# Commitments → rationale map

This map ties each load-bearing commitment in [SPEC.md](../SPEC.md) / [CATALOGS.md](../CATALOGS.md) to the **research position** that grounds it. The spec stands alone (it carries no decision-record citations); this trace is where the *why* lives, one hop away. It is a **curated defensibility subset** — one position per major commitment, not the full research substrate — and it **grows by extension** as more commitments earn a ported rationale.

The positions are agent-authored research syntheses (`provenance: agent-synthesis`); each cites its own external grounding (the traditions, papers, and standards it rests on) via `[handle]{N}` citations that resolve against the consolidated bibliography in [references.md](references.md). They have been de-coupled from the originating deployment (decision-record citations and deployment-specific paths stripped) but their reasoning and external groundings are intact.

| Spec commitment | Grounding position |
|---|---|
| [§2](../SPEC.md#2-the-failures-ard-addresses) — anchor-and-drift as the canonical failure mode | [anchor-and-drift-three-tradition-framings](positions/anchor-and-drift-three-tradition-framings.md) |
| [§2](../SPEC.md#2-the-failures-ard-addresses) / [CATALOGS §1](../CATALOGS.md) — the locus-of-failure coordinate system (clean-room in the spec; de-anchored from PIES, the relationship recorded in this position) | [locus-of-failure-coordinate-system](positions/locus-of-failure-coordinate-system.md) |
| [§4.2](../SPEC.md#42-the-per-source-attestation-primitive) — the per-source attestation primitive | [attestation-primitive-as-testimony-chain](positions/attestation-primitive-as-testimony-chain.md) |
| [§4.6](../SPEC.md#46-strict-layer-directionality) — strict layer-directionality; [§11](../SPEC.md#11-known-limitations) — the paraphrase-drift residual | [directionality-fences-cross-tier-not-paraphrase-drift](positions/directionality-fences-cross-tier-not-paraphrase-drift.md) |
| [§4.6](../SPEC.md#46-strict-layer-directionality) — the hermeneutic-circle answer; [§11](../SPEC.md#11-known-limitations) — the fused-horizons gap | [hermeneutic-circle-settlement-externally-confirmed](positions/hermeneutic-circle-settlement-externally-confirmed.md) |
| [§4.7](../SPEC.md#47-per-claim-epistemic-status-markers) / [§10.4](../SPEC.md#104-citation-form-and-provenance) — per-claim markers vs file-level provenance as distinct axes | [marker-vs-provenance-separation-earns-its-place](positions/marker-vs-provenance-separation-earns-its-place.md) |
| [§3.1](../SPEC.md#31-the-floor--the-always-activated-core) — the reachability-indexed, shape-relative hard floor | [reachability-indexed-floor-membership](positions/reachability-indexed-floor-membership.md) |
| [§7](../SPEC.md#7-the-verification-stack-the-rigor-control) — the verification stack as a catalog with a hard-floor core | [verification-as-catalog](positions/verification-as-catalog.md) |
| [§8](../SPEC.md#8-the-positioning-dials-the-dial-space) — the dial-space / engagement primitives | [research-engagement-primitives](positions/research-engagement-primitives.md) |
| [§9](../SPEC.md#9-the-registration-contract) — the always-present registration schema | [registration-schema-store-pole-earns-its-place](positions/registration-schema-store-pole-earns-its-place.md) |
| [§9](../SPEC.md#9-the-registration-contract) — record-provenance gap (registry-conditioned) | [record-provenance-gap-shared-registry-conditioned](positions/record-provenance-gap-shared-registry-conditioned.md) |

