FMIE Next-Gen v2 Spec

Version: v2.0
Date: 2026-03-26
Status: Locked
Owner: FMIE core session
Runtime baseline: 0.42.7

Purpose

This document locks the next-generation FMIE plan against the actual Foundation-V2 implementation state.

FMIE should not expand into another application reasoning layer. Foundation-V2 already owns routing, intelligence briefs, partner profiling, contradiction handling, and product workflows.

Next-gen FMIE should become the source-side engine for:

• private evidence extraction
• provenance and authenticity
• deterministic disclosure transforms
• frontier-safe abstract packet generation
• local proof and verification for Midnight anchoring

North Star

FMIE must own the parts of the stack that Foundation cannot safely reconstruct after the fact:

1. source-grounded evidence
2. source-grounded provenance
3. source-grounded disclosure transforms
4. source-grounded packet verification
5. runtime health and degradation semantics that Cortex can trust

Ownership Boundary

FMIE owns:

• media ingest, decode, chunking, live session runtime, and stream backpressure behavior
• OCR, ASR, face, object, scene, document, audio, sync, and authenticity compute
• per-asset evidence extraction and artifact generation
• Media Passport emission and local passport verification
• deterministic disclosure-transform artifacts and proof bundles
• frontier-safe abstract packet preparation from raw source material
• runtime status, health status, crash breadcrumbs, and degradation reporting

Foundation / Cortex own:

• KB/entity graph persistence and storage policy
• strategy twin state
• partner-fit scoring policy
• external-model orchestration
• trust scoring and governance decisions
• UI, product workflows, approvals, and publication logic

P0 Deliverables

1. Semantic Airgap Pack

FMIE emits deterministic, source-linked transforms that Foundation can route through the External Cognition Gateway without leaking raw source material by default.

New artifacts

summary.sensitivity_tags.json

{
  "schema": "fmie.sensitivity_tags.v1",
  "asset_id": "string",
  "spans": [
    {
      "id": "span_1",
      "modality": "text|ocr|transcript|image_region|video_region|audio_window",
      "source_ref": "page:3|segment:12|frame:144",
      "start": 120,
      "end": 168,
      "labels": ["pii_name", "company_confidential", "biometric_face"],
      "severity": "low|medium|high|blocking",
      "confidence": 0.97
    }
  ]
}

summary.disclosure_transforms.json

{
  "schema": "fmie.disclosure_transforms.v1",
  "asset_id": "string",
  "transforms": [
    {
      "transform_id": "tx_1",
      "mode": "public_research|abstract_reasoning|raw_exception",
      "input_refs": ["segment:12", "page:3"],
      "operations": ["redact_names", "surrogate_entities", "crop_region", "drop_audio"],
      "output_artifact": "abstract-packet-1.json",
      "sha256": "hex",
      "policy_hints": {
        "contains_biometrics": false,
        "contains_real_entities": false
      }
    }
  ]
}

Requirements

• every transform must be deterministic for the same source refs and transform recipe
• every transform must link back to source refs and forward to output artifacts
• biometric vectors remain default-deny and do not leave FMIE under P0 transforms
• transform artifacts must be discoverable in artifact-manifest.json and summary.passport.json

2. Meeting + Document Intelligence Pack v2

FMIE emits normalized strategic primitives so Foundation does not have to rebuild them from raw OCR and transcript fragments.

New artifacts

summary.document_need_vectors.json

{
  "schema": "fmie.document_need_vectors.v1",
  "asset_id": "string",
  "vectors": [
    {
      "id": "need_1",
      "source_refs": ["page:4", "table:2"],
      "strategic_themes": ["identity", "wallet", "privacy"],
      "required_capabilities": ["kyc", "selective_disclosure"],
      "constraints": ["telecom_distribution", "jp_market"],
      "timeline_horizon": "near_term",
      "urgency_score": 0.82,
      "confidence": 0.76
    }
  ]
}

summary.claim_blocks.json

{
  "schema": "fmie.claim_blocks.v1",
  "asset_id": "string",
  "claims": [
    {
      "id": "claim_1",
      "type": "claim|action_item|decision|blocker|deadline|quote",
      "text": "string",
      "source_ref": "segment:14",
      "speaker_ref": "speaker_2",
      "time_start_sec": 412.3,
      "time_end_sec": 428.9,
      "page_ref": null,
      "supporting_refs": ["ocr:7", "table:2"],
      "confidence": 0.88
    }
  ]
}

Requirements

• quotes, actions, decisions, blockers, and deadlines must retain source-local refs
• page, slide, chart, and table anchors must be stable enough for frontend playback and drill-in
• document_need_vectors complements existing meeting_need_vectors, not replacing them

Bridging Artifacts Already Shipped

The live Foundation-V2 runtime now has concrete Partner-Fit and Strategy Twin services, so FMIE also ships two additive bridge artifacts:

• summary.partner_fit_request.json
• wraps FMIE’s numeric projection into the live Partner-Fit FitRequest body shape
• summary.strategy_twin_signals.json
• wraps FMIE twin observations into the live Strategy Twin /api/v1/twin/ingest signals[] payload

These are not substitutes for richer future artifacts like abstract_packets or lineage_graph. They are compatibility sidecars that reduce Python translation churn while the rest of the next-gen pack continues to land.

