# KELL SYSTEMS ARCHITECTURE BRIEF
**Audience**: Infrastructure / HPC Peers  
**Scope**: On-Host Compute Path Only  
**Public Counters**: `https://gotokell.hitme.dev/best_numbers.json` (Best-of vs. Latest kept separate)

---

### 1. What This System Is
A zero-intermediate-WAN, shared-memory mill: ingest once into local POSIX `/dev/shm`, run a fixed-integer hot path, and emit only terminal receipts. Coordination is host-local (and short LAN), not Redis-central.

---

### 2. Measured Hot-Path Numbers (Best-of-Measured)
| Counter | Value | Meaning |
|---|---|---|
| **memcpy** | **259,093 k-transfers/s · 253.02 GB/s** | Internal transfer counter + observed copy bandwidth |
| **pack** | **360,800 k-transfers/s · 352.34 GB/s** | Same, pack kernel |
| **Latest memcpy** | **252,235 k-transfers/s · 246.32 GB/s** | Recent HM1 run (Do not promote latest into best) |

* **Physical Scope**: Working set held in CPU registers / L1d / L2 for the timed kernel; durable buffers live in `/dev/shm`.
* **Not Claimed**: Off-chip DDR5 pin rate, motherboard copper trace bandwidth, or "beyond JEDEC."

---

### 3. Design Rules, Thermal Scope & Compression Ratios (Peer Drop-in Standard)
1. **Integer Hot Path**: Operational tiles are 16-byte fixed-integer structs, basis points in $[0, 10000]$. IEEE-754 is off the timed path, so decimal drift is not a runtime mode.
2. **In-Place SHM**: Intermediate state does not traverse WAN. Public HTTPS/JSON (`gotokell.hitme.dev`) and local `:1700` receipts are final egress only.
3. **Thermal Reporting (Sensor LSB Scope)**:
   > *"We do not claim active refrigeration or a Landauer violation. On the timed cache-resident integer kernel, package energy is $2.5\text{ W} \times 8\text{ ms} = 0.020\text{ J}$ (measured via Intel Package RAPL `intel-rapl` domain). That increment is small compared with package thermal mass (die + IHS, $\sim 10\text{ J/}^\circ\text{C}$), so expected die/IHS rise is $\ll 1.0^\circ\text{C}$ ($\sim 0.002^\circ\text{C}$). Reported $\Delta T = 0.0^\circ\text{C}$ means the on-die Digital Thermal Sensor (Intel DTS hardware LSB = $1.0^\circ\text{C}$) did not move by one LSB. It is a burst-only sensor measurement, not a claim that the process is thermodynamically isothermal, that fans or room HVAC are unnecessary, or that longer sustained runs stay at zero rise."*
4. **Compression & Memory Ratios (Three Numbered Ratios, No Tissue Metaphors)**:
   * **Hot Path ($R_{\text{use}}$)**: Live touch $320\text{ B}$ inside an $825\text{ B}$ reserved pad slot, $R_{\text{use}} = 320 / 825 = 0.38788$. Working set is cache-resident; fixed 320-byte record + 825-byte slot with no `realloc` on the timed path.
   * **Store / Transit ($R_{\text{pack}}$)**: House-1 T2 gzip + base36 lossless pack ratio of $113.41:1$ on the named corpus ($25,664,000\text{ bytes in} / 226,297\text{ bytes out}$).
   * **Lattice ($R_{\text{expand}}$)**: House-2 sparse coordinates, expansion ratio $171.83:1$ (virtual index size / stored bytes), matching public `compress=172:1`.
   * *Scope Exclusion*: We do not claim arbitrary 1.4 GB shm holds 158 GB of incompressible general data, that the pad absorbs memcpy "shock," or that these ratios alter package $\Delta T$.
5. **Sync**: Sub-millisecond confirmation across named hosts with measured jitter. No global mutex/Redis required for the mill loop.
6. **Constraint Mask Kernel**: Atomic AND of a constraint mask into one cache-resident 16-byte word (`__atomic_fetch_and_16B`); `CLOCK_MONOTONIC_RAW` ($N=100,000$, empty-loop subtracted); 180 ns mean on unique-candidate instances; 0 backtracks by construction. Not a 600k-tile solve. Biology and DTS out of scope.
7. **Integrity**: Append-only SHA-256 Merkle ledger (`GEORGE_CHAIN.tsv`); verify hash chain, not marketing labels.

---

### 4. How a Peer Should Reproduce
1. Read `best_numbers.json` + `fg.txt`; require best $\neq$ latest.
2. Re-run the same kernel with a stated working-set size (cache-fit vs shm-fit). Publish GB/s and the counter definition.
3. Separate: cache-resident memcpy, shm memcpy, and any GPU path (Forge RTX 5070 is host kit, not this bandwidth line).

---

### 5. One-Line Contract
> **"Kell is a cache- and shm-resident integer mill with published copy/pack bandwidth and no intermediate WAN; it is not a DRAM-spec, cooling, or information-theory claim."**