Hypothesis-driven measurement of "did monomorphisation actually
buy us performance?" on a 100M-iter LCG hot loop, AILang mono'd
code vs. four C reference variants (direct-inlinable, direct-
noinline, indirect-monomorphic, indirect-polymorphic). Zen 3,
clang -O2, median-of-15.
Headline: H1 supported, but the mechanism is inlining, not
dispatch shape. AILang mono = hand-C direct (1.000x). Indirect-
monomorphic = direct-noinline (1.000x) — saturating branch
predictor makes the indirect-call cost vanish on this hardware.
Inlining is the actual 3.31x win; polymorphic indirect adds
another 21% predictor-miss penalty.
DESIGN.md Decision 11 gains a rationale paragraph reframing mono
as inlining-enabler rather than indirect-call-eliminator, with
explicit pointer to the bench. JOURNAL entry records the full
methodology, ratios, limitations, and the side-effect mono-pass
env.globals-seeding bug surfaced while building the AILang fixture
(separate RED-first debug iter to follow).
User feedback after first draft:
- family/iter not standard terms — milestone/iteration adopted
- ailang-docwriter folded into audit/ rather than left orphan
- explicit "Core rules adopted from Superpowers" section so each
inherited Iron Law is named at the spec level, not just implied
Five-skill pipeline (brainstorm, plan, implement, audit, debug) under
skills/, emitting durable artefacts to docs/specs/ and docs/plans/.
Codifies discipline currently spread across CLAUDE.md (TDD-for-bugs,
mandatory tidy-iter, bench-regression rules, feature acceptance) into
trigger-bound skills. Existing per-iter workflow preserved; the layer
formalises the pre-iter design step and the iter handoff contract.
Bootstrap spec — defines the system that future specs will pass through.
Records the iter close: AST schema floor (Def::Class/Instance) +
workspace registry with three coherence checks (Orphan, Duplicate,
MissingMethod), seven new tests (288 → 295), all three bench gates
green at close. JOURNAL queue updated: 22b.1 closed, 22b.2 next
(typecheck arms + FnDef.type.constraints + class-schema validation).
Adds DESIGN.md "Feature-acceptance criterion" as a top-level section:
a feature ships only if (1) an LLM author naturally produces code
that uses it, and (2) it measurably improves correctness or removes
redundancy. Aesthetic appeal and human ergonomics do not count.
Mirrored in CLAUDE.md as "Feature acceptance: LLM utility", paired
with the existing "Design rationale != implementation effort". The
two together narrow valid feature rationales to one thing: what the
LLM author gets out of the feature.
Trigger: the typeclass-design conversation around 22a. Rule was
implicit in many past decisions (Decision 1's JSON-over-text choice,
Decision 10's "what LLMs are good at" reasoning) but never stated
as a feature-gate. Codifying it now means future feature proposals
get evaluated against an articulated criterion instead of being
re-derived each time.
Documentation-only; no Rust, schema, or bench changes. Test state
288/0/3 unchanged.
Three drift items from ailang-architect, plus one false-positive
surfaced during verification:
1. DESIGN.md silent on closure-pair 4.14x finding (21'b).
Decision-10 ratified: "Workload scope of the 1.3x target"
paragraph scopes the retirement gate to linear/tree/poly-ADT
workloads; closure-pair carve-out documented as
representational cost (closure cell + env struct = 2 allocs
per step) until a slab/pool answer ships.
2. bench/compile_check.py corpus drift. Three fixtures added
(bench_compute_intsum, bench_compute_collatz,
bench_list_sum_explicit), re-baselined. Now 12 fixtures x
2 ops = 24 compile-time metrics.
3. baseline.json convention not codified. Note field gains
"max-of-distribution gets wider band than percentile"
convention discovered in 21'd.
4. (verification finding) bench_compute_intsum cross_lang
tolerances at 15%/12% fire on subprocess-spawn jitter for
sub-millisecond fixtures. Widened to 35% across all five
intsum metrics; convention recorded in baseline_cross_lang
note field (sub-ms fixtures need looser bands).
All three bench gates re-run sequentially after edits:
bench/check.py — 63 metrics; 63 stable
bench/compile_check.py — 24 metrics; 24 stable
bench/cross_lang.py — 25 metrics; 25 stable
Total: 112 metrics under regression coverage, all green.
288 tests passing, 3 ignored. No Rust changes.
Closes the apples-to-apples gap from 21'e. Adds:
- examples/bench_list_sum_explicit.ailx — same algorithm and sizes
as bench_list_sum, fully (borrow)/(own)/(drop-iterative)
annotated so codegen emits proper inc/dec instrumentation.
