Skip to content

Charter v0.2: two-track structure for bigger models and J-lens (issue #1) - #2

Merged
horner merged 1 commit into
mainfrom
plan/issue-1-bigger-models-jlens
Aug 10, 2026
Merged

Charter v0.2: two-track structure for bigger models and J-lens (issue #1)#2
horner merged 1 commit into
mainfrom
plan/issue-1-bigger-models-jlens

Conversation

@horner

@horner horner commented Aug 10, 2026

Copy link
Copy Markdown
Contributor

Closes #1

Amends the research charter to v0.2 per the decisions in the issue #1 discussion: adopt Track B, fold Pythia-410M into the authorized repeat, and make Pythia-1B/1.4B conditional on the 410M trend.

Changes

Two-track structure (new §1.2)

  • Track A — interoception ladder: unchanged Pythia program, Milestones 0–11 / G00–G11.
  • Track B — J-space connection: new Milestones B1–B2 / gates GB1–GB2 on the Gemma family. B1 = behavioral prerequisite on Gemma 2 2B + lens-machinery validation against the published solarkyle/jspace Gemma-4-12B artifacts. B2 = sparse J-decomposition of the mobilization/body_state mediator + the automatic-versus-flexible interoception battery, with mandatory shuffled-lens / rotated-dictionary / norm-matched-random controls and J-clamping mediation tests. Passive J-lens readout is never gate-eligible evidence.

Model ladder (§5)

  • Pythia-410M folded into EXP-003-v2 as a third scale point (dated amendment under G02; thresholds committed before its held-out split is scored; GATE-02.md remains frozen).
  • Pythia-1B/1.4B registered as a conditional extension under Milestone 9, only if the 410M trend is promising (EXP-017).
  • Track B model table: Gemma 2 2B (EmotionScope anchor) and Gemma-4-12B (fitted-lens anchor, cluster-only).

Reproducibility contract for Gilbreth (§4, new §5.2)

  • CUDA as a third parity-gated backend; CPU float32 reference moves to cluster CPU nodes for models exceeding 16 GB.
  • Run manifests gain gpu, cuda, driver, node_type, slurm_job_id, container_digest (Apptainer required on CentOS 7).
  • sbatch scripts must be thin wrappers around the same versioned CLI used locally.
  • Cluster scratch/Depot added to the storage policy for fitted lenses and large Track B artifacts.

Vocabulary

  • j_space added to docs/glossary.md as "verbalization-poised subspace" with usage requirements; "global workspace" explicitly excluded from accepted vocabulary.

Included documents

  • docs/j-space-global-workspace.md (method review that motivated the amendment) and its source PDF.

Reviewer notes

  • One judgment call to review: Milestone B2 begins after GB1 and does not hard-depend on Track A's G05, because Track B extracts its own directions on Gemma (EmotionScope-style) rather than reusing Pythia probe directions. This was the unresolved fork 2 from the issue discussion — flag if you want the stricter G05 prerequisite.
  • Track B opens only after G02 closes, so EXP-003-v2 remains the sole authorized next work.
  • make check and make test pass (44 passed, 2 deselected).

)

Track A (Pythia interoception ladder) unchanged; Pythia-410M folded into
the authorized EXP-003-v2 repeat as a third scale point, with a conditional
Pythia-1B/1.4B extension registered under Milestone 9.

Track B (Gemma J-space connection) added as Milestones B1-B2 / gates
GB1-GB2: behavioral prerequisite and lens-machinery validation on
Gemma 2 2B against solarkyle/jspace Gemma-4-12B artifacts, then sparse
J-decomposition of the mobilization/body_state mediator and the
automatic-versus-flexible interoception battery. Passive J-lens readout
is never gate-eligible evidence.

Reproducibility contract extended for Gilbreth: CUDA as a parity-gated
backend, Apptainer image digests and Slurm fields in run manifests,
sbatch scripts as thin CLI wrappers, cluster scratch in the storage
policy. j_space added to the glossary as verbalization-poised subspace;
'global workspace' remains outside accepted vocabulary.

Includes the J-space method review (markdown + source PDF) that motivated
the amendment.
@horner
horner merged commit f903f20 into main Aug 10, 2026
4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Bigger models and the use of J-lens

1 participant