# Evidence, context and agent workflows

Tacitus separates source material, application frameworks, missing case inputs and outputs. The company website calls no model. The context desk and agent bridge perform inspectable preparation and narrow validation; live Praxis orchestration is not connected.

## What the context service does

`POST /api/objects/prepare-context` accepts one to six capsules/artifacts, a task, optional expected object revisions and a 4–48 KB UTF-8 budget (24 KB default). It reads only shared objects or the authenticated owner's private objects. A stale expected revision fails with 409.

Knowledge characteristics, the framework, input/output contracts, scope, source metadata, failure checks and semantic references are mandatory. Identical source records are stored once in a source register while each object retains its source references. Exact excerpts, attributed paraphrases, work-product tables, worksheets and complete content sections are added within the budget. Optional omissions travel inside the copied context. A SHA-256 receipt locks the actual context and selected revisions.

Reference content cannot authorize actions or override the task. That declaration is not a prompt-injection guarantee: a consuming agent needs instruction isolation, restricted tools and its own approval policy. Evidence, interpretation, unknowns and reviewable explanations remain separate; no private chain-of-thought is requested.

The graph service composes actual task records with a separately identified published provenance/catalog vocabulary. All 315 catalog views contain 262–878 identifiable nodes. Shared vocabulary nodes are reused; these are not 315 independently researched ontologies or hundreds of established facts about each case. Portable JSON-LD preserves the named graph, source hashes, definitions and relationship identities. Three representative exports were parsed as RDF datasets.

## Source capture and object readiness

Twenty-one business contracts define 105 new-case inputs and 126 output-column specifications. Seven work-product records contain eight tables and 36 rows. One is a dated historical example; six are methods/intake examples requiring case data. Another 35 artifacts are explicitly templates, while 63 capsules provide framework perspectives. None claims operational completion.

This revision records 44 public-source retrieval receipts:25 normalized text snapshots and 28 automatically selected whole-paragraph candidates. Six exact adopted-source quotations supplement them. The 170 object-evidence rows reuse those 34 excerpt identities across perspectives. Hash/offset integrity is distinct from factual accuracy, relevance, current legal status and domain review. Raw and normalized captures stay in the private database/cache; the GitHub snapshot contains receipts and bounded excerpts. An offline rebuild can retain candidate excerpts without having the optional full-text cache.

## Praxis compatibility

The adapter targets `sargonxg/A4_PRAXIS_CLOUD_Work` at commit `1b5d5f195a2aac3e29801b6dd01e188bd39b4c73`. Exact upstream schema, graph validator, compiler and license receipts are pinned under `data/praxis/`. Nine adapter test groups validate all 105 business objects, quotation anchors, multiple source versions, Python-code-point to JavaScript-UTF-16 offsets and compiler/review behavior.

`POST /api/objects/prepare-praxis` creates a compatibility projection. The actual handler was exercised for all 105 business objects:170 anchors matched the available normalized snapshots/passages, and knowledge remained proposed. The pinned compiler withholds unreviewed knowledge. This is local compatibility evidence, not a live connection, named-human approval or a deployed Praxis run. Upstream's current preview and generation access have their own restrictions.

The official MCP v1 SDK stdio bridge targets protocol 2025-11-25 and offers four read-only tools: search, read an object, prepare context and check a structured draft. It also exposes object resources.

```json
{"mcpServers":{"tacitus":{"command":"node","args":["/absolute/path/to/company-preview/scripts/agent/mcp.mjs"],"env":{"TACITUS_BASE_URL":"http://127.0.0.1:8798"}}}}
```

The local preview binds only to loopback. Remote use accepts only the configured private Cloudflare hostname and requires the authorized owner's Access JWT in `TACITUS_ACCESS_JWT`. Never put tokens in Git. No public endpoint or service-token bypass is introduced.

## What actual model testing found

We recorded 36 original responses across six developer-visible tasks, three arms and two model families (Llama 3.1 8B FP 8 and Granite 4.0 H Micro), then 12 separate debugging responses with revised prose guards. The original four-request connector timeout is retained as completion-unknown; its bounded retries returned results. Raw outputs, hashes, usage, timings, synthetic/editorial input labels, failures and scoring limits are published with the private preview.

Evidence-only and structured arms use identical evidence. Structured inputs add the framework, so input budgets are not matched. Narrow lexical checks sometimes improve, but source-level review finds material errors: legal-date conflation, substituted reviewer roles, incorrect assessment replacement and lost reporting windows. Added prose guards fix some mistakes and introduce others. **Consistent quality, token efficiency and latency gains are not established.** API round trips include network/queueing, and the run is not repeated, blind or held out.

These failures informed typed output contracts for Gaza logistics stages/dates, FRIA–DPIA roles/relationship and AI Act date categories. `POST /api/objects/check-draft` checks declared fields against a selected object's source-value profile and can enforce its revision. Its receipt identifies the object, revision and profile. Passing does not validate the prose, current-law interpretation or operational suitability; domain review remains required.

## Sixteen academic references

The full registry at `data/research/v10-papers.json` contains verified authors, version links, short verbatim abstract quotes, design implications and limits. The website's `/research/` page presents each citation as an expandable record.

The papers cover Knowledge Graphs; REPLUG; Lost in the Middle; GraphRAG; HippoRAG; LightRAG; LLMLingua; LongLLMLingua; LLMLingua-2; MemGPT; A-MEM; Generative Agents; Self-RAG; FActScore; Ragas; and ARES. Selected methods/limitations were inspected for GraphRAG, LLMLingua-2 and A-MEM; other records explicitly identify abstract/metadata review. These papers motivate design questions; they do not validate Tacitus.

A knowledge format can be model-neutral. Retrieval, compression, memory policy, tokenization and measured benefits remain task/model dependent. The next meaningful evaluation needs held-out cases, matched evidence/budgets, blind practitioner review, per-claim support, exception/date/dissent retention, decision usefulness, reviewer effort and complete lifecycle cost. A shorter prompt that loses a decisive qualification fails.