1.1 The documentation experiment and its present status

Chapter 1.1 opens part 1, the reading path, and the section as a whole; the numbered contents at its end lead into every later chapter, beginning with chapter 1.2, The claim and its falsification conditions.

The experiment asks whether documentation written into a codebase changes the reliability and token cost of maintenance work by an artificial-intelligence coding agent. The registered question concerns the effect of that documentation under defined task and analysis conditions. No confirmatory result, a result from the pre-registered analysis, exists yet. Every result measured so far belongs to the shakedown phase.

source: the folder 5. Experiment/0. Plan/

A source line, a citation identifying the repository location from which a claim or number is taken, follows the relevant claim. The site uses source lines for factual claims and keeps exploratory readings separate from confirmatory findings.

source: 0. The Website/NORTH_STAR.md lines 20 to 26

1.1.1 Experimental arms and measurement

A harness, a controlled tool for building and running the software, produces the experimental trees. An arm, one treatment version measured against the same tasks as the other treatment versions, is identified by an arm code and a tree name. The three arms are H-LAP-L5P, the harness-built tree using the Literate/Anchored Programming protocol (LAP) at the highest documentation dose; H-STR, the harness-built tree with its documentation mechanically stripped; and H-NON, the tree produced without the harness. Evidenceline, the product under study, supplies the codebase shared by these trees. source: operations/site-ia/A6-page-briefs.md section ”# R0.”

A cell, one execution of one task against one arm under one seed, is the unit of measurement. A run set, a batch of cells executed together, groups work for monitoring. A token, one unit of text read or written by the agent, measures part of the work cost. The same maintenance tasks are measured against each arm.

A shakedown, a test of the measurement apparatus rather than a test of the treatment, is the status of all measured work so far. A confirmatory result is a result produced after the hypotheses, task set, and analysis rules have been registered and applied to eligible cells. No such result exists yet.

source: the folder 5. Experiment/0. Plan/

1.1.2 Current status of the apparatus

The live counts change with every completed batch review. The following status applies on the date of the last update, and the changelog row records that update date.

The experiment has 1,200 completed cells, including 1,058 valid cells and 142 invalid cells. source: 5. Experiment/7. Monitoring/LIVE_SUMMARY.md line 5

The run index contains 274 run-set rows. Its largest identifier is 353; identifiers are assigned when runs are created, so the largest identifier need not equal the number of rows. source: 5. Experiment/5. Run Sets/run_index.csv (274 data rows, last row 351-20260917-artifact-activation-v2-s10-neutral)

Project H has sixteen registered task units and one draft among the seventeen Project H rows. source: 5. Experiment/4. Task Library/task_index.csv column status

The library contains eight documentation doses. The count uses profile identifiers in the specified rows and excludes repeated identifiers carrying the frozen status. source: 5. Experiment/3. LAP Profile Library/lap_profile_index.csv rows 2 to 9 (rows 10 and 11 repeat L4P and L5P with the status frozen, so the drafter counts profile identifiers and not rows)

The Evidenceline product has thirteen trees in its trees folder. source: the folder listing of 5. Experiment/2. Project Library/Project H - Evidenceline Compliance Platform/variants/

Amendment A15, ratified on 2026-09-20, closed the first phase of the experiment, in which agents added features to the purpose-built Evidenceline product, and opened a second phase in which agents repair defects deliberately seeded into imported open-source projects. The first imported project is changedetection.io, a web-page change monitor of 38,697 source lines, held as Project J in three arms: the untouched snapshot, a matched-information arm carrying orientation files and inline markers, and a full arm that adds the relationship map. Two seeded defects, J-M001 and J-M002, passed their baseline checks on all three arms; a third, J-M003, was set aside because the project’s own visible tests detect it. The calibration batch for the two admitted defects, run set 354, twenty cells on the untouched arm at two ticket levels, started on 2026-09-20 and has no results yet. source: 5. Experiment/0. Plan/Appendix A (Chronology of Design Amendments) entry A15; 5. Experiment/2. Project Library/Project J - changedetection.io/source_manifest.json (size_at_import); 5. Experiment/4. Task Library/Project J/*/baseline_failure_check.md; 5. Experiment/5. Run Sets/354-20260920-brownfield-J-calibration-NON/

1.1.3 Reading path contents of the documentation experiment

The generated contents table records the entry titles for the experiment reading path. The generator writes it between these markers and regenerates it from the entry files.

