{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "$id": "urn:pythia-paths:schema:trajectory-evidence:0.2",
  "title": "Pythia Paths trajectory evidence",
  "type": "object",
  "additionalProperties": false,
  "required": [
    "schema",
    "evidence_id",
    "source_reviewed_at",
    "subject",
    "artifact",
    "metric_observation",
    "training_source",
    "selection",
    "claims",
    "authority",
    "unknowns"
  ],
  "properties": {
    "schema": { "const": "pythia-paths/trajectory-evidence/0.2" },
    "evidence_id": { "type": "string", "minLength": 1, "maxLength": 160 },
    "source_reviewed_at": { "type": "string", "format": "date-time" },
    "subject": {
      "type": "object",
      "additionalProperties": false,
      "required": [
        "provider",
        "repo_type",
        "repo_id",
        "declared_ref",
        "observed_branch_target_commit",
        "branch_target_observed_at",
        "step",
        "tokens_seen"
      ],
      "properties": {
        "provider": { "const": "huggingface" },
        "repo_type": { "const": "model" },
        "repo_id": { "const": "EleutherAI/pythia-70m-deduped" },
        "declared_ref": { "type": "string", "pattern": "^step[0-9]+$" },
        "observed_branch_target_commit": { "type": "string", "pattern": "^[0-9a-f]{40}$" },
        "branch_target_observed_at": { "type": "string", "format": "date-time" },
        "step": { "type": "integer", "minimum": 0 },
        "tokens_seen": {
          "type": "object",
          "additionalProperties": false,
          "required": ["value", "status", "method", "source_repo", "source_commit", "source_path"],
          "properties": {
            "value": { "type": "integer", "minimum": 0 },
            "status": { "const": "derived" },
            "method": { "const": "step * 2097152" },
            "source_repo": { "const": "EleutherAI/pythia-70m-deduped" },
            "source_commit": { "type": "string", "pattern": "^[0-9a-f]{40}$" },
            "source_path": { "const": "README.md" }
          }
        }
      }
    },
    "artifact": {
      "type": "object",
      "additionalProperties": false,
      "required": [
        "path",
        "storage",
        "bytes",
        "reported_lfs_oid",
        "retrieval_status",
        "execution_status",
        "format_equivalence",
        "relationship_to_evaluation"
      ],
      "properties": {
        "path": { "const": "model.safetensors" },
        "storage": { "const": "lfs" },
        "bytes": { "type": "integer", "minimum": 1 },
        "reported_lfs_oid": { "type": "string", "pattern": "^[0-9a-f]{64}$" },
        "retrieval_status": { "const": "metadata_only" },
        "execution_status": { "const": "not_run" },
        "format_equivalence": { "const": "not_tested" },
        "relationship_to_evaluation": { "const": "not_proven" }
      }
    },
    "metric_observation": {
      "type": "object",
      "additionalProperties": false,
      "required": [
        "status",
        "task",
        "metric",
        "direction",
        "value",
        "stderr",
        "task_version",
        "fewshot",
        "sample_count",
        "harness_revision",
        "comparison_status",
        "source_repo",
        "source_commit",
        "source_path",
        "source_blob_sha1",
        "evaluated_model_commit",
        "evaluated_artifact_digest",
        "metric_to_observed_branch_artifact_binding",
        "reported_run_config"
      ],
      "properties": {
        "status": { "const": "reported_upstream" },
        "task": { "const": "piqa" },
        "metric": { "const": "acc_norm" },
        "direction": { "const": "higher_is_conventionally_better" },
        "value": { "type": "number", "minimum": 0, "maximum": 1 },
        "stderr": { "type": "number", "minimum": 0 },
        "task_version": { "const": 0 },
        "fewshot": { "const": 0 },
        "sample_count": { "type": "null" },
        "harness_revision": { "type": "null" },
        "comparison_status": { "const": "not_proven" },
        "source_repo": { "const": "EleutherAI/pythia" },
        "source_commit": { "type": "string", "pattern": "^[0-9a-f]{40}$" },
        "source_path": { "type": "string", "minLength": 1 },
        "source_blob_sha1": { "type": "string", "pattern": "^[0-9a-f]{40}$" },
        "evaluated_model_commit": { "type": "null" },
        "evaluated_artifact_digest": { "type": "null" },
        "metric_to_observed_branch_artifact_binding": { "const": "not_proven" },
        "reported_run_config": {
          "type": "object",
          "additionalProperties": false,
          "required": [
            "model",
            "model_args",
            "batch_size",
            "device",
            "no_cache",
            "limit",
            "bootstrap_iters"
          ],
          "properties": {
            "model": { "const": "hf-causal" },
            "model_args": { "type": "string" },
            "batch_size": { "type": "integer", "minimum": 1 },
            "device": { "type": ["string", "null"] },
            "no_cache": { "const": true },
            "limit": { "type": "null" },
            "bootstrap_iters": { "const": 100000 }
          }
        }
      }
    },
    "training_source": {
      "type": "object",
      "additionalProperties": false,
      "required": ["repo", "commit", "config_path", "config_blob_sha1"],
