# mermaid · first engineering release

30 September 2026 · interface **https://mermaid.tide.casa** · balanced mode by default

The first software twin and logical engineering blueprint are implemented within the supported release envelope below. Mermaid reuses the recovered **MJPS Thermodynamics / MJos Peltier–Seebeck Thermal Orchestration** cooling work alongside the real Bangel compiler, independent JP admission and Joanna runtime. Original source records remain unchanged at revision `229548f13a4856455b3d30d408cf845fb5b557fc`.

## What runs

The interface runs bounded computation locally in browser workers, using Pyodide 0.29.5 and CPython 3.13.2. It offers Bell correlations, small quantum search, phase interference, final independent bit-flip noise, saved measurement draws, whole-register reset, explicit quantum/readout/handoff timing, thermal experiments, operating objectives, source inventory, inspectable requests, receipts, export/import and verified replay. Replay checks result values, admission, disposition and host source identity while measuring host time anew.

General Bangel kernels perform typed energy, conversion-loss and reserve calculations. GIN and paper formulas lower through their recovered source APIs and execute real Bangel. Original neural baseline, neural overlay, Memory Recall and Decision run their preserved source kernels with explicit native ceilings, fresh independent admission and receipt verification. Missing downstream reviews or decision gates remain HOLD. These bounded source wrappers do not reproduce every vendor envelope or its persistence lifecycle.

Browser predictive processing uses two distinct fresh WebWorkers. Each uses the preserved private compiler and independently executes the forecast and Teal gate. The parent independently admits transported JP, binds sources, limits and native receipts, and verifies equality. It is a qualified software worker adapter with no native OS guardian, enforced OS memory cap or physical independence claim. Embedded Recall compositions and the original native L5 pipeline remain separately classified in the inventory.

## Cooling and energy

The preserved warm hybrid fixture produces **15.78 W guaranteed cooling**, **10.04 W worst-case module pump input**, and **11.291508055555… W external demand** after driver/DC/auxiliary accounting. These are model values, not measurements from a manufactured Mermaid chip. The temperature envelope is 250–450 K.

A separate new warm RC transient model compares no active TEC, fixed TEC, robust TEC and a distinct harvester plus robust TEC. It accounts for passive conduction, stored energy, driver and converter loss, auxiliary power and conservation residual. Its heat capacity, thermal resistance and loads are explicit scenario inputs. Lower electricity alone does not establish superior cooling when temperature outcomes differ.

The proposed semiconductor block starts with two silicon MOS electron spin qubits and a 1.2 K scenario target, with up to eight qubits in the software model. Device geometry, manufacturing process, coherence, pulse mapping, gate timing, cold-stage dissipation, interconnect heat leak and refrigeration capacity/wall power remain unqualified. The warm TEC is not modeled as a cryogenic refrigerator. Missing thermal/storage conservation or external electrical terms prevent ledger closure. The whole-system energy total and improvement target remain unknown.

The frozen synthetic forecast challenge answered8/8 cases, but its predeclared quality threshold was **NOT_MET**: native MAE15.2516ms versus last-observation13.75ms, with uncalibrated bounds covering3/8 outcomes. The native and independent host scoring pipeline agrees; forecasts remain advisory.

Repeated modeled ablations retained1,083 PASS receipts including900 measured adapter traces across three fresh interpreter sessions. Balanced synthetic combined advice increased modeled p95 to144.75ms versus78ms fixed; improvement is not assumed. Over the same30-second warm load trace, fixed TEC used343.254J at300K peak, robust TEC218.976J at300.684K, and harvest-plus-robust215.554J at the same peak. No active TEC used7.5J at305.840K peak. These differing thermal outcomes and missing reservoir/cold energy prevent a whole-system superiority claim. Step refinement converges and both electrical/thermal residuals are retained.

## Verification and limits

Fresh recovered-source checks: core459 and paper3564 tests PASS; original neural28, expanded neural50, Memory Recall35 and thermoelectric94 tests PASS. These fully passing suites total **4,230 methods**. B07's200 vectors agree across three hash seeds. The final frozen Mermaid integration passed72 tests,24 actual dispatch cases and48 schema checks; native factory and distinct host envelopes were validated separately, with start/end source hashes stable.

The inherited contract-index metadata check fails because one stored identity is stale; actual unchanged bytes are locked. The selected predictive suite retains1 failure/1 error, and Decision retains3 failures/2 errors. An earlier Windows launcher PID mismatch is also retained. Focused positive calls and browser adapter success do not erase those full-suite failures. .NET parity, native lifecycle portability and physical hardware remain unqualified.