3. Passport / Authenticity Pack v2

FMIE deepens Media Passport into a stronger challenge and verification base for Midnight anchoring.

New artifacts

summary.authenticity_regions.json

{
  "schema": "fmie.authenticity_regions.v1",
  "asset_id": "string",
  "regions": [
    {
      "id": "region_1",
      "source_ref": "frame:144",
      "bbox": [0.12, 0.18, 0.42, 0.61],
      "signals": ["temporal_face_inconsistency", "compression_boundary"],
      "score": 0.71,
      "confidence": 0.69
    }
  ]
}

summary.lineage_graph.json

{
  "schema": "fmie.lineage_graph.v1",
  "asset_id": "string",
  "nodes": [
    {"id": "asset_root", "type": "asset"},
    {"id": "transform_1", "type": "transform"}
  ],
  "edges": [
    {"from": "asset_root", "to": "transform_1", "type": "derived_by"}
  ]
}

summary.challenge_bundle.json

{
  "schema": "fmie.challenge_bundle.v1",
  "asset_id": "string",
  "passport_sha256": "hex",
  "artifact_root_sha256": "hex",
  "midnight_anchor_payload_sha256": "hex",
  "proof_material": {
    "manifest_sha256": "hex",
    "lineage_sha256": "hex",
    "transform_sha256": "hex"
  }
}

Requirements

• authenticity outputs must explain which regions and signal families contributed
• the challenge bundle must be locally verifiable before Foundation anchors anything on Midnight
• proof-material hashes must align with the Passport / Gateway / Midnight contract set already locked in the external specs

4. External Cognition Prep Pack

FMIE emits packet-ready abstract artifacts so Foundation does not have to reconstruct safe external payloads from raw sidecars.

New artifact

summary.abstract_packets.json

{
  "schema": "fmie.abstract_packets.v1",
  "asset_id": "string",
  "packets": [
    {
      "packet_id": "pkt_1",
      "task_type": "public_trend_research|abstract_ranking|abstract_comparison|claim_validation",
      "modality": "text|document|image|video|audio",
      "source_refs": ["segment:14", "page:4"],
      "artifact": "abstract-packet-1.json",
      "transform_id": "tx_1",
      "safe_for_external": true,
      "contains_real_entities": false
    }
  ]
}

Requirements

• packets must always reference the transform that generated them
• packets must be discoverable from the manifest and passport
• P0 packets are advisory/prep artifacts only; Foundation still decides whether to send them externally

P1 Deliverables

5. Operational Runtime Pack

New artifact

summary.runtime_events.jsonl

Line-oriented event stream for:

• decode failures
• degrade-mode entry
• model pressure
• queue pressure
• backpressure drop counts
• live-session reconnects
• crash breadcrumbs promoted from /tmp/FMIE.log

Requirements

• events must be pollable through the ABI, not only file-based
• Cortex must be able to distinguish malformed media from system saturation

6. Cross-Camera Continuity Pack

Deliver continuity primitives beyond pairwise sync and link similarity:

• handoff windows
• reappearance hints
• drift-aware normalization
• event stitching candidates

Planned ABI Additions

• fmie_get_sensitivity_tags_json(...)
• fmie_get_disclosure_transforms_json(...)
• fmie_get_document_need_vectors_json(...)
• fmie_get_claim_blocks_json(...)
• fmie_get_abstract_packets_json(...)
• fmie_get_runtime_events_json(...)
• fmie_verify_challenge_bundle_json(...)

All of these are additive and should return FMIE-owned JSON strings plus normal fmie_free_string(...) cleanup behavior.

C++ Sprint Breakdown

Sprint 1

• sensitivity tagging for transcript, OCR, and image regions
• disclosure transform compiler scaffold
• emit summary.sensitivity_tags.json
• emit summary.disclosure_transforms.json

Sprint 2

• document need-vector extraction
• claim/action/decision/blocker/deadline anchors
• emit summary.document_need_vectors.json
• emit summary.claim_blocks.json

Sprint 3

• challenge bundle generation
• lineage graph export
• stronger proof-material hashing
• initial authenticity-region attribution scaffold

Sprint 4

• abstract packet generation for transcript, storyboard, and document structures
• manifest/passport linkage for packet artifacts
• biometric hard-deny enforcement for default transforms

Sprint 5

• runtime event stream
• degrade-mode telemetry
• model pressure and queue hints for Cortex

Sprint 6

• cross-camera handoff windows
• reappearance hints
• broader topology validation

Acceptance Criteria

• every transform is reproducible and hashable
• every abstract packet links to source refs and a transform id
• every challenge bundle can be locally verified before Midnight anchoring
• Foundation can consume need vectors and claim blocks without reconstructing them from raw evidence
• biometric vectors never leave through default P0 transform flows
• runtime pressure and degrade-mode events are visible to Cortex without spelunking FMIE internals

Blunt Rule

If Foundation can reconstruct it safely from FMIE artifacts later, it does not belong in next-gen FMIE.

If Foundation cannot prove or reproduce it later without the source media, FMIE should probably own it.