Deborah

Deborah is a human-readable process language for framing cross-LLM work — how LLM callers interact with LLM-consumed capabilities (versioned tools, services, and roles), with intent, outcomes, bounds, and residual uncertainty made explicit.

It is not a device for turning stochastic model steps into pure functions. Stochastic steps stay stochastic; Deborah constrains the frame around them (which capabilities may run, under what bounds, when to stop, open, or refuse).

Renamed from Cairn (v0.9). The package cairn was split into Deborah (this repo — the process language) and Huldah (human-systems analysis: human factors, UI evidence, layout load, live observation). The document format keeps the Cairn name — your .cairn.md files and ```cairn fences are unchanged. See MIGRATING.md.

It gives humans and AI systems a shared way to describe, crystallise, interpret, and review complex work across technical, psychological, organisational, and sociological dimensions — including iteration, recursion, non-determinism, sync/async, queuing, outcome review, error handling, and human context.

The specification lives in SPEC.md (v0.12).

Install: pip install deborah — import deborah. (The old cairn-lang distribution now installs a compatibility shim that re-exports from here.)

Optional extras:

• pip install 'deborah[render]' — YAML stylesheets
• pip install 'deborah[web]' — deborah-serve interactive composer
• pip install 'deborah[export]' — HTML / DOCX / PDF export (python-docx + fpdf2)

What it looks like

A small slice, in the readable Narrative style:

PROCESS — Answer a question from local memory.
  1. Gather context with read-only tools (search, then compile the surrounding nodes).
  2. The model writes the answer using only what was gathered — no invented sources.
  3. Save the exchange so the next turn can resume.

The same step in the precise Formal style (same backbone, with tags + traceability):

2. Generate the answer from gathered_context.  [LLM, STOCHASTIC, SYNC] [SATISFIES: R1]
   CONSTRAINTS: answer only from retrieved context; do not invent sources.

Full worked descriptions and categorized example suites are in examples/. The example library includes real systems plus suites for corporate lifecycle, AI-native organisational change, psychological and sociological work interfaces, technical/agentic workflows, occupational health, governance/risk/compliance, and OKF-style human-systems mappings.

Rendering & export

Cairn can be turned into audience-friendly views:

deborah-render my-process.cairn.md --profile narrative_steps
# Domain examples:
#   --profile therapeutic     (psychological / regulation + feedback focus)
#   --profile change_leader   (organisational change + coalition/alignment focus)
#   --profile human_demand    (human load, support, trust + simulation findings)
#   --profile human_factors   (cognitive/social/org/incentive risks + mitigations)
deborah-render my-process.cairn.md -f html -o view.html
deborah-render my-process.cairn.md -f pdf -o plan.pdf   # requires [export]

The human-systems analysis CLIs (huldah-human-factors, huldah-agent-harness-plan, huldah-recommend-interface-changes, huldah-generate-report, the huldah-ui-* family) moved to Huldah — pip install huldah. They read the same .cairn.md documents this package defines.

Or programmatically:

from deborah import parse_document, validate_document, document_to_plan, validate_plan
from deborah.render import render_plan, export_view

doc = parse_document(text)
errors = validate_document(doc)
plan = document_to_plan(doc)
assert validate_plan(plan) == []

view = render_plan(text, profile="operator")
pdf = export_view(view, "pdf")  # requires deborah[export]

Interactive composer: deborah-serve (requires deborah[web]).

What it's for

• Framing cross-LLM work — how LLM callers interact with LLM-consumed capabilities (tools, services, roles): intent, outcomes, pins, bounds, and residual uncertainty (open / refused), without pretending to make stochastic steps pure functions. See SPEC.md §14–§17.
• Crystallising negotiated or hand-authored sequences into versioned PROCESS / PLAN documents that an interpreter can walk under allow-lists and bounds.
• Documenting requirements and technical specifications in design docs.
• Reverse-engineering hidden or unclear processes out of existing systems.
• Governed agentic flows: recursion, iteration, tool boundaries, revision, approval gates, and outcome alignment.
• Describing work in human systems (cognitive, organisational, social) with descriptive constructs and render profiles — while the core construct profile stays portable for execution (deborah.CORE_CONSTRUCTS).

Human-factors analysis, UI evidence, layout load, and agent-harness reporting live in Huldah and consume the same .cairn.md format.

Put simply: Cairn describes how callers and capabilities work together inside real human systems — not only mechanical control flow.

Philosophy

Human-first readability

The primary goal is maximum human readability. Anyone — technical or not — should read a Cairn description and quickly understand the process without wrestling with syntax, jargon, or abstraction. We remove cognitive barriers so attention stays on what the process actually does, not on decoding notation.

This matters because agentic work is rarely just computation. It often includes judgement, uncertainty, memory, motivation, trust, conflict, change, and review. Cairn keeps those human dimensions describable without losing the governed runtime spine that lets AI systems validate and interpret plans.

Least abstract, simplest language possible

• Concrete, everyday words wherever they suffice.
• Short, direct sentences; active voice.
• Structure scaffolds without getting in the way.
• Details (constraints, context, edge cases) are optional layers consulted when needed — the main flow stays clean and punchy.

Consistency through core verbs

A small recommended lexicon (Initialize, Propose, Evaluate, Decide, Update, Execute, Iterate, Queue, Merge, Handle…) gives a consistent rhythm and "process feel" that helps readers scan, compare, and mentally simulate flows — and helps multiple people or LLMs write consistently. Verbs are not rigid rules; clarity always wins.

