Investigación sobre IX

Dominio: Experiments on IX, the GuitarAlchemist Rust workspace: streaming pipelines, expert routing under budgets, sandboxed stages, agentic loops and the oracles that judge them, each published with its receipts

Módulos

IXR-001 · Flujos con presupuesto: contrapresión, una compuerta que aprende y una verificación de Petri

Catálogo de cursos (previsto)

Head: skeptical-auditor Established: 2026-10-16 Subject: experiments on IX, the GuitarAlchemist Rust workspace

Every number in this department comes from a receipt: a pinned commit, a result file or a test run, named beside the number. A result whose receipt was lost is reported as lost. Code links are pinned to full commit hashes, so they keep pointing at the code that produced the result.

Courses

Code	Title	Level	Duration	Languages
IXR-001	Streams Under Budgets: Back-Pressure, a Learned Gate and a Petri Check	Intermediate	40 min	en, fr, es

IXR-001 is in en/ixr-001-streams-under-budgets.md, with its translations under fr/ and es/. It has no research-cycle record: it was written from a tracer slice run in a Claude Code session, and it waits for a human review.

Research index

Five studies of 2026-09-22 to 2026-10-16. The first is taught in IXR-001; the others are planned as IXR-002 to IXR-005.

1. ix-stream: streams under budgets

Code: crates/ix-stream at commit 14dccd0, branch exp/loop-nodes. SQL nodes followed in 944f42c: a batching node kind, sql::Window (a DuckDB query over each window of values, read as the table batch) and sql::ParquetSink. Then e44f9fe names the route runs’ folders after every flag that changes a result, and corrects the numbers in 14dccd0’s message.
What it is: JSON values flow through a graph of threads joined by bounded channels, with back-pressure up to a duckdb -jsonlines source and cancellation upstream; a lossy algedonic bus carries feedback; a gate routes each item to one expert with a per-context UCB1 bandit minus a shadow price paced on each budget, under a hard cap; an SSE feed with event ids, Last-Event-ID resumption and replay draws an ECharts page; and the graph is checked as a Petri net before it runs.
Result: 663 router questions (221, three passes), Jev answers replayed from recorded receipts, Jev budget $0.0320792472, which pays 265 calls (265.2 at the recorded cost per call). Waiting on the chosen lane, with Jev replayed at 40 ms, the learned gate scores a mean 0.833 (0.830 to 0.837) over seeds 1 to 5 against 0.791 (0.787 to 0.795) for “Jev until spent”. It wins on every seed, by 3.6 to 4.8 points (4.1 on average), with 265 paid Jev calls each; the local router alone scores 0.688. Without waiting (seed 7, Jev at 400 ms), Jev’s capacity binds: 84 calls happen and the fixed rule beats the gate, 0.729 against 0.710. One live run with 12 Claude Haiku calls scored 0.840, the same as that seed without Claude: no evidence of a gain, and its files were overwritten. A retry loop of waiting links deadlocks (1728 markings explored); a lossy link frees it.
Taught in: IXR-001.

2. ix-wasm: pipeline stages in any WASI language

Code: crates/ix-wasm, built in four commits on exp/loop-nodes: dfaaef6 (the sandbox, Go builds), 6229c60 (the wasm.run pipeline skill), 8ce2a29 (all of WASI preview 1 linked, files and sockets closed), 4cf71df (Zig beside Go, one content-addressed builder).
What it is: a stage is a WASI command module, JSON in on stdin and JSON out on stdout, run in the wasmi interpreter. Every preview 1 function is linked, so modules from any toolchain load, but no directory is preopened: path and descriptor functions answer EBADF, sockets and signals ENOSYS, and imports outside preview 1 are refused. Each run is bounded by fuel, memory and output caps; the clock moves 1 µs per reading and the random source is seeded, so two runs on one input give the same output and spend the same fuel. Builds are cached under a BLAKE3 hash of the language, the toolchain version and the source; Go builds with GOPROXY=off and cgo off, Zig runs build-exe on one file and never a build.zig.
Measured (commit messages): a Zig 0.17 ReleaseSmall stage with 3 imports runs correctly in 0.2 ms on 116,370 fuel; Zig’s std.Io (30 imports, 14 closed) and V 0.5.2 through C and zig cc now load and run correctly; V’s own wasm backend still produces an invalid module; a Zig program that opens a file gets EBADF and exits 1. Zig’s default 16 MiB stack made each run start by zeroing it (10 ms instead of 1 ms), so Zig stages are built with a 64 KiB stack.
A trap worth keeping: without wasmi’s auto-dispatch feature, an unoptimized build dispatches through tail calls that LLVM does not make at opt-level 0, and the host stack overflows on the first long wasm loop.
Planned: IXR-002.

