The capability. A partner asks the engagement a question — in plain words, from the Companion — and gets an answer drawn from everything the engagement holds: the documents that were uploaded and the record the team built, together, with every part of the answer citing where it came from, and with the answer itself admissible into the record marked as machine-assisted.
The one structural advantage. Hebbia answers from documents. Loomworks would answer from documents and the record — and the record contains what was never a document: the reference calls, the corrections, the dissent. The question that sells it: "what did the reference calls say about churn, and what does the contract actually commit?" — no document reader holds the first half.
The honest boundary. Engagement-scale, not bank-scale. Hundreds of documents per engagement, not ten thousand pages by this afternoon, and no chase of Hebbia's connector breadth or finance-tuned extraction. Parity is explicitly not the goal; the instrument-and-record composition covers the heavy-reading case more cheaply than building it.
What it takes. Three dependencies in strict order: recall must be truthful (the W-5 wrongness diagnosed and fixed), then searchable (the §18.2 retrieval substrate — the index that has never existed), then conversational (this capability). Ingestion completeness runs in parallel on its existing arc. When it lands, the no-live-recall-demonstration rule lifts, and the working session gains its second live moment.
Two different things travel under one name. Ingestion — getting a data room's contents in — is the smaller half; Loomworks's gaps there are already filed on the ingestion-interfaces arc and are engineering, not architecture. Corpus intelligence — indexing, retrieving, and reasoning across the corpus at scale with citations — is the moat half, built over six years. This investigation is about a bounded, differently-shaped version of the second half. Anyone who reads "ingestion channel" as the whole answer will under-scope it; anyone who reads "match Hebbia" will mis-scope it.
Fact one: recall today is wholesale — an unfiltered window of the fifty most recent committed assertions; the question never selects anything. Fact two: no search index has ever existed — the repo-and-migration sweep for text-search machinery returned zero hits across the schema's entire history. Fact three (W-5): at six assertions — inside the window, where the known limit cannot explain it — recall returned a superseded value with the wrong lifecycle state. That is active wrongness, not omission, and it is a separate defect from the §18.2 gap.
The open diagnosis: the unabsorbed June→July engine window carries test names indicating a recall-truthfulness change request landed. Either W-5's observation predates the walk's engine state in some path-specific way, the truthfulness work covers a different path than the one the walk exercised, or the work regressed. The scoping arc's Step 0 resolves this against the tree and the absorbed session records — it is exactly the kind of question the v0.75 absorption pass exists to answer, and this arc should not proceed past Step 0 without it answered. Truthful precedes searchable; searchable precedes conversational. Building conversational answering on a recall path that returns superseded values would industrialise the wrongness.
The corpus is the engagement's whole holding: committed assertions with their lifecycle states and correction trajectories; uploaded documents' extracted full text; and the structural objects (concerns, shapes, renders) as they exist. One index over all of it, engagement-scoped, respecting the fence — the corpus a person can ask is exactly the corpus they can read.
The ask is a Companion intent: a plain-language question, answered from retrieved-and-ranked corpus material, with every claim in the answer citing its source — an assertion by display number and contributor, a document by name and location. Supersession-aware by construction: a superseded value appears only as superseded, with its correction beside it. Unknowns stated as unknowns; an answer the corpus cannot support says so rather than reaching.
The answer is a first-class object. On the partner's word, it enters the record as a contribution — machine origin marked non-suppressibly, its retrieval citations carried as source threads, admitted like anything else. This is the direct answer to the category's re-check gap: the checking lands somewhere and stays.
The retrieval mechanism is the scoping arc's decision, not this document's. The honest note: the database in place supports mature text-search machinery natively, which is the obvious first substrate at engagement scale; semantic/embedding retrieval is a candidate layer above it, with real costs (infrastructure, index freshness, the AI-invisibility posture of the product identity) that the arc weighs rather than assumes. Engagement scale is what makes the simple substrate plausible — which is another reason the boundary in Section 4 is load-bearing, not modest.
Engagement-scale: hundreds of documents, a record of thousands of assertions — the working set of one deal or one fund, not a bank's archive. No connector-breadth chase; sources enter through the ingestion arc's adapters as that arc delivers them. No finance-tuned extraction models; extraction quality rides the ingestion skills. No answer without citation, ever — an uncited sentence is a bug, not a style choice. And no bake-off positioning: outward material never invites a retrieval-speed comparison; the demonstration is the question Hebbia structurally cannot answer, asked of a corpus that contains reference calls.
Not parity with Hebbia and not entry into the diligence-execution category — the comparison's posture stands. Not a replacement for the instrument-and-record composition; a fund with genuine ten-thousand-page reading needs uses a reading instrument, and its surviving findings enter the record marked. Not a general web-search or cross-fund capability — the fence scopes the corpus, and cross-engagement asking waits on the aggregation direction with its own access questions. And not a commitment to any retrieval technology — that is the arc's first real decision.
D-1 — The boundary. Confirm Section 4's commitments as the capability's standing definition. Recommendation: confirm. D-2 — The sequencing. Confirm truthful → searchable → conversational, with the absorption-pass gate on Step 0. Recommendation: confirm. D-3 — Commission the scoping arc. One arc absorbing §18.2, the W-5 diagnosis, and this capability into a single scoping note with its own Step 0 inspection brief — rather than three separate threads over the same substrate. Recommendation: commission; it is the highest-leverage build in the suite by the standing priority read, and this investigation is its charter.