Latest engineering review · 2026-09-20
See which systems are running, what the latest checks established and where the engineering is going next.
Engineering evidence · checked 2026-09-20
Source models, execution checks and component timings from the latest engineering review.
166,700 retained neurons
25,582,938 directed pairs
malecns-v1.0 is loaded in a running executor. The source import reports complete retention of its declared network.
Model dynamics, input mappings and anatomical coordinate coverage are specified separately.188,508 retained neurons
13,620,865 directed pairs
banc-888-v3 is loaded in a running executor. The source import reports complete retention of its declared network.
Model dynamics, input mappings and anatomical coordinate coverage are specified separately.30.3% lower CPU time
935.1 → 651.8 ms per 20 ms simulated interval in the recorded laptop comparison.
Same 200 physics steps per interval. Eight measured intervals per arm; shared-host component timing.6.2–7.1% lower CPU time
The narrow prefetch update improved both recorded laptop tapes. Complete mutable state and weights matched byte-for-byte.
Neural-kernel measurements with frozen and learning-enabled profiles. Whole-loop cost includes physics and other work.One shared-laptop ABBA comparison, eight measured intervals per arm, with retained physics cadence. Component timing.
Fixed retained-image tapes, eight measured intervals per variant. Complete state and weights compared byte-for-byte over four intervals. Neural component timing.
Actual simulator wiring, source identity, complete activity, image hashes and clean checkpoint/exit were checked.
ERAIS source: 6f20efa8bf6f6f2a86312d0fb80eef1460bb19a1. Each measurement retains its source receipt digest in the downloadable record.
Products, records and research
Start a project
Explore the Discord pilot, discuss a supported model conversion or define a research project around your workload and hardware.
Products and projectsInspect the results
7 retained systems records through 2026-09-05 cover training, memory and execution. Their original environments and dates remain attached.
Open the recorded resultsFollow the research
The next research question is whether richer cognition and greater stored capability improve useful quality while the active path stays constrained.
Explore the native programmeEach product page states its current availability. Each technical result states the model, task and conditions tested.
Live availability is shown on Doom, Minecraft AI, Unified World, Story Lab and the interactive lab. The Fracture API guide describes the public integration.
Open the detail you need
Latest systems receipts through 2026-09-05
Local CPU, cloud CPU, T4 and A10G results retain their source versions, measured workloads and limitations.
A same-host developmental diagnostic, not a matched-quality control, paper reproduction, whole-system energy measurement, or model-superiority result.
earlier component comparison / no quality promotion
Selected Gemma MLP execution measured 1.982× the speed of its previous implementation on 48 retained rows, with unchanged output for those rows.
Boundary: An earlier covered-bank component configuration, compared with its previous implementation rather than a dense-model baseline. It does not establish complete-bank quality, end-to-end generation speed, held-out quality, or current serving savings.
ERAIS receipt · ab5cf5bd39 · 366110742bb7…
paired short hardware comparison
Two cloud CPU configurations resumed the same retained 3M checkpoint and completed the same 27-update interval. The faster configuration measured 1.390163× the training throughput.
Boundary: This compares hardware, physical-core allocation and thread budget, not ERAIS architecture efficiency. One short training interval does not establish sustained full-run speed, energy savings, finished training, or model quality.
ERAIS receipt · ab5cf5bd39 · 366110742bb7…
descriptive diagnostic
At TinyStories paper-text scale on the same CPU host, the corrected ERAIS full-stack window measured 3.40× the dense reference throughput and 68.45% lower CPU-package energy per reported valid target.
Boundary: A same-host developmental diagnostic, not a matched-quality control, paper reproduction, whole-system energy measurement, or model-superiority result.
ERAIS receipt · e9d4068464 · 0aa8595157e2…
authenticated complete segment
A cloud CPU continuation completed 16.79M valid targets with exact terminal state, sustained throughput, zero swap events, and high use of its declared thread budget.
Boundary: An authenticated bounded CPU continuation. It does not establish calibrated cloud energy, reconciled billing, useful generation, accelerator superiority, or production readiness.
ERAIS receipt · 5d526e4e53 · 8a77b33856c4…
operator-authenticated mechanics pass
The repaired TinyStories full stack completed 205 updates on a T4 with bounded quality movement, active cognitive paths, compact sparse execution, and an exact-resume checkpoint.
Boundary: A single-worker T4 mechanics and resource result. It does not establish useful learning, total-system energy, dense superiority, multi-GPU scaling, or production readiness.
ERAIS receipt · d5c2794a65 · 311d52be34cb…
governed bounded training result
A governed T4 training intervention reduced peak accelerator memory by 45.13% while retaining more than 100,000 valid targets/s and terminal quality inside the registered window.
Boundary: One exact-source T4 training window. It does not establish a larger-batch or long-learning result, energy benefit, dense or inference superiority, multi-GPU scaling, or production readiness.
ERAIS receipt · da89d1bdb0 · eb5cb1dcd03a…
authenticated bounded systems pass
At 1.61B stored-parameter scale, a larger-batch A10G arm completed ten finite updates, exercised the full governed stack, retained exact restartability, and exposed GPU occupancy as the next optimisation target.
Boundary: A ten-update systems result at stored-parameter scale. It does not establish useful generation, equal-token learning improvement, distributed scaling, end-to-end energy, or commercial readiness.
