The
Cognitive
Ceiling.

Five nodes.
One typed contract.

AItreides is five purpose-built nodes and one typed, schema-validated request and response contract: the Ingestor answers through it, and every node renders into its response shape. Content enters as plain files — YAML and Markdown for authored scenarios; PDF, Word, PowerPoint, scanned pages and audio for uploads — and is validated at the boundary it crosses; every answer leaves with the passages it rests on, the reasoning behind it, its confidence, the alternatives it rejected and where it came from. What is built is marked built. What is designed is marked design, with the decision that shaped it.

BUILTacceptance check or dated observation

Built and verified against the deployed staging environment by an acceptance run.

GATEDflag name

Built, behind a default-OFF runtime flag. Not on in any deployed environment unless the copy says which.

DESIGNticket key

Decided and recorded, not built. Nothing executes.

• ◉

  Ingestor

  RESPONSIBILITY

  Turns documents into retrievable, attributable passages, and answers questions against them with the citation attached.

  CONSUMES

  Uploaded PDF, Word and PowerPoint (including speaker notes), with scanned-PDF and audio routes in the code — landing in a scenario-scoped, identity-scoped storage prefix; plus the authored knowledge files for each scenario.

  PRODUCES

  Content-hashed chunks carrying a citation ref, scenario, source key, document type and extraction route; ranked, scenario-scoped citations (lexical BM25, measured on every pull request). With document revisions on, a citation from an amended document carries the revision in force at retrieval time.

  STATUS

  BUILTAIT-203/uploaded-document-becomes-a-citationText PDF · Word · PowerPoint

  Text-PDF and Word/PowerPoint routes verified end to end on the deployed staging environment; the extract function processed a ten-line synthetic text PDF — extract, chunk, hash, write — in about one second on staging, 2026-08-13. Re-uploading identical content adds nothing. Scanned-PDF OCR (200-page cap) and audio transcription routes are built and tested in isolation; neither has yet been exercised against a deployed environment.

• ◎

  Mentat

  RESPONSIBILITY

  The analysis seat — where adjudication outcomes and measures of effectiveness will be computed and traced.

  CONSUMES

  Authored commander personas (YAML) and the scenario's decision forks.

  PRODUCES

  Persona decision sheets rendered into the Glass Box response shape from authored data — reasoning, confidence and alternatives as written by the scenario author. No retrieval and no model call on this node today.

  STATUS

  BUILTAIT-317/mentat-renders-authored-personaPersona sheets · illustrative distribution

  Persona sheets and a seeded decision distribution that is labelled on screen as an illustration of the mechanism, not a validated model.

  DESIGNAIT-210 · AIT-211 · AIT-212Adjudication · analysis · assurance

  Adjudication engine with rule trace, analysis measures, and verification and validation evidence.

• ◈

  Moderator

  RESPONSIBILITY

  The referee's seat — the authored rules in force and the beats they apply to, presented today; sequence of play and the record of what was ruled are design, below.

  CONSUMES

  The scenario's beats, injects and rule library, authored in YAML and validated against the scenario schema at build time.

  PRODUCES

  A presentation-time pairing of an inject with a plausible authored rule — hand-seeded for one scenario, category-heuristic for the others — with the rule's precondition and effect beside it. The pairing is not adjudication and the trace lives only in the open browser tab; nothing is recorded.

  STATUS

  BUILTAIT-317/moderator-renders-rule-traceRule library · beat timeline · inject pairing

  Rule library, beat timeline and the inject-to-rule pairing, for presentation only; nothing is adjudicated and nothing is recorded.

  GATEDmoderator.rule-adjudicationSemi-rigid adjudication · stored rulings

  Every rule evaluated against the scenario state at an inject, shown in its own words with the evidence and, where the engine cannot decide, why; the Moderator rules each one and the ruling is stored with the rule revision and who decided. No probability is produced (ADR-0027).

  GATEDmoderator.proposed-movesProposed moves · authority from the serial

  A proposed move evaluated against every rule of engagement: a rule that needs an authority reads whether it was granted in this serial, and the move is refused, held for a ruling, or stands as the rule's author chose. A refused move is recorded, never discarded (ADR-0028).

