Add the third opinion: an independent model, exhaustive, on 33.7M inputs - #2201
Merged
Conversation
#2199 shipped exhaustive cross-target checking and printed a limitation on every run: C against Rust only, so a fault shared by both backends -- a spec bug or shared front-end lowering -- was invisible. 'The backends agree' and 'the backends both match the specification' are different claims, and the first is easy to mistake for the second. tools/ternary_model.py is an independent transcription of the primitives, read from specs/ternary/ternary_ripple_adder.t27 and re-expressed in Python. It was NOT derived from the generated C or Rust; had it been, it would agree by construction and prove nothing. Transcribed faithfully, including what looks wrong: pack2 does not mask its arguments to two bits, so a value above 3 spills into the neighbouring trit lane. A model that 'fixed' that would report a divergence which is really the model disagreeing with the specification it exists to encode. full_adder(u8,u8,u8) 16,777,216 inputs model == C == Rust maj3(u8,u8,u8) 16,777,216 inputs model == C == Rust tmul(u8,u8) 65,536 inputs model == C == Rust pack2(u8,u8) 65,536 inputs model == C == Rust negate(u8) 256 inputs model == C == Rust 33.7 million inputs, three independent implementations, about eight seconds. Two negative controls, because the model arm needed its own. Perturbing C proves the C/Rust comparison has resolution; it says nothing about whether a model disagreeing with BOTH backends would be noticed, which is the entire reason the model exists. one-input perturbation of C 91a68892 -> b7b51485 one-input perturbation of MODEL 91a68892 -> cea51633, backends 91a68892 Measured rather than guessed: pure Python did not finish full_adder's 16.7M inputs in 600s -- nine dot27 calls per input, 27 lanes each, ~4e9 operations. Memoising dot27, a pure function whose semantics are untouched by it, took the rate from 179k calls/s to 2.2M/s. Targets added to TARGETS without a model entry stay two-way and are labelled [2-way: no model entry], so the weaker case cannot pass itself off as the stronger. Closes #2200 Refs #2198 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Contributor
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
Contributor
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
#2199 shipped exhaustive cross-target checking and printed a limitation on every run: it compared C against Rust only, so a fault shared by both backends — a spec bug, or shared front-end lowering — was invisible. "The backends agree" and "the backends both match the specification" are different claims, and the first is easy to mistake for the second.
This closes that gap.
tools/ternary_model.pyis an independent transcription of the primitives, read fromspecs/ternary/ternary_ripple_adder.t27and re-expressed in Python. It was not derived from the generated C or Rust — had it been, it would agree by construction and prove nothing.Transcribed faithfully, including the parts that look wrong:
pack2does not mask its arguments to two bits, so a value above 3 spills into the neighbouring trit lane. A model that "fixed" that would report a divergence which is really the model disagreeing with the specification it exists to encode.Result — three-way agreement, exhaustive:
full_adder(u8,u8,u8)maj3(u8,u8,u8)tmul(u8,u8)pack2(u8,u8)negate(u8)33.7 million inputs, three independent implementations, ~8 seconds total.
Two negative controls, because the model arm needed its own. Perturbing C proves the C/Rust comparison has resolution; it says nothing about whether a model disagreeing with both backends would be noticed — which is the entire reason the model exists. So there are now two:
A measurement that changed the approach. Pure Python over
full_adder's 16.7M inputs did not finish in 600 s — it callsdot27nine times per input, 27 lanes each, about 4 billion operations. Memoisingdot27(a pure function, so semantics are untouched) takes it to under a second. Measured rather than guessed: 179k calls/s uncached, 2.2M/s cached.Anything added to
TARGETSwithout a model entry is still checked two-way and labelled[2-way: no model entry], so the weaker case cannot pass itself off as the stronger one.Refs #2198
Closes #2200
🤖 Generated with Claude Code