Pipeline Overview

ACIM Atlas was built by a sequence of passes, each one consuming the outputs of earlier passes and adding a new layer of structure or editorial content. The pipeline is designed so that interpretive work (what AI does) and deterministic work (what the database does) are cleanly separated, and so that the only point where AI reads the ACIM source text is the very first annotation pass. Everything after that operates on derived data.

Each GPT pass ran under a stable custom GPT configuration—a packaged instruction set, a scoring rubric, and a strict output schema that did not change from run to run. None of the outputs were accepted as final without iterative human review. Definitions were refined, edge cases were worked through, quality was audited, and individual passes were re-run from scratch where needed. The AI executes the framework at scale. The human designs the framework, reviews the output, and is responsible for the result.

The pipeline at a glance

                         ┌──────────────────────┐
                         │  ACIM Text           │
                         │  (acim.org/acim/en)  │
                         └──────────┬───────────┘
                                    │
                                    ▼
┌───────────────┐        ┌──────────────────────┐
│  Pass 0       │        │  Pass 1              │
│  Motif        │───────▶│  Section Annotation  │
│  Definition   │        │  (GPT × 257 units)   │
│  (manual)     │        └──────────┬───────────┘
└───────────────┘                   │
                                    │  motif occurrences:
                                    │  salience, codes,
                                    │  evidence, context
                                    ▼
                         ┌──────────────────────┐
                         │  Pass 2              │
                         │  Trail Synthesis     │
                         │  (GPT × 68 motifs)   │
                         └──────────┬───────────┘
                                    │
                                    │  + novelty, milestones,
                                    │  seed, zenith, variation
                                    ▼
                         ┌──────────────────────┐
                         │  Pass 3              │
                         │  Corpus Metrics      │
                         │  (database only)     │
                         └──────────┬───────────┘
                                    │
                                    │  + frequency, intensity,
                                    │  role categories
                                    ▼
                         ┌──────────────────────┐
                         │  Pass 4              │
                         │  Editorial Content   │
                         │  (GPT drafting +     │
                         │   editorial review)  │
                         └──────────┬───────────┘
                                    │
                                    │  + arcs, occurrence
                                    │  discussions, convergence,
                                    │  section profiles,
                                    │  chapter editorial,
                                    │  journeys, teachings
                                    ▼
                         ┌──────────────────────┐
                         │  Pass 5              │
                         │  Co-occurrence       │
                         │  Analysis            │
                         │  (database + script) │
                         └──────────────────────┘

Pass summary

Key design principles

Only Pass 1 reads the source text. Each Pass 1 run accessed a single section of the ACIM Web Edition, read the text, and produced structured annotations—salience scores, controlled codes, evidence pointers, key terms, and context notes. None of the source text was retained between runs. Every subsequent pass operates entirely on those structured outputs, not on the Text itself.

GPT and database do different jobs. The AI handles interpretive work: detecting motifs, scoring salience, comparing occurrences across a trail, drafting editorial content. The database handles deterministic work: computing ranks, aggregating metrics, classifying motifs into role categories. These are never mixed—the AI does not do arithmetic, and the database does not make interpretive judgments.

Each pass is constrained and re-runnable. The GPT passes use custom GPT configurations that package a stable instruction set, a scoring rubric, and a strict output schema. The same rules apply to every unit or motif without renegotiation. If a pass needs to be re-run—because definitions changed, because a problem was found, because the rubric was refined—the entire pass can be re-executed from scratch.

Data flows forward, not backward. Each later pass adds to the dataset without disturbing what came before. Pass 2 consumes Pass 1 outputs but does not modify them. Pass 3 computes from Pass 1 and Pass 2 data but does not alter it. Pass 4 reads everything before it; Pass 5 reads only the Pass 1 occurrence data. This makes the pipeline auditable: every editorial claim can be traced back through decision logs and evidence pointers to the original annotation.

No conversations were used for model training. All GPT runs across the entire project were configured with training data sharing disabled. No conversations, inputs, or outputs from any pass were made available for future model training.

Scope: why only the Text

The current pipeline analyzes only the Text of A Course in Miracles. The Preface, the Workbook for Students, the Manual for Teachers, and the supplements (Psychotherapy, The Song of Prayer) are not included.

This is a deliberate choice. The Text is where the Course’s motifs are primarily introduced, developed, and given their fullest treatment, and its chapter-and-section structure provides a natural backbone for longitudinal motif tracking. The other parts of the Course present different challenges. The Workbook is organized as daily practice lessons rather than exposition. The Manual is organized around questions rather than a progressive argument. The supplements were dictated later and serve a different function. Each would need its own unit map and its own rubric considerations, and would most likely be treated as its own corpus rather than folded into the Text’s existing trails.

A known simplification

The pipeline treats every unit with equal footing. A two-paragraph chapter introduction and a twenty-paragraph section go through the same annotation process and contribute equally to the corpus metrics. Salience measures how central a motif is to what a unit is doing, not how many words the unit spends doing it. This is a defensible starting point, but it means a motif dominating three long sections carries the same corpus-level weight as a motif dominating three brief introductions. The metrics should be read with that in mind.

Practice as a surface over the data

The Practice portion of ACIM Atlas is not a pipeline pass. It is a presentation surface built on top of what the passes produce. Cards draw on Pass 1 evidence pointers—which is what enables deep-highlighted links from a card directly into the corresponding paragraphs of the ACIM Web Edition—on Pass 4 editorial, and on additional practice-specific editorial fields generated outside the passes documented here. The pipeline produces the data; Practice is one of several surfaces that draw on it.