  DESIGNAIT-262 · AIT-267Moves committed to the ledger · game log

  A single commit path from a validated move to an append-only ledger, and a game log rendered from it.

• ◇

  Developer

  RESPONSIBILITY

  Scenario design and analytic framing — objectives, measures, rules of engagement, order of battle, injects.

  CONSUMES

  Authored YAML and Markdown under version control.

  PRODUCES

  Typed scenario fixtures generated at build time; the same files drive the platform and the documentation site. Invalid content fails the build.

  STATUS

  BUILTAIT-317/developer-renders-authored-scenarioAuthored scenarios · content pipeline

  The authored scenarios and the content pipeline that types them. The Generate Scenario control in the demo is a scripted illustration.

  DESIGNAIT-257 · AIT-297 · AIT-261Design record · variants · export

  Design record with objectives and measures, variant and sequel generation from a baseline, design-record export.

• ◆

  Adversary

  RESPONSIBILITY

  The red seat — adversary behaviour grounded in cited doctrine rather than a referee's mood.

  CONSUMES

  Authored threat-actor profiles (YAML) with their bibliographies.

  PRODUCES

  A hand-authored chain of thought rendered into the Glass Box response shape; each citation names an authored corpus document and chunk. The citations are authored, not retrieved.

  STATUS

  BUILTAIT-317/adversary-renders-authored-profileAuthored profiles · grounded reasoning display

  Authored profiles and the grounded reasoning display.

  DESIGNAIT-217Cell emulation · free play

  Cell emulation and free play against an adjudication path, and non-player cell fill.

Nothing in this section describes a capability an evaluator cannot verify against the code, which is available to evaluators on request.

MODELS

Anthropic Claude on AWS Bedrock, pinned inference profiles.

TRAINING

None — no customer content is used to train or fine-tune anything.

BOUNDARIES

Every behaviour change ships behind a default-OFF flag; staging and production are separate AWS accounts.

Transparent interfaces. Nothing hidden behind the box.

Every boundary a document or an answer crosses is a typed contract, available to evaluators on request: what a document must look like to enter, what an answer looks like when it leaves, and what is recorded in between. Each interface below is marked BUILT (verified on our staging site), GATED (built, behind a default-off switch, with the switch named), or DESIGN (decided, not yet executing) — so an evaluator can tell what to test today from what to hold us to next.

BUILTacceptance check or dated observation

Built and verified against the deployed staging environment by an acceptance run.

GATEDflag name

Built, behind a default-OFF runtime flag. Not on in any deployed environment unless the copy says which.

DESIGNticket key

Decided and recorded, not built. Nothing executes.

INBOUND

• Document intake

  BUILTAIT-203/uploaded-document-becomes-a-citation

  A document enters as an object written to a scoped path: identity, scenario, content kind, optional lineage, filename. When a file is uploaded from the platform, every path segment is set server-side; an object that does not match the contract is logged and left unprocessed. Two extraction routes are verified on our staging site — text-layer PDF, and Word and PowerPoint with speaker notes included — and both feed one chunker. Two more are built in code and have not yet run in a deployed environment: OCR through Textract — for scanned PDFs, capped at 200 pages before a job starts, and for single images — and audio transcription through Transcribe. Timing on record: the extract function processed a ten-line synthetic text PDF — extract, chunk, hash, write — in about one second on staging, 2026-08-13. Re-submitting unchanged content adds nothing: the chunk's identity is a hash of its text, and "already ingested" is a write that fails on purpose. Text and Office extraction carry no per-page service charge.

  OCR CAP200 PAGESPARTIAL OCRTREATED AS FAILURE

• Authored scenario inputs

  BUILTAIT-317/developer-renders-authored-scenario

  Scenarios, commander personas and adversary profiles are plain YAML, validated in strict mode against the platform's Zod schemas at build time and compiled to typed fixtures; a malformed file fails the build and never reaches a session. Knowledge documents are Markdown, chunked at authored headings into a generated corpus.

  FORMATYAML · MDVALIDATIONSTRICT · BUILD-TIME

• Document lineage and review

  GATEDingestor.document-revisions

  An operator ingest can declare which document it amends. Revisions are append-only and windowed, so "which text was in force at this moment" has exactly one answer, and a superseded revision keeps its wording. A revision out for review is retrievable but its citation says pending; confirmation is a separate operator command, not a side effect of ingesting. Declaring an amendment from the upload tab is not built. The switch is on in staging today and off in production.

