art-ops

A toolkit for the analysis of cultural artefacts — texts, images, documents, practices. The aim: read as much out of an artefact as it will bear — without crossing into speculation. art-ops serves that aim by describing analysis as operations: single, nameable moves performed on the material. Because every operation carries a profile — what it needs, what it produces, what it forbids — analyses can be planned mechanically, moves from different traditions can be combined, and the reliability of every step is declared item by item.

---

The idea in three sentences

1. Analytical traditions contradict each other in their justifications — but many of their moves are identical or combinable. art-ops works at the level of the move.
2. Every operation is described as a profile: what it needs from the material, what it demands of the analyst, what it produces, what it forbids. From these profiles, analysis sequences can be planned mechanically — including the conflicts between operations.
3. Honesty is a construction principle: the register consistently distinguishes which statements are empirically tested and which are (still) reading-derived hypotheses — and every generated analysis plan declares this line by line.

What's inside

File	Content
catalogue.md / catalogue.yaml	Inventory: 24 operations, distilled from 21 analysis prompts of a curated collection — which traditions perform which move, where they overlap, how checkable their core claims are
operations-spec.md / operations.yaml	Mechanism: the 24 operation profiles, the conflict register (which operations exclude each other within the same pass) and the interop graph (which feed into each other)
orchestrator/conductor.md	Planner: takes an artefact + an analytical interest and produces an ordered, conflict-checked operations plan — with an evidence status attached to every recommendation. It plans; it does not execute.
operations/	Modules: executable single prompts per operation (so far the three of one real image analysis), plus a Region Citation Convention for images
examples/	Worked examples: the workflow on three artefacts (text / image / dialogue) — plan + execution per artefact
manifest.yaml	Machine-readable index: artefacts, versions, derivation relations, evidence state

How to read: the prose files are authoritative (they carry justifications and caveats); the YAML files are the machine-readable derivation. Where they disagree, the prose governs.

How to use this

These are LLM prompts, not a software package — "execution" means a session of an AI assistant that you run and judge. Three ways in:

1. Plan an analysis: give an LLM orchestrator/conductor.md together with operations.yaml and your artefact. The Conductor proposes a goal item from the register (you confirm it — never a silent choice) and produces a conflict-checked operations plan with an evidence status on every line. It plans; running the planned operations is your session.
2. Run a single operation: the three modules in operations/ are self-contained prompts (image segmentation · sign-binding · constructedness exposure) — paste one into an LLM with your artefact.
3. Work by hand: follow the four-step procedure in operations-spec.md §7 — the same mechanism without any AI in the loop.

The Conductor's entry triage references a tool (artifact-triage) from the source collection, which is not yet public; the Conductor carries a declared fallback for exactly this case (direct existence-marker read), so nothing here depends on files you cannot get. A rendered overview of the register lives at the project page.

What it can do

• Translate an analysis question into an ordered sequence of moves: "first segment, then sort sign-bindings, then expose naturalisations" — each step nameable, each step checkable.
• Make conflicts visible before they spoil the analysis: which moves demand mutually exclusive stances from the analyst, which orderings are forced.
• Analyse images, although the underlying prompts are text-shaped — via a Region Citation Convention that translates "quote the passage" for image regions.
• Declare every recommendation with its evidence status: empirically supported · refuted · reading-derived hypothesis.

What it cannot do (yet) — the limits, laid open

• The register is mostly hypothesis. The profiles and conflicts are derived from the wording of the source prompts, not tested against material. Only a few statements are empirically tested so far — among them the most remarkable finding: the first conflict of the register ever put to a test did not hold (two operations that, per derivation, ought to collide demonstrably run within the same pass — the "conflict" was an artefact of an underdetermined instruction). Consequence: the tool recommends and declares; it enforces nothing.
• The planner does not execute. A run mode is deliberately deferred until the register is more broadly validated — nothing should harden untested hypotheses into a hard gate.
• Interpretive operations remain interpretation. Operations like the naturalisation exposure have no hard falsifier; their outputs are justified readings with a mandatory counter-reading, not measurements. The tool makes this visible instead of hiding it.
• Modality gaps: sound, moving image, numerical data and code are not covered by the inventory; image analysis rests on the (itself still untested) region convention.

Origin and method

The register is a reading distillate: from 21 analysis-prompt files of a curated collection, the operations were extracted "de-named" — as verb + object, without theory brand names (those appear only where they are themselves the object of analysis). No prompt text was copied; this is a re-description at the level of the move, not a fork.

The empirical findings come from pre-registered test runs with blinded judging instances and constructed gold standards; the test identifiers are attached to the respective findings. The full test documentation is not part of this repo.

Status & licence

Under development (draft); API and register content may change. Licence: Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) — the same licence as prompt-library, the source collection this register was distilled from. See LICENSE.

No employer, client, or person-related materials: analyses of real private artefacts (e.g. photographs) remain outside this repo as a rule; a .gitignore net prevents accidental inclusion.