- bench/reference/list_sum_explicit_free.c — same algorithm
with explicit free() walking the chain after sum.
The full alloc+dec vs malloc+free comparison reveals two non-
trivial conclusions:
1. AILang's full RC pipeline is only 26% slower than glibc
malloc+free on this workload (rc/c = 1.26x). The implicit-
mode comparison's 1.42x was misleading — it counted neither
pipeline's free path. The fair ratio is 1.26x, materially
better than the previous read.
2. RC's dec is cheaper per cell than glibc free(). AILang
dec-tax: ~3 ns/cell. C free-tax: ~5.5 ns/cell. Plausible
cause: ailang_rc_dec operates on a known-shape cell with a
fixed-offset refcount and a static per-type drop fn — no
free-list bucketing, no header introspection, no global lock.
bump's advantage expresses fully: bench_list_sum_explicit.bump/c
= 0.42x means AILang at bump is 2.4x faster than C malloc+free.
Sets a useful upper bound on a slab/pool RC allocator's potential.
The 21'-family arc — bench-regression infrastructure — is now
substantively complete: 21'a (bench/check.py), 21'b (corpus
widening), 21'c (compile_check.py), 21'd (pure-compute fixtures
+ harness hardening), 21'e (cross-language hand-C), 21'f (explicit
apples-to-apples). 63 runtime metrics + 18 compile metrics + 25
cross-lang metrics under regression coverage. Any future iter
that regresses any axis beyond tolerance gets caught at the next
family close.
Remaining queue is back to substantive language work — Family 21
(typeclasses / polymorphic ADTs at runtime / pattern-binding
generalisation) is now an orchestrator-level fork that needs
direct user input.
Closes the question CLAUDE.md has carried since day one ("LLVM-
linkable, performance is extremely important") with data. Hand-C
variants of the four bench fixtures, compiled with clang -O2,
each carefully matching the AILang algorithm and explicitly
documenting representation differences (cell width, leak policy)
that affect the ratio.
Three substantive findings:
1. Pure-compute parity with C: bench_compute_collatz runs at
AILang/C = 0.99x across both allocators. AILang's IR composes
with LLVM's optimizer at the same level a hand-C source does.
This is the LLVM-linkable performance claim, backed by data
for the first time. bench_compute_intsum (1.05-1.18x) confirms.
2. AILang bump beats glibc malloc 2x on linear allocation:
bench_list_sum.bump/c = 0.50x. Bump's two-instruction inline
fastpath outperforms glibc's free-list-managed malloc on
no-free workloads. Quantitatively measured for the first time.
3. RC overhead vs C malloc quantified: bench_list_sum.rc/c =
1.49x, bench_tree_walk.rc/c = 2.61x. The 8-byte refcount
header + zero-init + libc backing add 50-160% over glibc
malloc on these implicit-mode workloads. Explicit-mode + a
free()-adding C variant (21'f, queued) will close the
apples-to-apples gap on dec-cost.
CLAUDE.md updated to list bench/cross_lang.py as the third
tidy-iter gate alongside bench/check.py and bench/compile_check.py.
20 new metrics in bench/baseline_cross_lang.json with 12-15%
tolerances (cross-language ratios are inherently noisier than
within-AILang ratios — two compiler stacks contribute variance).
Closes the third corpus blind spot (heap-allocation-only) by
adding two fixtures with no allocation pressure: bench_compute_
intsum (tail-recursive integer accumulator) and bench_compute_
collatz (Collatz step-counter, branchy).
Surprise on intsum: 50M-iteration loop runs in 1ms wall under
all three allocators. LLVM's induction-variable analysis applies
the closed-form triangular-sum reduction to AILang's IR — a
positive codegen finding (the IR composes with LLVM's optimizer
at the same level a hand-C loop would) but it makes intsum
useless as a runtime regression bench. Excluded from run.sh's
fixtures array; kept in examples/ as reference and as a future
cross-language comparison anchor.
Collatz survives optimization (data-dependent control flow). At
56ms wall, gc/bump/rc all within 2% — the canonical "pure-compute
is allocator-invariant" data point this fixture is meant to
prove. If a future codegen change leaks an allocation into the
inner loop, the 1.00x / 1.02x ratios diverge visibly.
Two infrastructure fixes the new fixtures forced:
- 6-decimal precision in run.sh's Python timing helper and median
averager (was 3-decimal; sub-ms times rounded to 0.000 and
crashed the ratio awk with Division durch Null).