## Not yet traced (grows by extension)

Several commitments do not yet carry a ported rationale here — they rest on the spec's own argument and the failure-shape inventory in CATALOGS §1 (ARD's own analytical artifact, in ARD's own locus-of-failure coordinates — see the [locus-of-failure-coordinate-system](positions/locus-of-failure-coordinate-system.md) position). The verb-set ([§6](../SPEC.md#6-the-act-vocabulary-the-verbs)), the decision-graph ordering ([§10.1](../SPEC.md#101-the-decision-graph-activation--ordering)), discipline propagation ([§5](../SPEC.md#5-discipline-propagation-the-invariant)), and the data contracts ([§10](../SPEC.md#10-data-contracts)) are specified directly in the spec. **So are five of the anti-fabrication core's eight disciplines**: the source-bound citation discipline / three-state rule ([§4.1](../SPEC.md#41-source-bound-citation-discipline)), the substrate test ([§4.3](../SPEC.md#43-the-substrate-test)), the composed-claim discipline ([§4.4](../SPEC.md#44-composed-claim-discipline)), contradiction-handling ([§4.5](../SPEC.md#45-contradiction-handling-discipline)), and corrections-vs-reversals ([§4.8](../SPEC.md#48-corrections-vs-reversals)) — each rests on the spec's own argument and the failure shapes it fences, with no ported position yet. As the framework matures, positions grounding these can be ported in — the trace is curated, not closed.

**v0.2 additions — grounding noted, positions not yet ported.** The newly-populated `WM` locus and the `GR`/`CO` extras (CATALOGS §1) were each surfaced against an external tradition in the originating deployment's analytical work — broadly: `WM.1` (per-artifact-warrant fallacy) against Quine's confirmation holism; `WM.2` (theory-ladenness of the substrate test) and `GR.7` (paradigm-trained gestalt-fit) against Kuhn; `GR.6` (attestation-abuse) against Austin's misfire/abuse cleavage; `CO.2` (premature normalization of abnormal discourse) against Rorty; `WM.3` (consumer-side decoding mismatch) against Austin's uptake conditions + Eco's aberrant decoding. The reasoning is intact in that deployment's substrate; de-coupled positions for these are not yet ported here. The **per-class IP profile** layer (CATALOGS §2) is operational/jurisdictional guidance rather than a tradition-grounded commitment — it rests on standard fair-use / license-compliance practice and carries the explicit "not legal advice" caveat in the catalog. These are honest gaps in the curated trace, not closed.

**v0.4.0 additions — grounding noted, positions not yet ported.** The optional **typed cross-reference** contract ([§10.5](../SPEC.md#105-typed-cross-references-optional) / [CATALOGS §9](../CATALOGS.md)) rests on established citation/argumentation ontologies — CiTO 2.8.1 (the citation-typing predicates), IBIS/Toulmin (argument-to-position: `grounds` / `supports` / `objects-to`), and SKOS (`related`) — plus the LLM-era knowledge-graph evidence for *small-closed-vocabulary + ad-hoc-fallback* over both CiTO-rich and free-form-extracted vocabularies (KAG's six closed predicates; OG-RAG's ontology-grounded-extraction gains over non-schema baselines). Its harvest signal is **two independent deployments converging on the identical twelve-predicate subset** — the cross-deployment convergence the bidirectional-harvest discipline treats as the tell that a convention generalizes, rather than a single deployment's quirk. The **handle-uniqueness** check ([CATALOGS §3](../CATALOGS.md)) is an integrity invariant of the citation wire-form itself — a handle must resolve to exactly one source — rather than a tradition-grounded commitment; like the per-class IP profiles, it is operational. De-coupled positions for these are not yet ported here.

**v0.5.0 additions — grounding noted, positions not yet ported.** The **decomposition-rationale** data contract ([§10.6](../SPEC.md#106-decomposition-rationale-multi-specialist--program-walks)) and the **proactive-enrichment acquisition discipline** ([§4.1](../SPEC.md#41-source-bound-citation-discipline) / `AQ.3`, CATALOGS §1) were harvested from deployment use: the decomposition-rationale artifact is the persistence layer that makes `decompose`'s multi-sampling non-inert against `PR.2` (its rationale is carried inline in §10.6 — the discipline is inert unless the candidates are persisted), and `AQ.3` extends the three-state rule (grounded by the [anchor-and-drift](positions/anchor-and-drift-three-tradition-framings.md) tradition work in the citation case) from the citation act to the acquisition-suggestion act. De-coupled positions for these are not yet ported here.