      "properties": {
        "repo": { "const": "EleutherAI/pythia" },
        "commit": { "type": "string", "pattern": "^[0-9a-f]{40}$" },
        "config_path": { "const": "models/70M/pythia-70m-deduped.yml" },
        "config_blob_sha1": { "type": "string", "pattern": "^[0-9a-f]{40}$" }
      }
    },
    "selection": {
      "type": "object",
      "additionalProperties": false,
      "required": [
        "model_checkpoint_frame_count",
        "published_evaluation_report_frame_count",
        "visualized_report_count",
        "visualization_scope",
        "detailed_receipt_steps",
        "published_report_without_detailed_receipt_count",
        "checkpoint_without_published_report_count",
        "detail_selection_method",
        "detail_selected_before_outcomes_seen",
        "detail_representativeness",
        "detail_selection_reason",
        "task_sampling_frame",
        "included_task_metric",
        "omitted_task_count",
        "task_selection_method",
        "task_selected_before_outcomes_seen"
      ],
      "properties": {
        "model_checkpoint_frame_count": { "const": 154 },
        "published_evaluation_report_frame_count": { "const": 27 },
        "visualized_report_count": { "const": 27 },
        "visualization_scope": { "const": "all-reports-in-pinned-directory" },
        "detailed_receipt_steps": {
          "const": [0, 1000, 73000, 143000]
        },
        "published_report_without_detailed_receipt_count": { "const": 23 },
        "checkpoint_without_published_report_count": { "const": 127 },
        "detail_selection_method": { "const": "illustrative-hand-selection" },
        "detail_selected_before_outcomes_seen": { "const": false },
        "detail_representativeness": { "const": "not_claimed" },
        "detail_selection_reason": { "const": "provide richer receipts for four sparse non-monotonic examples" },
        "task_sampling_frame": { "const": "87-common-reported-tasks" },
        "included_task_metric": { "const": "piqa.acc_norm" },
        "omitted_task_count": { "const": 86 },
        "task_selection_method": { "const": "illustrative-hand-selection" },
        "task_selected_before_outcomes_seen": { "const": false }
      }
    },
    "claims": {
      "type": "object",
      "additionalProperties": false,
      "required": [
        "checkpoint_reference",
        "artifact_pointer",
        "artifact_bytes_verified",
        "artifact_executability",
        "digest_status",
        "metric_value",
        "metric_to_observed_branch_artifact_binding",
        "metric_comparability",
        "safety",
        "deployment_readiness",
        "resume_file_completeness",
        "resume_replay_equivalence"
      ],
      "properties": {
        "checkpoint_reference": { "const": "observed_remote" },
        "artifact_pointer": { "const": "observed_remote" },
        "artifact_bytes_verified": { "const": "not_tested" },
        "artifact_executability": { "const": "not_tested" },
        "digest_status": { "const": "reported_lfs_oid" },
        "metric_value": { "const": "reported_upstream" },
        "metric_to_observed_branch_artifact_binding": { "const": "not_proven" },
        "metric_comparability": { "const": "not_proven" },
        "safety": { "const": "not_proven" },
        "deployment_readiness": { "const": "not_proven" },
        "resume_file_completeness": { "const": "not_proven" },
        "resume_replay_equivalence": { "const": "not_tested" }
      }
    },
    "authority": {
      "type": "object",
      "additionalProperties": false,
      "required": [
        "carried_by_this_record",
        "established_by_this_view",
        "external_authority_status"
      ],
      "properties": {
        "carried_by_this_record": { "const": false },
        "established_by_this_view": { "const": false },
        "external_authority_status": { "const": "not_checked" }
      }
    },
    "unknowns": {
      "type": "object",
      "additionalProperties": false,
      "required": [
        "exact_evaluation_harness_revision",
        "evaluation_sample_count",
        "optimizer_state_for_exact_resume",
        "scheduler_state_for_exact_resume",
        "dataloader_cursor_for_exact_resume",
        "rng_state_for_exact_resume",
        "gradient_scaler_applicability",
        "data_order_index_for_exact_resume",
        "software_environment_for_exact_resume",
        "distributed_state_completeness",
        "artifact_format_equivalence"
      ],
      "properties": {
        "exact_evaluation_harness_revision": { "const": "unknown" },
        "evaluation_sample_count": { "const": "unknown" },
        "optimizer_state_for_exact_resume": { "const": "not_established" },
        "scheduler_state_for_exact_resume": { "const": "not_established" },
        "dataloader_cursor_for_exact_resume": { "const": "not_established" },
        "rng_state_for_exact_resume": { "const": "not_established" },
        "gradient_scaler_applicability": { "const": "not_checked" },
        "data_order_index_for_exact_resume": { "const": "not_established" },
        "software_environment_for_exact_resume": { "const": "not_established" },
        "distributed_state_completeness": { "const": "not_checked" },
        "artifact_format_equivalence": { "const": "not_tested" }
      }
    }
  }
}