# ARA World Model — Engine Contract (the shared interface) > This file is the **single source of truth** the engine binds to. The two agent contracts > (`RETRIEVE.md`, `PREDICT.md`) conform to the grounding references, read-only isolation, and > invariants defined here. If two documents disagree, **this file wins**. > > Scope = **single-ARA** operation: the engine reads **one ARA directory** directly. Multi-ARA > federation is a later layer; the references below are chosen so it can be added **without > renaming anything** (see §2). --- ## 0. First principles 1. **The ARA is the substrate.** The engine reads an ARA's native files **directly** — there is no projection step, no intermediate distilled database, and no global-ID scheme. The ARA produced by the compiler (`PAPER.md`, `logic/`, `trace/`, `evidence/`, `src/`) is already a structured, machine-readable artifact; the World Model reasons over *that*, not over a lossy copy of it. Grounding is the ARA's own native anchors (§2), used verbatim. 2. **Field-agnostic.** The engine knows the ARA's *structure*, never its domain. Which quantity is the objective, whether more is better, and what any domain term means arrive as **data in the ARA's own text** (the `problem.md` objective, the method, the evidence) — never from a rule in the engine, and never from your own priors about the field. The same engine works on a biology ARA, an HCI ARA, anything. 3. **Agentic core.** The two contracts (`RETRIEVE.md`, `PREDICT.md`) are **agentic**: each reads the ARA's prose and reasons over it. There is **no deterministic projection, distance, or index layer**; relevance and inference are semantic, over the ARA's real prose. 4. **Native-ref grounding.** Every assertion a contract makes carries one or more **native refs** into the ARA (§2) — used verbatim, with **no minted IDs**: trace node `trace:N109`; claim `logic/claims.md#C05`; concept `logic/concepts.md#`; method/problem `#
`; evidence `evidence/.../.md`. A ref promises the cited location actually contains the claim. 5. **Grounded answering over the ARA.** The Predictor reads the ARA and **answers any question grounded in it** — predictive, explanatory, comparative, evaluative, or anything else; it separates `grounded_inference` from a named `speculative_leap` and commits to a checkable point. The engine is **read-only** (§4b). --- ## 1. Engine layout ``` research-foresight/ # ENGINE — the skill, all-English contracts, no per-ARA data SKILL.md # entry point: retrieve precedent, then answer the question references/ CONTRACT.md # this file — the foundation both contracts bind to RETRIEVE.md # agentic retrieval over ONE ARA's native files PREDICT.md # grounded answerer over one ARA — any question shape / # THE SUBSTRATE — read directly, never copied or projected PAPER.md # the narrative logic/ # claims.md, concepts.md, problem.md, experiments.md, # related_work.md, and solution/* (the actual method layer) trace/ # exploration_tree.yaml — the FULL DAG: # experiment, dead_end, question, decision, pivot evidence/ # figures/, tables/, ... — grounded evidence src/ # code / configs / data as the work warrants ``` The engine carries no data and no per-ARA config: it operates on a bare ``, builds no intermediate database or index, and has no write path (§4b). --- ## 2. Grounding references Grounding is the ARA's **native anchors**, used **verbatim** — no minting, no `://…` strings. Every assertion a contract emits carries one or more of these refs: | what is referenced | native ref form | example | |---|---|---| | trace node | `trace:` | `trace:N109` | | claim | `logic/claims.md#` | `logic/claims.md#C05` | | concept | `logic/concepts.md#` | `logic/concepts.md#warmup-schedule` | | method / problem | `#
` | `logic/solution/method.md#step-3`, `logic/problem.md#objective` | | evidence | `evidence/.../.md` | `evidence/figures/figure3.md` | **Ref-honesty rule.** A ref is a *promise* that the cited location actually contains the claim it is attached to. A contract may not attach a ref it has not read and confirmed; an assertion with no resolvable ref is not allowed (§4a). **Multi-ARA hook (design only, not built now).** Single-ARA refs are the zero-prefix case — in single-ARA operation the prefix is simply omitted: `trace:N109`, `logic/claims.md#C05`. A later multi-ARA layer disambiguates by ARA name (`:trace:N109`); the contracts say "the ARA in scope", so no field is renamed to add it. --- ## 3. What each contract reads and writes | contract | reads | writes | |---|---|---| | `RETRIEVE.md` | the ARA's native files directly (`logic/`, `trace/`, `evidence/`, `src/`, `PAPER.md`) | **nothing** | | `PREDICT.md` | the `retrieval` object from `RETRIEVE.md` + the bodies of every ref it cites | **nothing** — its `Answer` output may carry a `contradiction_report` (PREDICT.md §7): sanctioned *output* for the caller to transport, not a write | `RETRIEVE.md` is **not** limited to `experiment`/`dead_end` nodes — `question`/`decision`/`pivot` nodes, `problem.md` framing, the `solution/` method, and evidence figures are all in scope and are often the highest-value context. --- ## 4. Invariants These bind every module; a contract that cannot satisfy one must say so rather than violate it. a. **Forced native-ref grounding.** Every assertion a contract emits carries one or more native refs (§2) that resolve into the ARA in scope. No resolvable ref ⇒ the assertion is not made. b. **The engine never writes.** Every contract (`RETRIEVE.md`, `PREDICT.md`) produces output only; the World Model writes nothing into the ARA — not `logic/`, not `trace/`, not `staging/`, nowhere. Write-back into the ARA's canonical `logic/` is owned by the ARA's own `research-manager`. c. **No domain branching.** No rule or inference keys on a domain, metric, or parameter name; domain specifics enter only as data read from the ARA in scope. d. **Honest disclosure.** Uncertainty is disclosed, never suppressed: the Retriever lowers its coverage verdict when the ARA does not cover the query, and the Predictor lowers confidence for any answer beyond what its cited refs support — a low-confidence bold answer is allowed; a confident ungrounded one is not.