- Zero-guard in the ratio awk (defensive even with the precision
bump, since LLVM-eliminated workloads can still hit zero).
Latency baseline: implicit_at_rc.max_us tolerance 25% -> 30%.
Three captures today (477 / 456 / 609 µs) show natural run-to-run
dispersion wider than the original tolerance accounts for. Not a
softening to dodge regression — the original baseline was the
first capture; a fairer tolerance across natural max-of-1000-
samples width is what the harness needed from the start.
Baseline file: 47 -> 55 metrics. 21'e (cross-language reference,
clang -O2 hand-C ratios) is the natural next dispatch.
Closes the second axis the user named: every typechecker / codegen
perf change was previously invisible to the tidy-iter gate. With
Family 21 typeclasses (queued) and 21'b's poly-ADT additions both
pushing on the typechecker, naive substitution loops would have
landed silently and decayed the compile path.
bench/compile_check.py is a separate script from bench/check.py
because the methodology differs: sub-process spawn timing on small
workloads (1ms scale for `ail check`, 65ms for `ail build`) vs.
allocator-stress on large ones (multi-second). Tolerances differ
by an order of magnitude (25% / 20% here vs. 5-15% there).
Empirically: ail check is sub-ms across the corpus, dominated by
subprocess spawn (~5-10ms on Linux); ail build is 63-69ms,
dominated by clang's link step. The bench is a catastrophe
detector (10x slowdowns visible) — finer regressions need a
profiler. CLAUDE.md updated to list both scripts as co-equal
tidy-iter gates alongside the architect drift report; exit 0/1/2
semantics are uniform across both.
JOURNAL queue: 21'd (pure-compute fixtures) and 21'e (cross-
language reference / hand-C ratio) remain to land the LLVM-
linkable performance claim.
Two new throughput fixtures targeting blind spots in the 21'a
corpus:
- bench_closure_chain exercises the build_pair_drop_fn codegen
path (the 18c.4 doubled-braces trigger). Each iteration of
run_loop allocates a {thunk, env} closure pair via the
let-rec-name-as-value escape route. Sizes 10k / 100k / 500k.
rc/bump = 4.14x — materially worse than the 2.91x / 2.59x of
the linear/tree fixtures, exposing that closure work pays the
RC alloc tax twice (pair + env-struct).
- bench_hof_pipeline exercises poly-ADT instantiation and
indirect dispatch via fold_with_fn over List<a>. Sizes 100k /
1M / 3M elements. Ratios essentially match bench_list_sum,
confirming the 13b static-template-plus-ctor-inline design
has zero measurable overhead at this scale.
Baseline file extends from 31 to 47 metrics. The two new fixtures
build clean under all three allocators; the rc-arm build exercises
the per-type drop fn for the closure-pair, providing a tripwire
for any future 18c.4-class IR malformedness.
JOURNAL records both surprises (4.14x closure tax, ~zero HOF/poly
overhead) and explicitly notes the dispersion observation on
explicit_at_rc.p99 — three captures today (357.5 / 294.6 / 251.5)
confirm wide run-to-run variance on that fixture. Methodology
upgrade (n>=10 captures or tighter fixture) deferred to 21'c.
bench/run.sh fixtures array updated. bench/check.py needed no
changes — its parser handles the wider table by metric name.
Closes the structural gap between bench/run.sh (one-shot capture
into JOURNAL) and a continuous tripwire. baseline.json records 31
metrics (16 throughput, 15 latency) with per-metric one-sided
tolerances tuned to absorb run-to-run noise on a quiet developer
machine; check.py spawns run.sh, parses both the throughput
pipe-table and the per-arm latency stanzas, diffs against the
baseline, prints a per-metric report, and exits non-zero on any
regression beyond tolerance.
User-facing flags: --from-file, --stdin, --baseline, and
--update-baseline (re-run + overwrite baseline.json after
intentional improvements).
Validation: captured the baseline, then re-ran end-to-end. All 31
metrics within tolerance. The harness also caught a single-capture
explicit-rc p99 spike (357.5 us) that was first read as drift vs.
yesterday's 18g.tidy.fu2 numbers but came in at 294.6 us on the
follow-up run — exactly the kind of noise the tolerance band is
there to absorb. Without 21'a we would have either chased a
phantom or buried the signal; with it, single noisy runs are data
points, not verdicts.
Tidy-iter discipline addition (run check.py at every family close
alongside the architect drift report) is recommended in JOURNAL
but not enacted in this iter — that's an orchestrator-level
update to CLAUDE.md.
