SMC-001 RD1: freeze adaptive-depth protocol and establish P0–P2 apparatus #37

Closed
opened 2026-09-21 19:48:39 +00:00 by pyrrha · 0 comments
Member

Operator direction

Cassie approved the broad RD1 structure on 2026-09-21 and authorized work through P0, P1, and P2. P3 remains the first scientific kill gate. Learned halting is not authorized unless P3 establishes meaningful oracle headroom.

Scope

  • Canonicalize the SMC-001 / RD1 protocol, including the final corrections: difficulty is evaluation metadata only; P3 separates independently trained fixed-depth frontiers from one recurrent model evaluated at truncation depths; oracle success uses a preregistered stability criterion rather than first-correct; at least one state-dependent sequential family such as pointer chasing; post-hoc matched-histogram random-halting control first; same examples/tokens and optimizer schedule as primary training comparison.
  • Independently verify the adaptive/recurrent-depth literature ledger before canonizing positioning.
  • P0: pinned environment and honest hardware/dispatch characterization.
  • P1: minimum resource-accounting harness with independent/golden validation and per-example records.
  • P2: bounded training smoke proving finite training state, checkpoint save/reload, resumed-step equivalence, and accounting integration.

Hard boundaries

  • No paid compute or rental.
  • No substantial training run.
  • No learned halting implementation or P3 scientific run in this task.
  • Preserve algorithmic FLOP savings and realized hardware latency as separate claim layers.
  • Public repository: no credentials, private host paths, raw predictions, model weights, or datasets.

Ownership

  • Pyrrha: protocol, literature/reproducibility, integration, PR.
  • Eido: P0 compute/environment evidence and feasibility review.
  • Vesper: baseline/control critique and adversarial review.

Acceptance

  • Exact source/version/status/claim/scale/comparison/limitation ledger for required lineage.
  • Protocol and phase gates agree with operator decisions above.
  • P0 evidence is bounded and labeled by measurement basis.
  • P1/P2 tests demonstrate the claimed behavior; full repository gates pass.
  • Different Forgejo identity reviews the exact PR head before merge.
## Operator direction Cassie approved the broad RD1 structure on 2026-09-21 and authorized work through P0, P1, and P2. P3 remains the first scientific kill gate. Learned halting is not authorized unless P3 establishes meaningful oracle headroom. ## Scope - Canonicalize the SMC-001 / RD1 protocol, including the final corrections: difficulty is evaluation metadata only; P3 separates independently trained fixed-depth frontiers from one recurrent model evaluated at truncation depths; oracle success uses a preregistered stability criterion rather than first-correct; at least one state-dependent sequential family such as pointer chasing; post-hoc matched-histogram random-halting control first; same examples/tokens and optimizer schedule as primary training comparison. - Independently verify the adaptive/recurrent-depth literature ledger before canonizing positioning. - P0: pinned environment and honest hardware/dispatch characterization. - P1: minimum resource-accounting harness with independent/golden validation and per-example records. - P2: bounded training smoke proving finite training state, checkpoint save/reload, resumed-step equivalence, and accounting integration. ## Hard boundaries - No paid compute or rental. - No substantial training run. - No learned halting implementation or P3 scientific run in this task. - Preserve algorithmic FLOP savings and realized hardware latency as separate claim layers. - Public repository: no credentials, private host paths, raw predictions, model weights, or datasets. ## Ownership - Pyrrha: protocol, literature/reproducibility, integration, PR. - Eido: P0 compute/environment evidence and feasibility review. - Vesper: baseline/control critique and adversarial review. ## Acceptance - Exact source/version/status/claim/scale/comparison/limitation ledger for required lineage. - Protocol and phase gates agree with operator decisions above. - P0 evidence is bounded and labeled by measurement basis. - P1/P2 tests demonstrate the claimed behavior; full repository gates pass. - Different Forgejo identity reviews the exact PR head before merge.
Sign in to join this conversation.
No labels
R03
research-b
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
smolmodelco/thesmolmodelcompany#37
No description provided.