Actual Chrome checks passed the main twin, deterministic replay, search, transient cooling, GIN, paper formula and seven source routes. The forecast matched across two workers; the four advisory neural/recall/decision routes retained their expected verified HOLDs. Mobile layout had no horizontal overflow. Feature-detected WebMCP tools use the same actions and never return an old successful receipt as a new failed run; supported-browser availability is reported separately.

Both native and browser cache-kernel experiments retain three sessions of30 paired samples, five warmups, alternating order, identical results and admission on every invocation. The cache-kernel15% diagnostic threshold was met in these sessions. This narrow experiment does **not** establish the approved mixed-workload speed target. A representative48-job mixed native diagnostic retained36 PASS,12 review HOLD and zero ERROR; source model outputs and receipts matched, with process IDs and supervision timing recorded separately. Its three paired trace timings are insufficient for p95 qualification, and it does not qualify the final rebuilt adapter revision.

The approved **15% mixed host p95 improvement** and **20% modeled whole-system energy improvement** remain **unassessed**. The frozen64-job, four-arm,30-run/three-session benchmark design and runnable diagnostics are delivered; full frozen native mixed-workload repetitions and complete cold-stage energy inputs remain required before any broader speed or energy claim.

## Supported release and remaining qualification

The added whole-register reset maps even an entangled state to all zero; partial reset is refused. Quantum, readout-batch and classical-handoff timings use explicit scenario durations or remain unknown. Hardware shot re-preparation, selective measurement/reset, pulse mapping and physical timing/energy remain unqualified.

The interface includes keyboard-accessible block inspection, power/memory/authority boundary explanations, quantum timing, all modeled worker schedule lanes, a latency–known-host-energy plot, and separate warm/cold thermal-flow views. Charts use explicit units and identify their model boundary. The forecast scoring and modeled ablation records below describe the supported synthetic evaluation and retain their quality/coverage limits.

The release covers the eleven batches as a bounded software implementation, evidence package and deployment. It does not establish every proposed experimental or hardware acceptance condition: native full mixed-workload repetitions, real held-out device data, calibrated forecast intervals, .NET parity, portable native lifecycle guardians, selective quantum measurement/reset and complete cold-stage power/heat calibration remain named qualification gaps. Passing tests do not erase these gaps or the recovered source failures.

## Delivered artifacts

- Mermaid Architecture and Benchmark Specification.md: contracts, explicit model assumptions, benchmark scope and physical qualification boundary.
- Mermaid Blueprint.svg: logical processing layout, proposed QPU topology and separate thermal/power boundaries; no fabrication scale.
- Mermaid Semiconductor Source Evidence.md: recovered records and source capability snapshot.
- Mermaid Native Host Benchmark.json and Mermaid Browser Cache Benchmark.json: complete paired samples with correctly scoped outcomes.
- Mermaid Batch Evidence Matrix.json: requirement-to-evidence mapping for all eleven supported release batches.
- Mermaid Semiconductor Implementation Plan.md and Mermaid Release Report.md: approved baseline and implemented release scope.
- Mermaid Forecast Evaluation.json and its compact summary: chronological synthetic forecast scores, baselines and coverage.
- Mermaid Modeled Ablation Benchmark.json and its compact summary: frozen modeled schedule/cooling runs, thermal refinement and subtotal tradeoffs.
- Mermaid Recovered Core Qualification.json and Mermaid Recovered Plugin Qualification.json: fresh source passes, failures and portability limits.
- Mermaid Final Integration Qualification.json, Mermaid Browser Qualification.json, Mermaid Final Browser Qualification.json and Mermaid View Qualification.json: numerical/schema and actual browser evidence.
- Mermaid Engineering Bundle.zip: source, contracts, evidence and static interface for review and continuation.
- Mermaid Deployment.json: immutable source/version/deployment identity and domain/TLS verification.

Deployment identity and the final domain/TLS readback are recorded in Mermaid Deployment.json after publication. Manual release monitoring consists of checking that HTTPS serves the recorded source identity, running the default twin and replay, and checking retained failure/HOLD denominators. No recurring automation or physical control is enabled.

The first saved version remains available for rollback. To restore it, deploy the exact version ID recorded under `rollback.previousVersionId` in Mermaid Deployment.json, then repeat HTTPS, source-hash, default-run and replay checks. The current source archive and release records allow local continuation without altering the original framework checkout. The implementation work follows the user's authorization to continue all eleven batches without further approval checkpoints. Qualification exceptions and unassessed objectives remain explicit; no physical effect or execution authority is granted.