthquake: the first CAP went out in minute 2 instead of 85; 26 of 26 passed. PM2.5: permits over the quota fell to zero; 28 of 28 passed. All figures come from simulation.
The fourth case study passes its first 6 rounds on synthetic data only, including sealed ballots opened only under approval and detection of cross-group vote coordination.
All 26 scenarios pass, with 4 real-event replays, a 24-hour soak and an 872-node scale test, plus a separate report replaying the Hat Yai flood of November 2025, as it happened versus with HEAIN. Thai and English editions.
The third case study simulates 4 provinces and 29 nodes. The first two rounds pass in full, including local warnings while cut off and sensors outside the trust boundary.
A 19-node, 4-tier simulation passed 306 checks; a 24-hour soak with 71 random faults processed 272,321 jobs with none lost or run twice; the 330-node scale test passed in full. Thai and English reports now have DOIs.
An English edition is now published; both editions are on Zenodo. See the Articles section.
The second case study after CineNexus Pro tests that raw data never leaves the district and that every decision affecting an individual passes an independent Approver.
All 244 checks pass. Over 24 hours with 70 random faults the system processed 216,080 jobs with none lost or run twice, and the study led to 9 fixes in heain-core.
The whole framework, from four papers and heain-core v1.0 to v1.3 proven on real hardware, ready to move to Alpha. Thai and English editions on Zenodo.
Automatic role return, operation through network partitions, and hierarchy-wide configuration governance, now on Zenodo.
The first AI media studio app built on heain-core. Follow its progress at app.cine-nexus.com.
All 232 checks passed, covering Master loss and return, a studio cut off, upgrades while running, PDPA erasure, backup and restore, and active security.
Multi-level succession, automatic role hand-back, and operation through network partitions.
Cases that weakened a gate of the Triad are closed at protocol level in heain-core v1.3.
A framework against cascading failure, authority creep and alert storms.
The Go heain-core passed live tests on an HP Z840 over mTLS.
All tests so far ran on a single workstation. Multi-machine results will be announced here.
Members receive news, take part in case studies, and get early access to heain-core when the binary opens.