Skip to content
Draft
Show file tree
Hide file tree
Changes from all commits
Commits
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
352 changes: 176 additions & 176 deletions REFERENCE.md

Large diffs are not rendered by default.

2 changes: 2 additions & 0 deletions experiments/qwen35_acceptance/.gitattributes
Original file line number Diff line number Diff line change
@@ -0,0 +1,2 @@
# Preserve terminal output byte-for-byte without treating progress padding as errors.
results/**/result/*.log -whitespace
355 changes: 355 additions & 0 deletions experiments/qwen35_acceptance/PROPOSAL.md

Large diffs are not rendered by default.

23 changes: 20 additions & 3 deletions experiments/qwen35_acceptance/README.md
Original file line number Diff line number Diff line change
@@ -1,8 +1,24 @@
# Qwen3.5 acceptance experiment

The committee-facing design and review questions are in
[`PROPOSAL.md`](PROPOSAL.md). Both authorized runs are preserved as execution
failures. V1 exposed an interpreter-path bug before scientific execution. V2
fixed that problem and completed both 4,097-token evaluations, but the emitted
checkpoint failed the mandatory official Transformers reload because the two
prepended layers' tensor names and MLP shapes do not match the saved Qwen
architecture. The terminal v2 evidence is in
[`results/v2-20c3330d0b3e-execution-failure/`](results/v2-20c3330d0b3e-execution-failure/).
The permanent evidence index is
[`f35OgXu…ARZTU`](https://arweave.net/f35OgXuOoulk9EcndJBNWIgZ5tmaQ6DRzeINP7ARZTU/).

The layer-prepending installer therefore remains experimental. No v3 run is
authorized. The model revision, corpora, thresholds, seed, and spending ceiling
for the completed v2 attempt remain frozen in
[`launch-manifest.json`](launch-manifest.json).

This directory turns the open Qwen integration questions into a frozen ilxyr
project for one owner-authorized, bounded attempt. It does not claim that the
full run has passed. It generates the
project and preserves its bounded executions. It does not claim that the full
run passed. It generates the
hypothesis, methodology contributions, experiment contract, two explicit model
forecasts, funding record, and ordered ilxyr commands for one real checkpoint.