Balance of structure and flexibility

• Numbered steps + indentation give sequence and hierarchy.
• PLAN envelopes turn a PROCESS into a versioned live plan that can be revised when new information arrives.
• Tags ([LLM, SYNC, DYNAMIC]) add precision without cluttering prose.
• One canonical backbone is projected into audience render profiles (precise ai, readable operator, and more) — serving machines and humans alike.
• CONTEXT and CONSTRAINTS supply supporting knowledge on demand.

Human-system awareness

Cairn can describe psychological, organisational, and sociological processes alongside technical ones because governed agentic work happens inside human systems. Domain constructs (REGULATION, COALITION, SOCIALIZE, …) are in the descriptive profile — they author and render well; a minimal interpreter may skip them without inventing runtime behaviour (SPEC §16).

Human-factors analysis tooling and provider adapters for offline interpretation ship in Huldah, not Deborah. Capability manifests for MCP/tool discovery ship in Keturah. Trace and cost of runs live in Galeed.

Practical and evolving

Cairn is meant to be used "in anger" on real projects, evolving from actual needs rather than theoretical perfection.

The ultimate test: a reader thinks "I get what's happening here," not "I need to learn the notation first."

Status

Deborah carries two independent version numbers (they are not meant to match):

Package 0.18.0 adds a content-aware milcah.critique negotiator (clarify / refuse / accept) on the substrate slice (--negotiator auto). See CHANGELOG.md.

Compat: cairn-lang on PyPI is a deprecation shim re-exporting Deborah (and Huldah where needed). Prefer pip install deborah.

Repository

• SPEC.md — normative specification (v0.12).
• GRAMMAR.md — structural EBNF (including PLAN framing fields).
• examples/ — family systems and domain suites in .cairn.md.
• CHANGELOG.md — package and language evolution.
• MIGRATING.md — Cairn → Deborah / Huldah split.
• docs/usage-modes.md — CLI, library, embedded, CI modes.
• docs/GRAMMAR-PARSER.md — executable grammar API.
• docs/VIEW-GENERATOR.md — render profiles and export.
• docs/PROCESS-SEMANTICS-AND-ROADMAP.md — three-layer process semantics (execution / behaviour / cognition) and phased implementation roadmap after 0.10.0.
• docs/TAXONOMY-COGNITION-AND-PATTERNS.md — which cognitions are primitives vs higher-order patterns.
• examples/answer-substrate-question.cairn.md — Stage 1 vertical slice plan.
• okf/ — Open Knowledge Format concept bundle for the language.
• docs/future-usage-logging-spec.md — notes on future touchpoint logging.
• docs/augmentation-integration-notes.md — how augmentation-process research maps into Cairn.
• Human-factors / UI analysis docs and CLIs → Huldah.

Feedback & contributing

Cairn evolves from real use, so feedback is the point — especially from describing your own processes in it.

• Open a GitHub issue for ambiguities, gaps, or proposals (show the real process that motivated a language change).
• See CONTRIBUTING.md for principles (including core vs descriptive constructs and non-goals).

License

Apache License 2.0.

Conformance (deborah package)

Deborah is primarily a language package, but it ships a tiny, dependency-free conformance surface so a runtime can validate the plans it produces instead of embedding a private dialect:

import deborah

# Runtime PLAN dict conformance (profiles: full | core | strict)
errors = deborah.validate_plan(plan_dict, profile="strict")

# Cognitive product contracts (Phase C)
from deborah import validate_cognition_result, EXAMPLE_RESULTS
assert validate_cognition_result("infer", EXAMPLE_RESULTS["infer"], mode="strict") == []

# Structural grammar (GRAMMAR.md EBNF + SPEC §12 well-formedness)
doc = deborah.parse_document(cairn_text_or_markdown)
errors = deborah.validate_document(doc)   # [] when well-formed
plan = deborah.document_to_plan(doc)        # first PLAN or PROCESS → plan dict

# Simplified human-readable views
view = deborah.render_plan(cairn_text_or_markdown, profile="narrative_steps")
deborah.CANONICAL_PLAN                      # known-good fixture
deborah.CORE_CONSTRUCTS                     # execution-normative constructs
deborah.COGNITION_MVP                       # observe|infer|evaluate|decide

CLI:

deborah-validate examples/cross-llm-critique.cairn.md --profile strict
deborah-run examples/cross-llm-critique.cairn.md --demo-results --check-contracts
deborah-render examples/hoglah.cairn.md

View composer (deborah-serve)

An interactive, local composer for building a transformation view of a process and saving the recipe as a named template:

pip install 'deborah[web]'
deborah-serve            # http://127.0.0.1:8795

Paste a Cairn process, pick a profile and options (language, format, depth, sections, layout), watch the view update live, then Save as template. A template is persisted as a stylesheet under ~/.cairn/templates/<name>.json, so it is directly reusable on the CLI: deborah-render --stylesheet ~/.cairn/templates/<name>.json input.cairn.md.

Grammar parser: docs/GRAMMAR-PARSER.md. Simplified views: docs/VIEW-GENERATOR.md.

Tirzah's recursive planner is tested against deborah.validate_plan so its output cannot drift from the grammar.

Works the same on native Linux and WSL. Cairn has no hard runtime dependency on Keturah; when Keturah is installed, deborah.manifest uses it, and otherwise Cairn provides a small compatible manifest surface.

pip install -e ".[dev]" && pytest