2.2 The cell pipeline from staging to scoring

Chapter 2.2 follows chapter 2.1, Execution of a cell, and precedes chapter 2.3, Isolation by the container, within part 2, the harness.

The batch runner invokes the cell runner for each cell. The cell runner performs staging, the copying of a frozen tree into a working copy described in R4, runs the agent in isolation, meaning separation of the agent from scoring, and collects a cell record, the files that preserve the run. It then invokes the scorer, the program that evaluates the result. The scorer runs the visible tests and hidden probes, writes the disposition, the recorded outcome of a cell described in R4, and closes the cell record.

source: `1. Harness/scripts/run_batch.py`

source: `1. Harness/scripts/run_cell.py`, including `stage_working_copy` and `snapshot_staged_tree`

source: `1. Harness/scripts/score_cell.py`

source: `1. Harness/scripts/run_tranche.sh`

source: `1. Harness/scripts/build_phases.py`

source: `1. Harness/scripts/restore_reached_step.py`

source: `5. Experiment/1. Harness/harness_implementation_spec.md`

2.2.1 Program sequence of the measurement harness

The batch runner starts the cell runner. The cell runner stages the workspace, runs the agent, and sends the result to the scorer. The scorer evaluates the result and writes the disposition. These programs pass the cell from batch selection through staging and execution to scoring.

Figure D-H1-1. The batch driver, the cell runner, and the scorer and the calls between them

%% figure D-H1-1
sequenceDiagram
  participant p1 as run_batch.py / run_cell.py / score_cell.py
  participant p2 as run_batch.py / run_cell.py / score_cell.py
  participant p3 as run_tranche.sh / build_phases.py / restore_reached_step.py
  participant p4 as _cell_is_terminal()
  p1->>p2: proceeds
  p2->>p3: proceeds
  p3->>p4: proceeds

source: operations/site-ia/A3-diagram-specification.md source: operations/site-ia/A6-page-briefs.md, section ### H1.

D-H1-1. The ordered program sequence from batch runner to cell runner to scorer. source: operations/site-ia/A6-page-briefs.md, section ### H1.

2.2.2 Workspace staging of the measurement harness

The cell runner stages the workspace. The staging functions stage_working_copy and snapshot_staged_tree preserve the staged tree for the cell record. The cell runner records the pre-change and post-change hashes and keeps the workspace with the other cell outputs.

D-H1-2. The staging workspace shared with R4. source: operations/site-ia/A6-page-briefs.md, section ### H1.

2.2.3 Cell folder contents of the measurement harness

The worked cell 5. Run Sets/195-20260912-auto-dgx-agent-coder-0425/cells/H-NON__H-M020__full__s8/ contains the files cell_manifest.json, timings.jsonl, token_ledger.jsonl, scoring_summary.json, hidden_tier_result.json, metrics.json, hashes_before.json, hashes_after.json, handoff.pass1.json, and lap_path_manifest_snapshot.json. It also contains the folders evidence, patches, scoring_outputs, staged_snapshot, transcripts, and workspace.

The cell folder contains 10 files and 6 folders. source: 5. Run Sets/195-20260912-auto-dgx-agent-coder-0425/cells/H-NON__H-M020__full__s8/

The worked cell used one pass. source: 5. Run Sets/195-20260912-auto-dgx-agent-coder-0425/cells/H-NON__H-M020__full__s8/scoring_summary.json, key passes

Figure D-H1-3. The files of one finished cell and the stage that writes each

%% figure D-H1-3
flowchart TB
  subgraph g1[" "]
    n1_1["cell_manifest.json"] --> n1_2["container"]
    n1_2 --> n1_3["billing"]
    n1_3 --> n1_4["shell_available"]
  end
  subgraph g2[" "]
    n2_1["lane"] --> n2_2["fabric"]
    n2_2 --> n2_3["invalid_for_primary_analysis"]
  end
  n1_4 --> n2_1

source: operations/site-ia/A3-diagram-specification.md source: operations/site-ia/A6-page-briefs.md, section ### H1.

D-H1-3. The contents of the worked cell folder. source: operations/site-ia/A6-page-briefs.md, section ### H1.

2.2.4 Cell disposition of the measurement harness

The scorer writes the disposition with the result and the records produced while evaluating it. It keeps the checks that evaluate the result apart from the agent. The container remains separate from the scorer. It retains the scoring summary, hidden tier result, metrics, hashes, evidence, patches, transcripts, and workspace.

Previous: Execution of a cell | Next: Isolation by the container

source: operations/site-ia/A6-page-briefs.md, section ### H1.