**From Interveil Alpha to a scalable framework**  
  
**Strategy**  
  
Interveil should scale by proving that its distinctions survive new domains, multiple reviewers, adversarial inputs, and production constraints. It should not scale by expanding the ontology in anticipation of every possible case.  
  
**Gate A — Alpha ratification and usability baseline (0–6 weeks)**  
  
**Objective: establish that the current packet is accurate enough to test and that users can answer the six core questions.**  
  
Deliverables:  
  
	●	founder attribution and corpus ratification decision;  
	●	source-extraction review by a photographic-forensics or archival researcher;  
	●	exact passage selectors for every source-borne claim;  
	●	accessibility review against WCAG 2.2 AA interaction requirements;  
	●	12-participant comparative usability study described in EVALUATION.md;  
	●	signed release manifest and archived public release;  
	●	corrections as append-only version 0.2.1 or 0.3.0.  
  
Go/no-go evidence:  
  
	●	≥90% trace accuracy across the six core questions;  
	●	no critical forbidden-collapse failure;  
	●	median completion time no more than 1.5× a conventional summary for the test tasks;  
	●	every source extraction independently reviewed.  
  
**Gate B — Cross-domain replication (6–14 weeks)**  
  
**Objective: learn which vocabulary generalizes.**  
  
Packets:  
  
	1.	**LK-99** for replication, methods, preprints, independent laboratories, and scientific revision.  
	2.	**A bounded public-policy or historical packet** selected for primary records, amended interpretation, and source reuse; avoid live high-stakes medical decisions until governance is ready.  
  
Deliverables:  
  
	●	inter-annotator guidelines;  
	●	blind dual encoding and adjudicated disagreements;  
	●	vocabulary usage frequency and “could not represent” log;  
	●	typed dependency dimensions;  
	●	refined loss-state controlled vocabulary;  
	●	migration tool from Alpha schema.  
  
Go/no-go evidence:  
  
	●	≥0.75 agreement on primary epistemic role before adjudication;  
	●	fewer than five new primary concepts across both packets;  
	●	every added term justified by two or more real examples;  
	●	no domain-specific term added to the core when a profile can contain it.  
  
**Gate C — Provenance hardening (12–22 weeks)**  
  
**Objective: make source identity and transformations stronger without implying truth.**  
  
Deliverables:  
  
	●	C2PA manifest import and validation-result display;  
	●	W3C Web Annotation selectors for quoted spans, regions, and temporal segments;  
	●	content-addressed identifiers and lawful source snapshots;  
	●	physical custody-event profile mapped to PROV;  
	●	RO-Crate conformance checks;  
	●	founder-controlled signing and optional SCITT-compatible transparency receipts;  
	●	redaction and withdrawal semantics that preserve accountable tombstones.  
  
Go/no-go evidence:  
  
	●	round-trip export/import without semantic loss for the three packets;  
	●	compromised, expired, absent, and valid signatures displayed distinctly;  
	●	test users never infer that signature validity equals claim truth above a pre-registered error threshold.  
  
**Gate D — AI-assisted authoring (18–30 weeks)**  
  
**Objective: reduce packet-construction cost while keeping AI suggestions noncanonical.**  
  
Workflow:  
  
```text  
ingest → segment → propose roles/claims/links → dependency audit  
      → human accept/edit/reject → invariant checks → signed revision  
```  
  
Deliverables:  
  
	●	source-grounded claim extraction with exact selectors;  
	●	AI proposal queue and role-specific reviewer controls;  
	●	prompt/model/tool receipts;  
	●	uncertainty-preserving summarizer;  
	●	omission and qualification diff;  
	●	disagreement interface;  
	●	evaluation corpus with known forbidden-collapse traps.  
  
Go/no-go evidence:  
  
	●	no autonomous canonical writes;  
	●	≥95% of accepted summary sentences have complete claim traces;  
	●	measured reduction in annotation time with no material fall in accuracy;  
	●	all model changes trigger re-evaluation.  
  
**Gate E — Query service and product integration (24–40 weeks)**  
  
**Objective: make Interveil the evidence layer that enriches and constrains downstream AI answers.**  
  
Core services:  
  
	●	packet registry and immutable version store;  
	●	provenance-aware query API;  
	●	subgraph receipt for every answer;  
	●	claim-level retrieval with dependency-aware evidence counts;  
	●	lens generation separated from the record plane;  
	●	diff and notification service for evidence, interpretation, and governance revisions;  
	●	embeddable inspector components.  
  
Reference answer contract:  
  
```json  
{  
  "answer": "bounded natural-language response",  
  "packetRevision": "stable identifier",  
  "claimIds": ["..."],  
  "supportIds": ["..."],  
  "dependencyClusters": ["..."],  
  "assumptions": ["..."],  
  "losses": ["..."],  
  "frontiers": ["..."],  
  "renderReceipt": {"...": "..."}  
}  
```  
  
Go/no-go evidence:  
  
	●	every answer reproducible from a named packet revision;  
	●	p95 trace query latency target set from real user journeys, not guessed in advance;  
	●	no source-count feature that ignores dependency clusters;  
	●	privacy and licensing checks at ingestion and rendering.  
  
**Gate F — Federation and governance (9–18 months)**  
  
**Objective: allow independent institutions to exchange and contest packets without a central ministry of truth.**  
  
Deliverables:  
  
	●	Interveil application profile and conformance suite;  
	●	namespace and extension policy;  
	●	institutional signer and reviewer roles;  
	●	protected disagreement and appeals process;  
	●	packet merging without silent conflict resolution;  
	●	trust-policy configuration separate from evidence content;  
	●	public change log and security disclosure process.  
  
Go/no-go evidence:  
  
	●	two external organizations independently produce conforming packets;  
	●	conflicting packets can coexist and be compared;  
	●	governance decisions are traceable and reversible;  
	●	red-team review covers manipulation, reputation laundering, citation cascades, prompt injection, and abusive dossiers.  
  
**Commercial sequence**  
  
	1.	**Design partners:** investigative teams, research integrity groups, archives, and model-evaluation teams with bounded, high-value evidence packets.  
	2.	**Interveil Inspector:** hosted and embeddable reader for receipted evidence packets.  
	3.	**Interveil Studio:** reviewed authoring, extraction, revision, and collaboration environment.  
	4.	**Interveil Evidence API:** claim traces, dependency-aware retrieval, and render receipts for AI products.  
	5.	**Interveil Profiles:** governed domain packages for science, journalism, history, compliance, and organizational decisions.  
  
Do not lead with “truth certification.” The value proposition is inspectability, source independence, defensible revision, and AI answers that show their work.  
  
**Metrics that matter**  
  
	●	claim-trace completeness;  
	●	forbidden-collapse rate;  
	●	source-dependency detection precision/recall;  
	●	reviewer agreement and adjudication time;  
	●	frontier resolution rate;  
	●	percentage of summaries with complete receipts;  
	●	correction latency;  
	●	user accuracy on the six core questions;  
	●	provenance-aware answer adoption and retention;  
	●	number and severity of governance reversals.  
  
Avoid vanity metrics such as raw node count, graph density, number of ingested pages, or unqualified “consensus” counts.  
  
**Immediate next decision**  
  
## The highest-value next step is not infrastructure. It is a founder and expert review of Alpha 0.2.0, followed by the small comparative usability test. If the distinctions do not improve human answers in this packet, the model should be revised before another line of platform code is built.