Repository navigation
feat(tri): emit determinism -- name the specs whose output wobbles, not just how many - #3007
Merged
Merged
Conversation
#3006 measured that three of four emitters are not byte-deterministic and said the cheapest next step is naming the differing files rather than counting them. This is that command. Measured over 650 specs, same binary, two runs: C gen-c differing: 4 specs/igla/coder/_tmp_pipeline_import.t27 specs/igla/coder/pipeline.t27 specs/physics/sacred_verification.t27 specs/physics/zamolodchikov_4d_conjecture.t27 Rust gen-rust differing: 2 (pipeline, sacred_verification) Zig gen differing: 4 (the same four as C) Verilog gen-verilog differing: 0 The names change the finding. Two specs wobble in all three non-Verilog backends and two more in two of them -- four specs in total, largely the same ones. That points at something shared rather than at three independent emitter defects, which is what the counts alone suggested. The counts are themselves unstable: an earlier ad-hoc harness read 1 / 3 / 2 where this reads 4 / 2 / 4. Which files wobble is a draw, so only the union is meaningful. Three refusals, each naming what it refused: `--runs 1` ("one run cannot disagree with itself"), an unknown backend, and a missing compiler -- the last because a determinism report of zero taken with no binary reads exactly like health. It reports and never gates. Refs #3006
Contributor
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
Contributor
Contributor
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
Contributor
PR DashboardGenerated at: 2026-09-03 07:54:01 UTC
Summary
Seal Status
|
Contributor
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
Contributor
PR DashboardGenerated at: 2026-09-03 08:27:05 UTC
Summary
Seal Status
|
Contributor
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
Contributor
PR DashboardGenerated at: 2026-09-03 08:35:47 UTC
Summary
Seal Status
|
This was referenced Sep 3, 2026
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.
#3006 measured that three of four emitters are not byte-deterministic and named
its own next step: "the cheapest next step is naming the files: run the pair
again and print the differing basenames rather than the count." This is that
command.
What it found, and why the names matter more than the counts
Four specs, largely the same ones.
pipeline.t27andsacred_verification.t27wobble in all three non-Verilog backends;_tmp_pipeline_import.t27andzamolodchikov_4d_conjecture.t27in two of them.That points at something shared — a path all three emitters route through —
rather than at three independent emitter defects, which is what the counts alone
suggested.
And the counts are not stable. The ad-hoc harness that produced #3006 read
1 / 3 / 2 where this reads 4 / 2 / 4. Which files wobble is itself a draw, so a
count taken once is not a measurement of the population; only the union is. That
is the argument for naming.
Three refusals, each naming what it refused
Honesty
the_four_emitters_are_all_listed: a backend missing from the table is abackend this command silently does not measure — the shape
tri mods orphanhad when it watched two crates of five, and the shape gen-verilog: break lowers to
disable fork;and continue to a comment -- 17 no-ops in 8 files, and noforkexists anywhere in the corpus #2988's guard had.skipped, so a spec that generates on one run and refuses on the next still
shows up as a difference instead of vanishing from both sides.
emitand notgen:Genis already the single-spec emitter.fixing or an ordering nothing downstream reads is not a walker's question, and
a gate that reddens on it before anyone has decided gets muted.
iteration order (an earlier finding saw
gen-verilog-for-simulationvarying inthe order of two adjacent
reg _t27_call_tmp_*declarations); whether thesefour are the same cause is unmeasured, and the command does not guess.
Refs #3006
🤖 Generated with Claude Code