3. ix-loop: loops that write loops

Code: pull request 383 on GuitarAlchemist/ix, open, head dd6a42a (ADR-0011).
What it is: agentic loops written in sandboxed Lua 5.4 (only table, string, math, coroutine and utf8; load takes text only; 64 MB per state; only JSON crosses between loops) that run, judge and write other loops. A budget counts fuel, calls, a deadline and depth (at most 3); a child’s share is carved out of its parent’s and the unspent part is refunded, so the root’s budget bounds the whole tree. The semantic node runs claude -p headless with no tools on the person’s plan. tasks.judge runs a draft on instances that no loop in the run has seen, which closes the Goodhart trick of the experiment’s first run, where the writer returned the shown instance’s answer instead of a loop. Loops can only propose; a person accepts.
Evidence (PR body): one test per line of the slice’s “done when” list, 14 tests in all; six fuel-escape routes closed and tested, including two xpcall handler routes found by a probe during review. Known limit: one long call into C, such as a backtracking string.find, runs to its end before the hook can stop it.
Planned: IXR-003.

4. The Jev router study, pre-registered

Data and results: state/router-spike/RESULTS.md, jev/full/stage3-comparison.json and jev/musicse/stage4-comparison.json, at ix commit 7c37c8c on main.
What it is: TypeSafe’s hosted zero-shot classifier jev-1.13.0 against GA’s local embedding router, each stage pre-registered with its decision rule before any call, scored with exact McNemar tests.
Stage 3, all 32 skills (221 model-written prompts): Jev 209 right, the arena router c2-merge 152, production 145; Jev is right where c2-merge is wrong on 64 prompts and the reverse on 7 (p ≈ 1.3e-12). Escalating to Jev on a decline or a margin below 0.05 gives 211 but sends 133 of 221 prompts (60%) to Jev.
Stage 4, human-written questions (87 scored Music Stack Exchange titles): Jev 82, c2-merge 54, production 42. Production routes 45 of the 86 questions no skill can answer to a skill; Jev routes 5.
Spend: 970 calls in all experiments, about $0.060 computed from reported tokens at the rate card, not an invoice.
Planned: IXR-004.

5. Node design: a family times a verb

Source: a French design draft of 2026-10-16 (Flux, experts et nœuds Claude Code : le tri), written in the session that built ix-stream. It is not in a repository; this entry is its summary.
The rule: a node is one family times one verb. A family is one resource, for example claude, jev, wasm, artifact, skill, mcp, gh, voice or ga; its verbs are read, write, run and manage; every node is off by default and has its own budget.
What follows: writing is claude.ask with an output schema, not a new node; reviewing is a host oracle (validation, tests, a tool’s schema), with any LLM judge panel used only to refuse by default; managing is the existing store, where a loop proposes and only a person accepts; running needs two nodes because it touches two resources, skill.run (headless Claude Code with one skill) and mcp.call (one tool of a named MCP server).
Agentic roles are skills, not node kinds: a review, a proposal or a web search is skill.run with an existing skill, and each call declares the tools it grants: none by default, the web only when granted, never personal data. A new role is a new skill, written and reviewed through the artifact family.
Status: a proposal. claude, wasm and the proposal store exist in ix-loop and ix-wasm; skill.run and mcp.call do not.
Planned: IXR-005.

Computer Science (agentic loops, CS-002), Cybernetics (the algedonic channel, CYB-001), MetaQA (oracles and held-out tests), Systems Engineering (interfaces between stages), Psychohistory (preregistration, PSY-025).