Swordnael commited on
Commit
0117235
·
verified ·
1 Parent(s): baba559

Document structural smoke boundary

Browse files
Files changed (1) hide show
  1. README.md +8 -0
README.md CHANGED
@@ -19,6 +19,8 @@ This public repository documents and reproduces the router-only curriculum for a
19
  - `fable-router-curriculum.v2.json`: immutable training/data policy.
20
  - `build_fable_router_curriculum.py`: deterministic local builder.
21
  - `REPRODUCIBILITY_MANIFEST.json`: pinned sources, input/output hashes, row counts, exclusions, and build command.
 
 
22
 
23
  Given independently authorized access to the two public sources, these files reproduce the exact materialized artifacts by SHA-256. They also expose the split, deduplication, overflow, anti-loop, duplicate-tool, benefit-gated routing, and anti-collapse decisions for review.
24
 
@@ -53,3 +55,9 @@ python scripts/bench/build_fable_router_curriculum.py
53
  The builder rejects malformed or overlong examples, deduplicates reasoning prompts, samples cumulative agent prefixes by immutable trajectory, creates train-only anti-loop preference pairs, and writes a provenance-bound `result.json`. Final training must retokenize with the actual chat template and reject overflow above 2,048 tokens rather than truncate it.
54
 
55
  Publication of this package does not authorize training, routing, deployment, or promotion in Autonoma.
 
 
 
 
 
 
 
19
  - `fable-router-curriculum.v2.json`: immutable training/data policy.
20
  - `build_fable_router_curriculum.py`: deterministic local builder.
21
  - `REPRODUCIBILITY_MANIFEST.json`: pinned sources, input/output hashes, row counts, exclusions, and build command.
22
+ - `training/`: the non-routing router shell, pure provenance validators, and free-GPU structural-smoke runner.
23
+ - `fable-router-cloud-smoke.v1.json`: immutable host, bank, curriculum, hash, hardware, and evidence gates for that runner.
24
 
25
  Given independently authorized access to the two public sources, these files reproduce the exact materialized artifacts by SHA-256. They also expose the split, deduplication, overflow, anti-loop, duplicate-tool, benefit-gated routing, and anti-collapse decisions for review.
26
 
 
55
  The builder rejects malformed or overlong examples, deduplicates reasoning prompts, samples cumulative agent prefixes by immutable trajectory, creates train-only anti-loop preference pairs, and writes a provenance-bound `result.json`. Final training must retokenize with the actual chat template and reject overflow above 2,048 tokens rather than truncate it.
56
 
57
  Publication of this package does not authorize training, routing, deployment, or promotion in Autonoma.
58
+
59
+ ## Structural smoke boundary
60
+
61
+ The structural smoke validates every one of a bank's 2,880 tensor records, then materializes only layer 29's 32 experts around the real frozen Fable host. It checks exact host-disabled identity, strict trainable-parameter isolation, nonzero router/scale gradients, explicit host-only and positive-benefit counterfactual targets, router checkpoint parity, and persisted resource telemetry. This is intentionally small enough for a free 16 GiB Colab/Kaggle GPU.
62
+
63
+ Passing it proves that a donor bank is mechanically trainable. It does **not** prove full-30-layer memory fit, final quality, throughput, or promotion fitness, and it cannot authorize a training campaign. Every run is uploaded to a private evidence dataset so an ephemeral notebook cannot turn an incomplete or failed result into a pass.