ERAIS receipt · 47b9e613b9 · 576aa658e07b…
Decision map
A compact view of the public programme. Every item separates technical maturity from commercial availability and names the evidence needed to move.
Bounded work or evidence that can be inspected today.
commercial
Four scoped routes can enter qualification now: 10-Day AI Fit Test, Private AI Deployment Sprint, AI Evidence & Release Audit, Research & Benchmark Partnership.
Evidence boundary
NeuroForge is pre-revenue and has not completed an external pilot. Every route starts with written scope, evidence requirements and stop conditions.
Evidence exit
Accept the first qualifying engagement only when scope, data access, evidence and owner conditions are explicit.
research
Two governed native-byte arms completed 262,400 updates and 536.3M valid byte-level targets each on the same bounded four-thread CPU envelope. The richer cognitive stack learned materially differently, but the fixed evaluation panel did not support a quality promotion.
Evidence boundary
The native-byte mechanism and exact restart path passed, but the preregistered fixed quality panel did not promote a quality advantage.
Evidence exit
Test whether the richer learning trajectory transfers to qualified fixed holdouts, then compare stored-capacity and active-path matched dense controls under repeated independent evaluation.
research
The current public state contains 7 source-pinned systems receipts and 6 architecture milestones spanning bounded local CPU, cloud CPU, T4 and A10G observations.
Evidence boundary
These receipts establish observed execution and resource behaviour in named runs. They do not establish matched-quality superiority, universal scaling or a production service envelope.
Evidence exit
Repeat the decisive native comparisons under matched quality, resource accounting and independent evaluation before promoting comparative claims.
The nearest experiment or operating gate that can materially change the decision.
research
The next research question is whether richer cognition and greater stored capability improve useful quality while the active path stays constrained.
Evidence boundary
Schedule, representation and geometry results remain experimental until fixed holdouts, fair controls and resource accounting agree.
Evidence exit
Complete qualified fixed-holdout, stored-capacity-matched and active-path-matched comparisons with repeated independent evaluation.
Claims and surfaces deliberately withheld until their own gates close.
governance
The /chat/ workspace and /api/chat gateway need their own reviewed backend, access controls, monitoring and model labelling before a protected workspace pilot.
Evidence boundary
This hold applies to /chat/ and /api/chat only. The public /v1 native text preview and Unified World follow their own live availability and capability contracts.
Evidence exit
Qualify the protected workspace backend, identity, retention, security, monitoring, rollback and capacity controls.
research
No universal quality, speed, cost, energy or production-superiority claim is promoted from the current native or systems results.
Evidence boundary
Mechanism evidence and source-pinned receipts remain public with their scope; promotion waits for fair, repeated, matched-quality comparisons.
Evidence exit
Reconcile qualified dense controls, fixed holdouts and full resource accounting under an approved claim review.
The full roadmap includes every governed workstream and retained boundary.
Inspect the complete evidence-gated roadmapCommercial availability
10-Day AI Fit Test (A$4,800 ex GST) tests a defined workflow. Deployment, evidence review and research projects have their own scope and acceptance requirements. Fracture and Discord enquiries start on their product pages.
One repeated non-safety-authoritative workflow, one decision owner, an approved target or reference host, three to ten lawful non-sensitive cases, clear data/network/output limits, and a buyer review inside the delivery window are required before work starts.
A signed statement of work, approved data/model rights, representative frozen inputs, target environment, and acceptance owner are required before work starts.
The release must be identifiable and the customer must lawfully provide enough artifacts, decision history, and reviewer access to reconstruct it.
Named industry and method owners, lawful rights, a frozen baseline and holdout, negative-result protection, compute funding, IP/publication terms, and independent review must be agreed.
Design-partner boundary
Discovery is permitted. Delivery or co-development starts only after the named pre-contract evidence closes.
Exact-candidate external security and denied-path closure Representative load, admission, and at least 24-hour soak evidence Partner-environment restart, restore, rollback, and independent-operator completion
Full 14-stage fresh-workstation execution Second-operator completion with founder interventions measured End-to-end artifact portability and customer-operability evidence
A real manual external audit or equivalent repeated workflow A minimum cross-customer schema proved against the bounded use case Second-reviewer completion and repeatability evidence
Historical public-evidence assembly · reviewed 2026-07-28
This is a historical release-candidate record. It remains inspectable without being mistaken for the date or state of the current native and systems programme.
The July sanitised handoff passed canonical integrity, but the release remained held because active phase-discovery validation failed during assembly.
Boundary: This is a historical release-candidate record, not the date of the current systems programme and not a public release authority.
Active validation failed.
Exit: Close or explicitly retain the named validation failure in a newly assembled exact-SHA evidence release.
Active validation failed.
Exit: Close or explicitly retain the named validation failure in a newly assembled exact-SHA evidence release.
Active validation failed.
Exit: Close or explicitly retain the named validation failure in a newly assembled exact-SHA evidence release.
Historical claim reconciliation · 2026-07-12
The terminal handoff reconciles 44 claim rows and 99 evidence links with 0 recorded reconciliation issues. It remains historical: newer systems receipts do not prove or promote those claims.
Four bounded engagement routes can enter qualification now. The current native programme has two completed long training arms and is testing whether richer cognition transfers under fair controls. Latest systems receipts run through 2026-09-05; the July detailed-claim and evidence-assembly records remain historical where marked.