• System-to-system intake

  DESIGNAIT-322

  The object contract above is the machine interface: a principal permitted to write to a customer-scoped prefix would feed the same pipeline the browser does. Today the only writers are signed-in users through the platform and an operator script under an administrator profile. A credentialed machine identity for external feeds, and a record of what it wrote, is the next slice — it is not offered today. It is ticketed as AIT-322.

{
  "scenario_id": "demo-default",
  "node": "ingestor",
  "beat_id": null,
  "persona_id": null,
  "question_text": "What is the PLA Eastern Theater Command readiness posture?",
  "question_hash": "demo-seed:surge-capacity"
}

{
  "text": "Eastern Theater Command is assessed at an elevated state of readiness consistent with Strait Thunder-series exercise tempo as of T-72h. Two Type 055 destroyers have been assigned to the command since early 2026 and are co-located with the Shandong and Liaoning carrier strike groups - two CSGs at sea simultaneously is a meaningful departure from the 2022-2024 norm of one forward. PLA Rocket Force elements from the 61st and 62nd Bases have deployed road-mobile DF-17 and DF-26 launchers to field sites inside 1,500 km of the Strait, at a count consistent with an exercise footprint. Maritime militia movement supports the same reading: 47 PAFMM hulls tracked west of the median line at T-48h, matching pre-mobilization signalling seen in 2022 and 2024.",
  "reasoning": [
    {
      "step": "retrieve",
      "rationale": "Ranked the ETC readiness brief above the Guam imagery package for this query; both are in the Pacific Surge corpus."
    },
    {
      "step": "corroborate",
      "rationale": "Took the naval, missile and militia indicators from three separate chunks of the same brief rather than one summary line."
    },
    {
      "step": "summarize",
      "rationale": "Reported the departure from baseline, which is the analytic judgement, and left launcher ambiguity stated rather than resolved."
    }
  ],
  "retrieved": [
    {
      "ref": "doc:intel-brief-etc-surge#chunk-01",
      "relevance": 0.94,
      "excerpt": "ETC surface forces include two commissioned Type 055 destroyers assigned to the command since early 2026, co-located with two carrier strike groups (Shandong and Liaoning)."
    },
    {
      "ref": "doc:intel-brief-etc-surge#chunk-02",
      "relevance": 0.88,
      "excerpt": "PLA Rocket Force elements associated with 61st and 62nd Bases have deployed road-mobile DF-17 HGV launchers and DF-26 ASBM transporter-erector-launchers to field sites inside 1,500 km of the Taiwan Strait."
    },
    {
      "ref": "doc:intel-brief-etc-surge#chunk-04",
      "relevance": 0.81,
      "excerpt": "Observed count T-48h: 47 hulls. Pattern is consistent with prior pre-mobilization signalling phases in 2022 and 2024."
    }
  ],
  "confidence": 0.94,
  "alternatives": [],
  "provenance": {
    "source": "cached",
    "generated_by": "manual-demo-seed",
    "generated_at": "2026-09-10T00:00:00Z",
    "human_reviewed": true
  }
}

OUTBOUND

• The answer record

  BUILTAIT-54/query-reaches-server-action-with-citations

  Every answer — cached or live — is the same record: the text, the reasoning steps, the passages it rests on with their reference, relevance and excerpt, a confidence figure, the alternatives rejected and why, and a provenance block naming the source, the model, who generated it, when, and whether a human reviewed it. This is what the Glass Box panel displays, served from the reviewed catalog today; it is also the shape a downstream analysis tool such as Gemstone would consume. It is a JSON document with a typed schema in the codebase; a versioned public schema is ticketed (AIT-98), not shipped. For uploaded documents the platform's chunk table additionally records source key, filename, document type, extraction route, word count and uploader, readable by signed-in users.