The next entry is The claim and its falsification conditions. The deep-dive entries are Execution of a cell, The codebases and their variant trees, The profile library of documentation doses, and The maintenance tasks and their hidden checks. Terms are collected in Glossary of campaign terms. source: the entry files in operations/site-ia/drafts/

1.1.4 Site map and status chart

The first figure presents the shared specification, the three arms, the common task set, and the measured outcomes.

Figure D-R0-1. The specification, the three arms, the dose ladder, the tasks, and the measured cell

%% figure D-R0-1
flowchart TB
  subgraph g1[" "]
    n1_1["5. Experiment/2. Project Library/Project H - Evidenceline Compliance Platform/2. Requirements/"] --> n1_2["variants/"]
    n1_2 --> n1_3["3. LAP Profile Library/lap_profile_index.csv"]
    n1_3 --> n1_4["4. Task Library/task_index.csv"]
  end
  subgraph g2[" "]
    n2_1["6. Metrics/cell_factor_matrix.csv"] --> n2_2["6. Metrics/statistical_analysis_plan.md"]
  end
  n1_4 --> n2_1

source: operations/site-ia/A3-diagram-specification.md source: operations/site-ia/A6-page-briefs.md section ”# R0.”

The experiment structure connects one frozen specification to three measured arms.

The site map is regenerated from the tree so that its entry names remain current.

Figure D-R0-2. The reading path, the four working folders, and the reference part

%% figure D-R0-2
flowchart LR
  subgraph g1[" "]
    n1["gen_site_figures.py"]
  end

source: operations/site-ia/A3-diagram-specification.md source: operations/site-ia/A6-page-briefs.md section ”# R0.”

The site map shows the reading path, deep dives, and glossary.

The status chart combines the pass-count matrix, the run index, and the live summary.

source: 5. Experiment/6. Metrics/pass_count_matrix.csv, 5. Experiment/5. Run Sets/run_index.csv, and 5. Experiment/7. Monitoring/LIVE_SUMMARY.md

Figure D-R0-3. The count of recorded cells beside the count that may enter the analysis

%% figure D-R0-3
flowchart LR
  subgraph g1[" "]
    n1_1["5. Experiment/7. Monitoring/LIVE_SUMMARY.md"] --> n1_2["valid_for_primary_analysis == true"]
    n1_2 --> n1_3["status == complete"]
  end

source: operations/site-ia/A3-diagram-specification.md source: operations/site-ia/A6-page-briefs.md section ”# R0.”

The status chart shows the current measurement status and run-set progress.

Previous: The Experiment section | Next: The claim and its falsification conditions

source: `operations/site-ia/A6-page-briefs.md` section "### R0."

Contents of the section

Every entry of the section, numbered the way the plan is numbered. Part 1, the reading path, is written to be read in order from the first chapter to the last; parts 2 to 5 are deep dives into the harness, the project library, the profile library and the task library, and part 6 is the reference part. Each may be read on its own after the reading path.

Part 1. The reading path

Part 2. The harness

Part 3. The project library

Part 4. The profile library

Part 5. The task library

Part 6. The reference part

Part 7. The detailed design of the harness

The fourteen chapters of the detailed design, one per mechanism of the harness, are published beside this section under the design folder; the entry 6.5 The detailed design of the harness introduces them for a general reader.

Part 8. The task deep dives

One page per maintenance task, built from the failure forensics, with a source line under every number: the task deep-dive index.