The dual-allocator policy from 2026-05-08 made Boehm the CLI default
"for general workloads", with rc selected explicitly for real-time-
sensitive code. That asymmetry contradicted Decision 10: rc + uniqueness
inference is the canonical memory model AILang's typechecker enforces,
and the CLI was teaching the opposite mental model to users and prompt-
fed LLMs.
This iter flips the asymmetry without removing Boehm:
- main.rs: default_value gc -> rc for build/run; help text rewritten
so rc is the canonical path and gc is the parity oracle.
- DESIGN.md Decision 9 retitled "RC canonical, Boehm parity oracle"
and rewritten end-to-end. Migration plan step 6 updated to a
two-step retirement (default-flip done, full removal gated).
- DESIGN.md pipeline diagram: rc-first, gc-as-oracle.
- JOURNAL: 2026-05-09 entry recording the call, what shipped,
the bug the flip surfaced, the gating condition for full
Boehm removal.
Bug surfaced by the flip and fixed in the same iter:
codegen/drop.rs build_pair_drop_fn emitted `getelementptr inbounds
{{ ptr, ptr }}, ...` via push_str (not format!), so the doubled
braces leaked verbatim into the IR. LLVM parsed it as a struct-of-
struct, the GEP referenced a non-existent field, and clang failed.
Invisible under the old gc default (no per-type drop fns) and
invisible to the 8 explicit _with_alloc("rc") tests (none of them
escaped a closure-with-captures). Caught immediately by two corpus
tests once rc became the baseline: closure_captures_let_n,
local_rec_as_value_capture_demo. Fix: single braces. Tests stay
as the regression guard.
The episode is the canonical justification for keeping the GC arm
as a parity oracle for now — a months-old codegen bug found by
flipping the baseline column.
288 passed / 0 failed / 3 ignored, unchanged.
Architect drift review after Family 20 surfaced four items; (1) and
(2) are doc drift fixed here, (3) and (4) are recorded as acceptable.
DESIGN.md changes:
- Data model section rewritten from "Data model (MVP)" claiming
only fn/const exist to "Data model" reflecting all three Def
kinds. Added the missing schema fields the architect named:
`tail` flag on app/do, the `seq` / `clone` / `reuse-as` / `letrec`
terms, `paramModes` / `retMode` on Type::Fn, `suppress` on FnDef,
`drop-iterative` on TypeDef, `args` on Type::Con, `vars` on
TypeDef. Each additive field's hash-stable serialisation is
called out once at the top so individual mentions stay terse.
`Suppress`, `ParamMode`, `Literal`, `Pattern` get their own
sub-blocks.
- Pipeline diagram now shows `--alloc=gc` (links libgc;
transitional) and `--alloc=rc` (emits ailang_rc_inc / _dec;
canonical) as two backends sharing the same MIR.
- Ecosystem inventory gained a "Surface forms" bullet for
ailang-surface (Iter 14c, lossless Form-A printer/parser) and
ailang-prose (Family 20, lossy Form-B projection). CLI bullet
now lists parse, render, prose, merge-prose.
JOURNAL entry records the architect findings, the resolutions for
items (1) and (2), and the acceptable-drift status of (3)
FnDef::synthetic factor-out (defer to next schema-additive
FnDef field) and (4) linearity.rs spurious warning (not
corpus-triggered).
Pure documentation iter — no code changes. 288 workspace tests
unchanged and green.
Closes a design hole shipped in 20d: the merge-prose prompt instructed
the LLM to emit JSON-AST and gave a 12-line schema-essentials reminder.
JSON-AST is the canonical hashable artefact, not a writing surface; the
reminder was not a language spec. Foreign LLMs had no realistic shot.
20f makes two coupled changes:
1) The LLM now emits Form-A (the canonical authoring surface fixed by
Decision 6). merge-prose loads the original via ailang_core::load_module
and re-renders via ailang_surface::print before embedding (round-trip
is a gating contract on the surface crate, so this is lossless). The
user runs `ail parse foo.new.ailx` to recover JSON, then `ail check`.
2) crates/ailang-core/specs/form_a.md is the complete LLM-targeted
Form-A specification — grammar, every term/pattern/type/def keyword,
schema invariants, pitfall catalogue, four few-shot modules from
examples/*.ailx. Exported as ailang_core::FORM_A_SPEC via include_str!
and embedded verbatim in every merge-prose prompt.
Drift detection in crates/ailang-core/tests/spec_drift.rs: every variant
of Term, Pattern, Type, Def, Literal is reached via exhaustive `match`.
The arms are not the assertion — adding a new variant without updating
the match is a compile error before the test runs. Once matched, an
anchor string is asserted to appear in FORM_A_SPEC. 8 tests, all green.
