Horizon Fidelity Monitor

"Quality is not a model property — it is a conversation property."

Horizon is a real-time conversation health monitor for AI agents. It tracks the structural dynamics of multi-turn conversations — semantic drift, information gain, ontological gap width, temporal desynchronisation, circadian cognitive load, conversation velocity, and causal reachability — dimensions that LLMs do not reliably surface from inside the conversation.

Horizon ships two measurement planes. The conversation plane (above, always present) measures the health of a dialogue turn by turn. The optional mission plane — Memento Mori — measures elapsed calendar time against goals: ages, deadlines, stalls, per-entity latency, and share of a finite horizon. It is inert until you configure a store. See Mission plane.

Horizon is not a manipulation, sycophancy, or human-influence detector — it measures conversation dynamics, not whether an agent is steering or flattering the user. See LEGAL.md §1.

Why an external monitor? LLMs have limited and unreliable self-knowledge: introspection research shows partial self-access that is brittle and degrades on complex or out-of-distribution tasks (Binder et al. 2024; arXiv:2512.12411). So rather than depend on a model reporting its own conversation dynamics, Horizon measures them externally with cheap, deterministic, always-on arithmetic that does not call the model at all.

Why this exists

Multi-turn AI agents lose accuracy. The ICLR 2026 Outstanding Paper "LLMs Get Lost In Multi-Turn Conversation" (Laban, Hayashi, Zhou & Neville — Microsoft Research / Salesforce Research) reports 39% average accuracy degradation across multi-turn evaluation — a structural property that standard observability tools (LangSmith, RAGAS, DeepEval) cannot see because they measure responses, not conversations.

Horizon was built to close that gap. It is observability first: it surfaces conversation dynamics that response-level tools miss, using cheap deterministic arithmetic with zero model calls. In four controlled A/B scenarios where Horizon events drove a re-grounding intervention we measured a +15.7% composite quality lift and 87% fewer hallucination events — but those are synthetic, scripted scenarios with a hand-tuned controller, not a production result. Treat them as promising in-house evidence, not a guaranteed outcome (see Validation and LEGAL.md §5). Every signal — information gain, divergence, estimated ontological gap width, causal reachability — is a standard information-theory or arithmetic measure computed on text embeddings and timestamps; see 4D Spacetime Signals for the full definitions.

• Read the demand proof → ICLR 2026 Outstanding Paper (Laban et al.)

• Read the category argument → docs/content/naming-the-category-conversation-dynamics-monitoring.md

• Read the engineering case → docs/content/why-every-production-agent-needs-conversation-dynamics-monitoring.md

Related MCP server: nautilus-compass

Getting started

Three paths — pick the one that fits your workflow:

Path 1 — Hosted MCP (fastest, zero install)

The fastest way to add Horizon to any Cursor, VS Code, or Claude Desktop workspace. No Python required.

Request an alpha key → open a Discussion, then add the config for your client:

Cursor (~/.cursor/mcp.json):

{
  "mcpServers": {
    "horizon": {
      "url": "https://horizon.leocelis.com/sse",
      "headers": { "Authorization": "Bearer YOUR_KEY_HERE" }
    }
  }
}

VS Code / GitHub Copilot (.vscode/mcp.json in your workspace):

{
  "servers": {
    "horizon": {
      "type": "http",
      "url": "https://horizon.leocelis.com/sse",
      "headers": { "Authorization": "Bearer YOUR_KEY_HERE" }
    }
  }
}

VS Code note: Use "servers" (not "mcpServers") and "type": "http" — VS Code tries Streamable HTTP first and falls back to SSE automatically, so "type": "http" works with the /sse URL.

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "horizon": {
      "url": "https://horizon.leocelis.com/sse",
      "headers": { "Authorization": "Bearer YOUR_KEY_HERE" }
    }
  }
}

That's it. Reload your MCP client and three tools appear: new_conversation, process_turn, configure_session.

Alpha access: Horizon's hosted endpoint is in private alpha. Keys are distributed to agent developers who want to monitor real projects. Open a Discussion to request one — describe your use case and we'll send a key.

Path 2 — pip install (library integration)

Not yet published to PyPI — until it is, use Path 3 (install from source) below.

pip install horizon-monitor

Verify your install (exercises the full pipeline on 5 canonical scenarios, ~25s):

horizon-validate

Path 3 — MCP server from source

pip install 'horizon-monitor[mcp]'
horizon serve                             # stdio — for Cursor, Claude Desktop
horizon serve --transport sse --port 3847 # SSE — for web/team deployments

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "horizon": { "command": "horizon", "args": ["serve"] }
  }
}

Full Cursor and Claude Desktop setup guides: docs/integrations/

What it monitors

Standard observability tools evaluate individual response quality. Horizon evaluates conversation quality — a structurally different problem:

Horizon does not replace per-response or LLM-judge quality tools. The differentiator is how it measures: deterministic, zero-LLM-call arithmetic on every single turn — effectively free and always-on — versus the alternative of sampled LLM-judge evaluations, which cost per sample and typically run offline or async rather than in real time.

