EPIC: T27 backend — rings from the decision core to silicon

[gHashTag]

Everything below the interface is written in T27 and generated to its target. Swift keeps the interface and nothing under it. The single exception is the seed — t27/bootstrap, the minimal Rust compiler t27c, where hand-written Rust is the point.

A rule transcribed into Swift, Rust, Zig and Verilog is four rules that agree until someone edits one. Generated from one .t27 it is one rule.

Rings, innermost first

Like the layers of a trunk. A ring depends only on rings inside it, and is not started until the one inside it runs in production.

Ring	What	Depends	Target
T27-00	Decision core: retry, review, merge gate, capacity. No clock, no network, no store, no strings	—	Rust + Verilog
T27-01	A2A protocol: agent card, message, task, registry semantics	00	Rust + Postgres
T27-02	Orchestration: the Queen's tick, delegation, review sweep, bounded send-back	00, 01	Rust + inngest
T27-03	Transport: SSE, per-agent buffer, event ids	01	Rust
T27-04	Scoring: salience, reliability, latency — where TNF belongs	00	Rust + Verilog

Ring 00 needs no floating point at all: enums and integers, which is why it is also the ring that synthesises without argument.

Division of labour — read before touching anything

Another agent owns the t27 repository and the FPGA boards (wave W931, branch claude/igla-fpga-improvements). Their ledger shows an active yosys/nextpnr sweep across 19 tracts, per-arm XDC, and the TNF publication gate G8.

Therefore:

• Do not edit /Users/playra/t27. Compiler gaps are raised with them, not patched here.
• Do not program a board. A pnr or route job may be in flight.
• t27/fpga/HARDWARE_SSOT.md is authoritative for hardware. This epic does not keep a second register.

This epic owns the trios side: .t27 ring sources, the parity harnesses, and the service that runs them.

Verified here on 19 August 2026

The compiler builds; targets exist: gen-rust, gen (Zig), gen-c, gen-verilog, icarus-simulate.

Open toolchain, no Vivado: openFPGALoader, iverilog, yosys, nextpnr-xilinx.

Two stands answer over JTAG:

• -c ft232 / digilent_hs2 → idcode 0x3636093, Artix-7 200T. Their SSOT names this board QMTech Wukong V1 (XC7A200T-FGG676); the operator called it an ALINX AX7203. Same idcode, same die — the naming needs one word from the operator, and their SSOT wins until it comes.
• -c ft2232 / digilent → idcode 0x3727093, Zynq xc7z020 + Cortex-A9 (0x4ba00477) — tri-net's board.

Blocked on the seed

Ring 00 generates Rust today, but not valid Rust. Found by generating and reading:

1. Enum.variant emits as Enum.variant; Rust needs Enum::variant.
2. .variant shorthand emits as variant::.
3. switch as a function body emits an empty match x { }; — the arms are dropped.
4. ; comments inside an enum body become variants. Use // inside braces.
5. pub module name; was not parsed from a //-commented file; gen-verilog named the module unknown.

(1) and (2) block every ring. All five are in the seed, which is the other agent's repository.

Already built — do not rewrite

A2A is already a client/server split on Postgres: nine routes (register, unregister, heartbeat, agents, matrix, message, task/assign, task/update, stream), table agents, view agent_matrix. Ring 01 restates this contract; it does not invent one.

Railway project 999 already runs Postgres, Redis, inngest, postgrest and MinIO. The Queen's loop is a durable workflow, not a five-minute timer inside an app that has to stay alive.

TNF

Sign, exponent in balanced ternary trits, M-bit mantissa, 1 + E_t + M = N. No implementation was found in the trios tree; the other agent's ledger reports it measured at gate G8 (19/19 tracts routed, 14/15 within seed noise, bin16 exactly 1.00×, LNS16 an honest 1.46× outlier). Their numbers, not ours. This epic does not claim TNF is applied until a .t27 module in trios uses it and the answer is checked.

[gHashTag]

Two of the five seed gaps are closed. gHashTag/t27#2247 — [GOLD-RING], branched from origin/master, seal moved per FROZEN.md §5, docs/NOW.md synced.

• Verdict.escalate now emits Verdict::escalate.
• The shorthand .escalate now takes its type from the function's declared return type instead of emitting escalate::.

Both forms compile: rustc --crate-type lib produces an rlib from each, and from the ring-00 source a bee wrote (14912 bytes). Before the change neither did.

Corrected finding. Gap 5 is not about comment style. pub module name; is rejected at the top level — "Unexpected top-level token: KwModule" — while module name; is accepted. test_highlight.t27 uses the pub form, which is why it reads as supported; that file is a highlighting fixture and is never compiled.

Still open in the seed: switch as a function body emits an empty match; ; comments inside an enum body are parsed as variants.

Swarm status. The Queen picked #1280 herself once she could see this epic, and put bees on #1280, #1281, #1282 and #1283. #1282's bee committed 288 lines of tests/t27/ring00_verilog.sh before timing out. Two record-keeping defects surfaced and are under investigation before anything is merged:

• RING-00: the decision core in T27, generating valid Rust #1280 is recorded accepted with committedFiles: 1, and its branch does not exist.
• RING-00: Verilog from the same source, checked by simulation #1282 is recorded failed with committedFiles: None, and its branch carries the 288-line commit.

Nothing from this epic will be merged until those two are explained.