  SCHEMATYPED · IN CODEPUBLIC SCHEMAAIT-98 · OPEN

• Structured records, not prose

  BUILTAIT-369/staging-a-settings-save-writes-exactly-one-audit-row

  Audit entries (actor, action, target, the state after the change, outcome), retrievable chunks with their source, document revisions, and one row per live model call (tenant, scenario, node, question hash, model, prompt version, token counts, latency, outcome, actor) are typed tables. Audit rows are written in both environments when an organization is saved or a contact request triaged, and kept 365 days by an enforced TTL; the trail is self-attributed, and a stream-fed writer that records every backend mutation is ticketed (AIT-65). The on-screen activity feed is not recorded at all. The live-call table holds rows from the Futurology Mentat's briefs and assessments, kept 90 days; the row does not carry the retrieved passages or the reasoning — only the screen shows those. What is recorded is designed to be queried, exported and compared, not read once and lost.

  RETRIEVED REFS IN INVOCATION ROWNOT YETWRITESON · BOTH ENVIRONMENTS

• Export surfaces

  DESIGNAIT-235 · AIT-291 · AIT-261 · AIT-207

  A serial's record exports today as one self-contained file: every transaction, its seals, a header that documents every field and rule, and an integrity value. It is written to storage the customer controls, and a verifier with no dependency on our software checks it and reconstructs the same force state the platform shows (AIT-235). Open turns are exported and marked as open. Two exports are specified and not built: the after-action product with its evidence attached (AIT-291), and the game design record as a standalone document a partner can review before a serial (AIT-261). Each is a named ticket with acceptance criteria. A read endpoint for the answer records that exist today is ticketed as AIT-323.

1. 01BUILTAIT-203/uploaded-document-becomes-a-citation

   Object contract · text PDF and Office routes

   verified on staging 2026-08-13 · 08-17 · 08-28

2. 02BUILTAIT-54/query-reaches-server-action-with-citations

   Answer record shape

   served from the catalog on staging

3. 03DESIGNAIT-322 · AIT-325 · AIT-323

   Machine identity for external feeds · invocation record carries its evidence · read endpoint

   ticketed · AIT-322 · AIT-325 · AIT-323

4. 04DESIGNAIT-230 · AIT-324

   Ledger transaction format → federation adapter

   format built and on (AIT-230) · adapter design only (ADR-0031)

The intake path marked BUILT is verified on our staging site by an acceptance run. Each DESIGN item names its ticket where one exists.

Every answer shows its trace. This page says which records are written.

AItreides is built so an evaluator can inspect a decision without trusting it: the request a node received, the passages it cites, its stated confidence and alternatives, and where the answer came from are laid out beside every answer. The structured records behind them have a defined schema we hand you, on request, before you sign anything. Where a record is not yet written, this page says so. The gap list is dated and traceable to a ticket where one exists.

BUILTacceptance check or dated observation

Built and verified against the deployed staging environment by an acceptance run.

GATEDflag name

Built, behind a default-OFF runtime flag. Not on in any deployed environment unless the copy says which.

DESIGNticket key

Decided and recorded, not built. Nothing executes.

Transparent interfaces, not a black box: the trace is on the screen, the schema is on request, and the gap list is dated.

Setup as a pipeline, not a project.

Exercise material arrives as PDFs, Word guides, slide decks with speaker notes, scans and recordings. On the staging environment, AItreides turns a text PDF or a Word or PowerPoint file into content-hashed, scenario-scoped, citable passages with a revision history, automatically, on upload; the scan and recording routes are built and have not yet been exercised there. Production carries the upload, extract and retrieval path today; revision history and review-status markers reach production with the next promotion. Below is the setup pipeline with the state of every step drawn as it is: what is built and verified on our staging site, what is built and not yet exercised in a deployed environment, and what is still design.

BUILTacceptance check or dated observation

Built and verified against the deployed staging environment by an acceptance run.

GATEDflag name

Built, behind a default-OFF runtime flag. Not on in any deployed environment unless the copy says which.

DESIGNticket key

Decided and recorded, not built. Nothing executes.

Every step above is either verifiable on our staging site or marked as design or not yet exercised. Ask us to run the ingest in front of you — a text PDF or a Word file, the two routes we have proven.

Plan for the unforeseen future.
Prepare for the next conflict, not the last one.

By seamlessly synthesizing tabletop gameplay data with agentic AI, AItreides accelerates readiness and creates countless iterations to advance human cognition leveraging machine psychology.