State of the beta

Beta 1 (0.1.0b1) covers three packages: the compiler frontend and stdlib (this repo), the AVR backend (pymcu-avr), and the CircuitPython compatibility layer (pymcu-circuitpython). ARM/RP2040/RP2350, PIC, and RISC-V stay alpha on purpose. See Supported targets for the per-backend maturity labels and why.

This page collects the numbers from the five suites that back that claim, each with a link to the page that explains what it measures and how to reproduce it.

The five suites

Suite	What it measures	Result	Docs
User-program corpus	49 user-style AVR programs, each with an expected build outcome and a size gate	49 programs	pymcu-circuitpython/docs/corpus.md
CircuitPython API parity	Every digitalio/analogio/busio/pwmio/… symbol upstream defines, checked against this layer	230 symbols (175 provided, 55 allowlisted with a reason)	pymcu-circuitpython/docs/parity.md
MicroPython API parity	Every machine/utime/network/… symbol upstream defines, checked against this layer	292 symbols (74 provided, 218 allowlisted with a reason)	pymcu-micropython/docs/parity.md
HAL parity (tests/stdlib/test_hal_parity.py)	The register-level HAL’s own API, compared across all seven backend targets (avr, pic12/14/18, riscv, rp2040, rp2350)	191 facade/API deviations currently allowlisted, each tracked	docs/library/hal-parity.md
Differential oracle (tests/oracle/test_oracle.py, in the pymcu-avr repo)	109 probes compiled and run on the AVR emulator, diffed against CPython running the same source	109 probes: 78 match (7 documented divergences), 11 correctly refused, 20 tracked as filed compiler bugs (xfail(strict), suite green)	docs/language/oracle.md

What the oracle knows is wrong

The differential oracle does not just count matches. Its 20 tracked probes (as of 2026-09-15; re-verify at freeze, since this list is regenerated from docs/language/oracle.md’s “Compiler bugs, by cause” table) are 13 filed, open compiler bugs, each an xfail(strict) case so the suite stays green without hiding them:

Kind	Issues	What it means for a beta-1 program
Silently wrong value (no diagnostic, wrong answer)	#364, #390, #394 (one of its three probes), #395, #396, #397, #398, #401	An unannotated loop accumulator, a field read through with ... as, a nested comprehension, a ZCA __add__/__lt__, len(instance), two-index __setitem__, list[T].append() on the heap, and match on an array can each compute the wrong answer with no error. Each has a fixture pinning today’s wrong output so a silent fix does not regress. Two that were here, #393 (folded hex/bin/str) and #399 (a one-character index), were fixed on 2026-09-26 and their probes now match CPython.
Correctly refused, misleading reason	#391, #392, #400 (plus two of #394’s three probes)	The build fails with a CompileError rather than shipping a wrong answer, but the message names the wrong cause (a class with no explicit __init__, bytearray() assigned to self.field, and an Enum member read outside a plain assignment all give a diagnostic that points somewhere other than the real limitation).

The distinction from “silent” as the term is used for the beta-1 exit bar: every probe above is filed, disclosed here, and enforced by a test that fails the moment the bug disappears without a matching fixture update. What made three bugs found the night before freeze (silent-write-loss in @inline, a factory’s field going stale, super() with constant constructor args: #427, #429, #430) block the freeze instead of joining this table is that nothing had caught them yet. That was the standard applied at the 2026-09-15 freeze. It did not survive contact with the second instrument: the 23 issues of 2026-09-25 were all outside the table, and they were found by pointing a new kind of program at the compiler rather than by any suite already running. So the claim this page makes is the narrower and truer one. Beta 1 ships with every bug we know about written down on this page, and the honest reading of one day that produced 23 of them is that more instruments will find more.

What a second instrument found

The oracle above compares probes we wrote. A different instrument compiles libraries we did not write: the unmodified upstream Adafruit CircuitPython and MicroPython drivers, straight from PyPI, with the transactions they put on the bus diffed against the same driver running under the real interpreter. It found bugs the probe corpus could not, for the reason that makes it worth running: a probe exercises the construct an author already suspected, and a real driver exercises the combination nobody chose.

On 2026-09-25 that harness opened 23 issues in one day. Eleven are fixed and in the tree, and a string campaign closed five older ones alongside them.

What follows is what is still open and can reach a beta-1 program, with the ones that give a wrong answer and no diagnostic listed first, because that is the category this page exists to disclose. Three of the rows below came from the day after, on 2026-09-26, when this list was cross-checked against the probes the oracle still tracks rather than against the issues opened that week. Two of those three had been open for longer than any issue in the first draft of this table, and were missed for the dull reason that the draft was written by looking at recent work instead of at the measurement.

Issue	What it does to a program	Reaches
#490	The ATtiny register maps are shifted against the vendor headers, so analog_read() on an ATtiny85 writes ADMUX where ADCL lives	ATtiny only
#494	int() is a cast to int16, so int(46051.7) prints -19485	every target
#506	A str is stored in a one-byte slot, so returning one from a function, or taking one as a bare parameter, truncates its id	every target
#510	A method with a single call site is inlined, and the filled ellipse it expands draws the wrong pixels	every target
#495	A compile-time chip predicate stops folding once it is bound to a name in another module, so both branches are compiled into the firmware	every target
#500	Two per-chip HAL modules exist twice; the facade wires one copy and the other has drifted, so reading the unwired copy describes behaviour the compiler does not have	every target
#449	A type annotation reached through a module alias, t.uint8 after import pymcu.types as t, wraps where the bare name promotes: the same program prints 44 one way and 300 the other, with nothing said	every target
#446	A constructor reading a field off another instance it was handed reads 0 instead of the value: a driver whose __init__ takes a configured object and copies one field out of it gets a zero	every target
#439	match on a tuple pattern compiles under the Python front end where the C# one refuses it, and takes the wrong branch	PYMCU_PY_PARSER=1

Two more are about the toolchain rather than a program, and belong here because they decide whether the numbers on this page can be trusted at all. #496: the IR handed to a backend is a wire format with no version field, so a backend wheel built against an older frontend decodes an operator as a different operator and miscompiles in silence. #504: the diagnostic text format is an undeclared public interface that two shipped editor plugins already parse.

Two are quality of diagnostics, not correctness: #505 reports an internal compiler error as a diagnostic on line 1 of the user’s file, and #509 emits one diagnostic per invocation and drops a warning when the program also fails.

One is deferred by decision rather than left open by omission: #507, where enumerate(s) unrolls on length alone and eight characters of display.text() cost 12736 bytes while sixteen cost 4746. The cause is measured and the one-condition fix is verified; what is not done is the program-by-program account of which corpus programs change size, and that is not work to land days before a release.

What “beta” does and does not claim

• Beta means: the language surface is implemented and test-covered on the AVR backend, and it is validated on real silicon (Arduino Uno, logic analyzer differential harness). It does not mean every API symbol above is implemented: the parity suites above exist precisely to make the gap explicit and trackable, row by row, rather than asserted.

• Every allowlisted deviation and every tracked:#N / xfail entry in these suites names a GitHub issue. None of them are silent.

• keypad.KeyMatrix (pymcu-circuitpython#13) and the storage/os modules (pymcu-circuitpython#16) are out of beta 1 by decision, not by omission.

• The integer-width inference RFC (RFC 0002, tracked as PyMCU#364) ships only behind its feature flag if it lands before release, off by default, and documented as experimental. It is not part of the beta-1 surface claim above.

See also: Language Limitations for what stops each individual Adafruit CircuitPython library, and the release checklist for how this beta is built, smoke-tested, and rolled back.