**Later additions — grounding noted, positions not yet ported.** The engagement-economics arc rests on established external work, recorded here with positions deferred (curated, grows by extension). **`AQ.4`** (inadequate-attempt blocking) + the acquisition-mode-exhaustion fence ([§4.1](../SPEC.md#41-source-bound-citation-discipline)) draws on the information-seeking *stopping / abandonment* literature — good-vs-bad abandonment [li-good-abandonment-search]{47} and cognitive stopping rules [browne-cognitive-stopping-rules]{48} — read against the reliability distinction between transport-level (transient, retryable) and content-level (permanent) failure; it is the `AQ.3`-sibling extension of the acquisition discipline. (The abandonment literature grounds *that* premature stopping is a real failure; the specific transport-vs-content cut — session-gated 403 / anti-bot / JS-render as transport vs hard 404/410 as content-absence — is ARD's **operational** classification resting on standard transient-vs-permanent fault practice, not a tradition-grounded commitment, like the per-class IP profiles.) **`GR.9`** (metadata recall-sourcing) + the metadata-tier source-bound extension ([§4.1](../SPEC.md#41-source-bound-citation-discipline) / [§4.2](../SPEC.md#42-the-per-source-attestation-primitive)) rests on the measured finding that LLMs mis-fill bibliographic metadata of *real* sources — DOI accuracy ~16–20% on non-hallucinated references while titles run ~99% [chelli-hallucination-reference-accuracy]{45}, with ~24–43% of real citations carrying substantive metadata errors [walters-fabrication-citations-chatgpt]{46} — with the provenance-binding remedy framed by W3C PROV's entity/activity/agent model [w3c-prov-dm]{1}[w3c-prov-o]{2} (bind each metadata value to the fetch activity that produced it). The **model-diversity verification property** ([§7](../SPEC.md#7-the-verification-stack-the-rigor-control)) — the *partial* fence for the correlated-verifier limitation ([§11](../SPEC.md#11-known-limitations)) — rests on the evaluator-correlation literature: same-family judges share errors and self-preference [zheng-judging-llm-as-judge]{42}, intrinsic self-correction fails without an external signal [huang-llms-cannot-self-correct]{44}, and — the honesty caveat the fence carries — the most capable models correlate *even across providers* [kim-correlated-errors-llms]{41}, a special case of algorithmic-monoculture outcome homogenization [bommasani-picking-same-person-monoculture]{43}; cross-model review therefore *reduces*, never eliminates, shared blind-spots. The **`decision_relevance`** registration field + kickoff gate ([§9](../SPEC.md#9-the-registration-contract)) — the value-of-information yield gate — rests on decision-theoretic value of information [howard-information-value-theory]{49} and its expected-value-of-sample-information form [raiffa-schlaifer-applied-statistical-decision-theory]{51}, the marginal economics of search [stigler-economics-of-information]{50}, and sequential-analysis stopping rules [wald-sequential-analysis]{52}; the operative principle — *information has zero value if it cannot change the chosen action* — is the formal basis for the "what decision changes if this finds X?" test. The **`PR.3` class-complete sweep** ([§4.8](../SPEC.md#48-corrections-vs-reversals)) is a fence refinement grounded in the originating deployment's reconciliation-completeness evidence (a subset-flagged class whose unflagged siblings persist is the same non-propagation failure at the class level), not an external-tradition commitment. Substrate confidence on the new entries is **search-summary** (bibliographic existence verified; findings engaged via abstract / landing-page, not full source-direct reads) — itself an instance of the `GR.9` metadata discipline applied to ARD's own trace; de-coupled positions for these are not yet ported here.