The hand-written + mechanical-drift-test combo addresses the user's
"distance to code is too big" concern about a docs-only spec. Generator
overkill rejected: structural drift is mechanically caught, but prose
explanation, schema-invariant catalogue, pitfall list, and few-shot
corpus cannot be emitted from AST shape alone.
Tests:
- ailang-core: +8 spec_drift tests; existing 12 unchanged
- ail unit (3): rewritten in lockstep — assert (own T)/(effects IO)
landmarks, FORM-A SPECIFICATION header, FORM_A_SPEC body verbatim
- ail e2e merge_prose_prints_framed_prompt: rewritten — assert
`(module foo` + `FORM-A SPECIFICATION` instead of `ailang/v0`
- Workspace: all green
Richer integration paths (LLM tool-use, MCP server, LSP) were named
in the design discussion and deferred per "kiss". All three layer
additively on the static-prompt path; static prompt remains the
lowest-common-denominator fallback.
Per CLAUDE.md tidy-iter at family boundaries. ailang-architect
flagged the canonical 'DESIGN.md silent' finding for the entire
19a-arc — same shape as family-20's 20e tidy.
Decision 10 schema-additions block gained 'Iter 19b —
FnDef.suppress' entry. New subsection 'Advisory diagnostics —
Iter 19a-arc' documents the over-strict-mode lint rule (incl.
heap-type filter rationale), Severity::Warning introduction, CLI
exit-on-Error gating, and the suppress mechanism. New subsection
'Why advisory + suppress instead of inference' pins the three
substantive reasons (annotation-states-intent, drift-bremse,
LLM-authoring forcing function) so future iters can't reopen the
decision without engaging the recorded rationale.
Migration plan extended with step 7 covering 19a/19a.1/19b.
Decision 10's mandatory-annotation rule is explicitly reaffirmed
as unchanged.
No code changes. Two architect findings deferred (codegen FnDef
fan-out, snapshot-fixture coupling). Data model (MVP) drift
predates 19b — separate iter.
Corpus-signal upgrade: orchestrator ran 19a's lint across 65
fixtures, zero warnings fired. The conservative match-scrutinee
carve-out filtered every (own T) fixture because every such body
destructures the param via match. Real-corpus signal warranted
the precise variant.
Precise rule: for an Own param with consume_count == 0, walk each
match arm whose scrutinee is the param; if any HEAP-TYPED
pattern-binder has consume_count > 0, stay silent (sub-consume
forces ownership). Otherwise fire.
Heap-type filter is the implementer's design call (not in the
brief): reading a primitive (h: Int) is load-by-value, no RC, no
heap-data move; reading heap-typed t: IntList IS a pointer move
requiring outer-cell ownership. Documented in module-doc.
Corpus signal after upgrade: 5/65 fixtures fire — all RC
codegen-test fixtures with deliberate over-strictness for the
codegen path. Justifies 19b (mode-strict-because suppression).
ailang-check test count: 53 → 55. e2e + everything else green.
Known debt: deeply-nested-match-on-sub-binder (no fixture hits).
Family-20 tidy-iter: ailang-architect drift review surfaced
DESIGN.md silence on the new prose surface as the highest-leverage
gap. Decision 6 now has a 'Form (B) — human prose projection'
subsection naming the lossy-vs-lossless contract, the no-parser-
by-design choice, and the LLM-mediated round-trip. CLI block
gained 'ail prose' and 'ail merge-prose'.
No code changes. Form (B) does not weaken any Decision 6
invariant: JSON-AST remains the only hashable artefact, form (A)
the canonical authoring surface, 30-production grammar unchanged.
Two architect findings deferred (prompt-template lockstep test;
snapshot-fixture cross-family coupling). Family 20 closes.
JOURNAL queue: empty.
Closes family 20: prose-edit cycle is end-to-end. ail merge-prose
<original.ail.json> <edited.prose.txt> composes a shaped LLM prompt
to stdout. User pipes to their LLM (Claude/etc.), gets new .ail.json,
runs ail check, saves.
No built-in API client — AILang stays a compiler + tooling, the LLM
mediator is supplied externally. PROSE_ROUNDTRIP.md documents the
6-step cycle, the prompt contract (preserve modes/effects/tail
flags from original), failure modes, and the design choice to omit
an API client.
Family 20 now closes: 20a renderer + 20b polish + 20d mediator
shipped; 20c rolled into 20a. Three deferred polishes (let-inlining,
print sugar, doc-wrap widow control) wait for corpus signal.
Brummel