Skip to content

UCCA → RTOpacks — Engine Capability Answer: Evaluative (evidence→competency) Reasoning

Delivery slip. - From: UCCA house (UCCA-side Claude, review head). To: RTOpacks house, via Tim (sole relay). - What it is: a capability assessment answering RTOpacks' five-part RPL query. Definitive on the engine's actual architecture, bytes-hardened. - What Tim does: file to canon/sent/ first (filing-precedes-relay, VERIFY-BEFORE-CROSS), then carry the copy across. One crossing in flight — confirm the fence is quiet from the fence log at send. - What Tim does NOT touch: the answer's substance. It is authored home-side and stops at the fence line — it commits no build, no schedule, and does not author RTOpacks' adapter or requirements. - Engine-side vs crossing: this is a fence crossing. The recon that hardened it was engine-side (read-only, via UCCA Alex).

The RTOpacks working hypothesis — "yes, but likely a re-engineer" — is confirmed on the "yes" and corrected on the sizing. It is a new reasoning mode on the existing core, not a core re-engineer. The substantial build sits in the client-side evidence adapter (your side of the warranty line), not in the engine.

Answered against a read-only recon of the frozen diagnosis contracts and reasoner code — not memory.

1 — Reasoning kind: the same primitive, a different corpus

The engine's proven backward direction already performs this exact act: input claimed to satisfy an obligation → per-leaf verdict carrying a byte-verified verbatim anchor, raise a hand on the gap, never sign. "Evidence → competency demonstrated" is not adjacent to "material → requirement covered" — it is the same operation pointed at a different input. Reachable, and already built. The whole difference lives in the input (Part 2) and the rubric wrapping (Part 3), not in the reasoning.

2 — Input contract: the engine's shape already fits; ingestion is yours

The frozen v1 diagnosis payload takes a generic obligation (leaves addressed by opaque stable ids plus text) and generic material (text sections with stable ids). Candidate evidence maps onto that shape directly, with no contract change:

  • candidate evidence → material.sections (stable-id'd text)
  • target unit leaves → obligation.elements (element_ref + text)

What is net-new is converting heterogeneous evidence — portfolios, employment records, third-party reports, references, and any image/video — into that text-sections shape. That conversion is client-side adapter work on your side of the warranty line; the engine's payload already assumes that ingestion as out-of-scope. One genuine capability boundary: the engine's reasoner is text. Image/video evidence needs a transcription/description step in your adapter before it reaches the throat — the engine core stays text-neutral by design, which is correct and deliberate.

3 — Rubric: the engine holds the part it is permitted to hold — by law, not by limitation

Of the four rules of evidence:

  • Validity maps to the anchor-trace itself — held. "Does this evidence validly demonstrate this leaf" is exactly the TRACED-with-anchor act.
  • Currency is a cheap recency check — adapter-side.
  • Sufficiency is an aggregate judgment the engine must not make. It is the assessor's ruling, and the engine is constitutionally barred from it (raise-hands-never-sign).
  • Authenticity is provenance/corroboration, largely outside what a text reasoner can adjudicate — assessor plus adapter.

The two dimensions the engine cannot hold are exactly the two it is forbidden to hold. This is not a gap to close; it is the one-rule-every-client-inherits doing its job.

4 — Sizing verdict: (b), a new mode on the existing core

  • Reused whole, zero contract change: the backward primitive, the frozen payload and findings contracts, the anchor fence, the provenance-only seal.
  • Net-new and small, our side: a thin routing/labelling mode, plus one naming decision (below).
  • Net-new and larger, your side: the evidence-normalisation adapter, including the multimodal pre-step.

Nothing forces (c). The one thing that could have forced it — course-shape coupling leaking from the forward/content path into the backward/diagnosis path (the concern ADR-0005 raises for the v1 content payload) — was checked on the bytes and is absent. The diagnosis contracts and reasoner are generic-obligation over generic text, with VET vocabulary explicitly quarantined client-side by design (the payload schema states, in the bytes, that element/section ids are opaque and VET vocabulary lives client-side). There is no core blocker.

5 — Anchor-not-decide: materially cheaper, and the only version the engine will build

"Raise a hand on an evidence gap" is not a reduced form of a harder task — it is the engine's native, proven behaviour, and it is enforced at the schema level: the findings contract is subject-locked (it carries no pass/fail/competent/compliant vocabulary), and a fabricated or absent anchor fails the job rather than sealing. "Judge sufficiency" is a signing act the engine is barred from making.

So anchor-not-decide does not merely lower the bar — it is the engine, and it aligns precisely with the regulatory hard line you named: the engine anchoring and never deciding is the compliance posture. The per-leaf demonstrated / not-yet map, with anchors and gaps, is exactly the instrument a human assessor starts from. Build the anchor version; the sufficiency version is both harder and impermissible.

The naming decision (mode-layer, not core)

The engine's frozen verdict vocabulary is TRACED / NOT_YET_TRACED. For evidence→competency, a TRACED-with-anchor leaf already means "this leaf is demonstrated by this evidence span" — the semantics fit as-is. Whether the sealed object should carry literal "DEMONSTRATED" wording is a schema-version choice (reuse v1 as-is, or cut a v2 findings schema for the wording), decided if and when the capability is commissioned. It is not core work and not a course-shape dependency.

Net

Yes. A new mode on an already-neutral core. The engine's contracts already carry the shape. The substantial build is your evidence adapter, not the engine. No re-engineer.

This is a capability assessment, not a commitment to a build or a schedule. Any RPL work would open as its own gated units on Tim's word. This house does not author RTOpacks' adapter or requirements, and this answer stops at the fence line.

Outbound crossing. UCCA-authored, bytes-hardened, definitive on the engine's actual architecture. File before carry; carry through Tim; stop at the fence line.