Skip to content

test(ci): gate that the generated trainer LEARNS -- XOR 4/4 + nonlinear held-out >=90%, incl. deep 3-layer (Refs #1764) #1909

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
gHashTag merged 1 commit into from
Aug 8, 2026
Merged

test(ci): gate that the generated trainer LEARNS -- XOR 4/4 + nonlinear held-out >=90%, incl. deep 3-layer (Refs #1764) #1909

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 3 additions & 0 deletions .github/workflows/emit-bitexact-gate.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -57,6 +57,9 @@ jobs:
- name: Prove the WHOLE trainer bit-exact in C (== model == Verilog)
run: python3 tools/verify_trainer_c.py

- name: Prove the trainer LEARNS (XOR 4/4 + nonlinear held-out >=90%, incl. deep 3-layer)
run: python3 tools/gft_backprop_microcode.py

- name: Differential-fuzz the trainer (random topologies, edge inputs)
run: python3 tools/fuzz_trainer.py 40

Expand Down
8 changes: 7 additions & 1 deletion docs/NOW.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,7 +1,13 @@
# NOW — Vivado closure kit: the multicycle constraint that kills the seed-lottery (2026-08-08)
# NOW — CI now gates that the trainer LEARNS, not just that it is bit-exact (2026-08-08)

Last updated: 2026-08-08

## test(ci): gate that the generated trainer LEARNS -- XOR 4/4 + nonlinear held-out >=90%, incl. deep 3-layer (Refs #1764)

- The bit-exact gates prove the generated RTL == model, but nothing CI-enforced that the algorithm actually LEARNS non-trivial tasks. The generator's self-tests already do (XOR 4/4; (2,4,1) noisy nonlinear held-out 58/60; (2,4,2) multi-output argmax 56/60; deep [2,4,3,1] 3-layer 59/60 -- all >=90%), but they only ran when the file was executed by hand
- Added `python3 tools/gft_backprop_microcode.py` as a CI step in emit-bitexact-gate.yml, so the learning + generalization claims are enforced per pull request
- Strengthens the whole story's thesis: the METHOD learns and scales beyond XOR (multi-layer, held-out generalization) in the model -- the only limit is the open silicon flow's placement marginality (now met with the Vivado closure kit), not the algorithm. CI-config + no code change. Refs #1764

## feat(docs): Vivado closure kit -- the set_multicycle_path constraint the open flow cannot express (Refs #1764)

- The root-cause arc concluded the seed-lottery is a global placement effect of the timing-relaxed open-flow placement; the cure is not another register but telling a timing-driven P&R the truth about the path. nextpnr-xilinx XDC = create_clock only, so it cannot; Vivado can
Expand Down
Loading