Expand All @@ -22,7 +38,8 @@ python experiments/qwen35_acceptance/generate.py \
/absolute/path/to/Qwen3.5-0.8B \
/absolute/path/to/installation-corpus.txt \
/absolute/path/to/evaluation-corpus.txt \
/absolute/path/to/ilxyr-project
/absolute/path/to/ilxyr-project \
--experiment-version 2
```

The two corpora must have different contents.
Expand Down
127 changes: 127 additions & 0 deletions experiments/qwen35_acceptance/launch-manifest.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,127 @@
{
"schema": "lecore.qwen35-launch.v1",
"state": "terminal_execution_failure",
"attempt": {
"experiment_version": 2,
"maximum_formal_runs": 1,
"retry_after_admission": false,
"seed": 0,
"authorized_at": "2026-08-13T00:09:30Z",
"predecessor": {
"experiment_id": "lecore.qwen35.install.20c3330d0b3e.d6905f043e78.a6c9d1135a04.a04ab5692be3.v1.acceptance",
"outcome": "execution_failure",
"scientific_result": false,
"corrective_change": "preserve the selected virtual-environment interpreter path"
}
},
"source": {
"repository": "https://github.com/atimics/holostuff.git",
"commit": "2f32151c6d44c80a7e2ff98c0696650cd1d9c6f1"
},
"ilxyr": {
"repository": "https://github.com/cenetex/ilXyr.git",
"commit": "e92382ff2a5e8714466533a160f6609b4ef9cee8"
},
"model": {
"repository": "Qwen/Qwen3.5-0.8B",
"revision": "2fc06364715b967f1860aea9cf38778875588b17",
"license": "Apache-2.0",
"critical_files": [
{
"path": "model.safetensors-00001-of-00001.safetensors",
"bytes": 1746942600,
"sha256": "04b1c301231dd422b8860db31311ab2721511346a32cb1e079c4c4e5f1fe4696"
},
{
"path": "tokenizer.json",
"bytes": 12807982,
"sha256": "5f9e4d4901a92b997e463c1f46055088b6cca5ca61a6522d1b9f64c4bb81cb42"
}
],
"complete_manifest": "generated and frozen before ilxyr admission"
},
"corpora": {
"installation": {
"handle": "git:a04ab5692be38f06120aba4b0bc5e2a284eb2c79:REFERENCE.md",
"license": "MIT",
"bytes": 2300089,
"sha256": "d6905f043e7856b93b2dd72dac5fa0dc593898c55d6c54c51f3153f4317d6b7f"
},
"evaluation": {
"handle": "https://www.gutenberg.org/ebooks/18",
"retrieved_at": "2026-08-12T21:32:31Z",
"last_modified": "2026-08-01T07:33:54Z",
"copyright": "Public domain in the USA",
"bytes": 1213410,
"sha256": "a6c9d1135a04d10955fe11d210b7f642e1c2341d4f2c8369b9a832cc97839d94"
}
},
"method": {
"minimum_paired_positions": 4096,
"reference_logit_relative_tolerance": 0.001,
"maximum_regression": 0.01,
"confidence": 0.95,
"bootstrap": "paired moving block",
"spectral_filtering": false,
"installer_experimental": true
},
"dependencies": {
"python": "3.12",
"numpy": "2.4.6",
"pillow": "12.3.0",
"torch": "2.11.0+cpu",
"transformers": "5.14.0",
"resolved_environment": "recorded before ilxyr admission"
},
"aws": {
"region": "us-east-1",
"availability_zone": "us-east-1a",
"ami": "ami-07a5b367e8dc8bd92",
"ami_name": "al2023-ami-2023.12.20260803.3-kernel-6.1-x86_64",
"instance_type": "r7i.4xlarge",
"vcpu": 16,
"memory_gib": 128,
"market": "on-demand",
"hourly_compute_usd": 1.0584,
"root_volume": {
"type": "gp3",
"size_gib": 100,
"encrypted": true,
"delete_on_termination": true
},
"inbound_ports": [],
"experiment_timeout_seconds": 21600,
"instance_lifetime_seconds": 28800,
"instance_initiated_shutdown_behavior": "terminate",
"maximum_compute_usd": 8.4672,
"total_ceiling_usd": 10.0
},
"publication": {
"temporary_staging": "s3://zero-training-022118847419/qwen35-acceptance/2f32151-2fc06364-a6c9d113/v2-formal/",
"review_index": "https://github.com/atimics/holostuff/pull/3",
"permanent_transport": "Arweave",
"permanent_root": "https://arweave.net/f35OgXuOoulk9EcndJBNWIgZ5tmaQ6DRzeINP7ARZTU/",
"permanent_root_transaction": "f35OgXuOoulk9EcndJBNWIgZ5tmaQ6DRzeINP7ARZTU",
"publication_manifest_sha256": "21874555c7f9e73d20b9d11599c79236f8f40267ee975f703ffc053eb6b1f9ab",
"publish_all_outcomes": true,
"exclude_model_weights": true
},
"terminal": {
"completed_at": "2026-08-13T04:51:28.293Z",
"formal_duration_seconds": 15891.556,
"runner_exit_code": 1,
"timed_out": false,
"resolved_outcome": "execution_failure",
"scientific_result": false,
"failure_gate": "official_transformers_checkpoint_reload",
"aws_instance": "i-059aab8bf1d3d87c9",
"estimated_compute_usd": 4.71,
"evidence_bundle_sha256": "df2ba31608f9039e9528a43a93d4766c8e8feb023668a1b9ffc1c85553210c1c",
"ledger_verification": {
"valid": true,
"objects_checked": 15,
"events_checked": 16
},
"retry_authorized": false
}
}
Original file line number Diff line number Diff line change
@@ -0,0 +1,57 @@
# Qwen3.5 acceptance experiment: v1 terminal record

- Experiment: `lecore.qwen35.install.20c3330d0b3e.d6905f043e78.a6c9d1135a04.a04ab5692be3.v1.acceptance`
- ilxyr outcome: `execution_failure`
- AWS instance: `i-0c549643b970e896a`
- Formal execution: started, exited `1`, did not time out
- Ledger verification: valid (`15` objects and `16` events)
- Scientific conclusion: none

## What happened

The corrected AWS preflight verified the frozen source, model, tokenizer, corpora,
minimum-token contract, timeout, and disabled spectral path. ilxyr then accepted
the project after all twelve admission gates passed and started the formal run.

The runner stopped after 51 milliseconds while importing NumPy:

```text
ModuleNotFoundError: No module named 'numpy'
```

The generator had resolved the virtual environment's `python` symlink to the
system interpreter. Dependencies were installed in the virtual environment,
but ilxyr was instructed to execute `/usr/bin/python3.12`. Commit `0bda4f1`
preserves the virtual-environment interpreter path and adds a regression test.

This terminal record does **not** contain reference-logit, installation,
bootstrap, memory, checkpoint-reload, text-generation, or vision-smoke metrics.
It cannot support an acceptance or rejection claim about Qwen or the
layer-prepending installer.

## Evidence map

- `project/` contains the exact ilxyr project admitted on AWS.
- `result/` contains preflight, runtime, status, and environment records. The
redistributable corpora are omitted from Git to keep the review small; their
exact hashes remain in `corpus-manifest.json` and `launch-manifest.json`.
- `ledger/` is the verified `.ilxyr` event/object store with the directory
renamed for Git publication.
- `evidence/` contains native ilxyr, RO-Crate, and in-toto exports.
- `preflight-001.json` records the earlier non-formal launch that rejected a
mistyped, nonexistent source commit before ilxyr execution.
- `publication-manifest.json` is a canonical hash-and-size index for this
directory and is generated by `../../build_publication.py`.

The original AWS evidence bundle for this formal attempt is retained at
`s3://zero-training-022118847419/qwen35-acceptance/a04ab56-2fc06364-a6c9d113/attempt-002/qwen35-evidence.tgz`
with SHA-256
`6182c0fd9646644d964e3d7b62a0ef46db8653b5b1e94901f03f22ea56386e11`.
The earlier preflight-only bundle has SHA-256
`836fd6b29fa63ca6a636a3f9b3f1e66b7928605e0166e868e871a8797357e76a`.

## Retry policy

The v1 contract prohibited an automatic retry under the same identity. A
corrected run therefore needs explicit authorization and a new `v2` experiment
identity. This record remains published regardless of whether v2 is run.
Loading