Quickstart

from horizon_monitor import FidelityMonitor
from datetime import datetime, timezone

monitor = FidelityMonitor()
session_id = monitor.new_conversation(metadata={"domain": "technical"})

result = monitor.process_turn(
    session_id,
    human_message="How does Python handle memory management?",
    agent_response="Python uses reference counting and a cyclic garbage collector...",
    timestamp=datetime.now(timezone.utc).isoformat(),
)

print(f"Fidelity:         {result.fidelity_score:.2f}")
print(f"Health:           {result.health_status}")
print(f"Circadian factor: {result.circadian_factor:.2f}")
print(f"Causal horizon:   {result.reachable_turns} reachable turns")
for event in result.events:
    print(f"  Event: {event.type} (confidence={event.confidence:.2f})")

Framework integrations

OpenAI SDK

from openai import OpenAI
from horizon_monitor import FidelityMonitor

monitor = FidelityMonitor()
session_id = monitor.new_conversation()
client = monitor.wrap(OpenAI(), session_id)

response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Tell me about quantum computing."}]
)

traj = monitor.get_trajectory(session_id)
print(f"Fidelity: {traj.current_fidelity:.2f}  T*: {traj.estimated_t_star}")

monitor.wrap() accepts custom timestamp and context providers for testing and replay.

Anthropic SDK

from anthropic import Anthropic
from horizon_monitor import FidelityMonitor

monitor = FidelityMonitor()
session_id = monitor.new_conversation()
client = monitor.wrap(Anthropic(), session_id)

response = client.messages.create(
    model="claude-3-5-sonnet-20241022",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Explain RLHF."}]
)

LangChain

from langchain_openai import ChatOpenAI
from horizon_monitor import FidelityMonitor
from horizon_monitor.integrations.langchain import HorizonCallback

monitor = FidelityMonitor()
session_id = monitor.new_conversation()
callback = HorizonCallback(monitor, session_id)

llm = ChatOpenAI(callbacks=[callback])
llm.invoke("Explain the CAP theorem.")
print(f"Fidelity: {callback.last_result.fidelity_score:.2f}")

OpenAI Agents SDK

from agents import Agent, Runner
from horizon_monitor import FidelityMonitor

monitor = FidelityMonitor()
session_id = monitor.new_conversation()
agent = Agent(name="assistant", model="gpt-4o-mini", instructions="You are helpful.")

for user_message in conversation:
    result = Runner.run_sync(agent, user_message)
    monitor.process_turn(session_id, human_message=user_message,
        agent_response=result.final_output, timestamp=datetime.now(timezone.utc).isoformat())

4D Spacetime Signals

"Spacetime" here is a metaphor, not physics. The relativity vocabulary (Minkowski interval, light cone, proper time) is design inspiration — it shaped which quantities we compute. Every signal below reduces to a standard information-theory or arithmetic measure on text embeddings and timestamps, listed in the Plain definition column. Nothing in Horizon's behavior or validation depends on the analogy being literally true, and the Lorentzian interval_class is emitted as descriptive metadata only — no event or score depends on it.

Every process_turn() returns a TurnResult with 32 fields across five signal families:

Core (always present)

Temporal (requires timestamp)

Pace (requires timestamp + turn ≥ 2)

Spacetime (requires timestamp + turn ≥ 2) — descriptive metadata only

Causal (requires timestamp)

Spatial (requires client_context)

16 Event Types (conversation plane)

All events default to observe mode (emitted, not acted on). Enable active mode via configure() once your event achieves ≥ 0.7 precision/recall on your domain.

Mission plane (Memento Mori)

"Progress is not a turn property — it is a calendar property."

The conversation plane answers is this dialogue degrading? The mission plane answers a different question: is this goal still moving, and against what clock? A month of individually healthy conversations that advance nothing is, to a conversation monitor, a month of perfect health.

The name is the design. Every store has exactly one root horizon — a finite end date you choose. Everything else hangs off it, so every day an item spends is a share of a budget that is visibly running out. Without a finite root, deferring work costs nothing and "later" is free forever. That is the failure this plane exists to make visible.

It is off by default. With no store configured, its six tools do not register and nothing in your integration changes.

The problem it catches

A task was given until 20 July. It is now 18 August and nobody has touched the mission since 2 July. A decision was parked "until things calm down" with a revisit date of 10 August that has quietly passed. The work has been sitting with one party for three weeks. None of that is visible in any conversation, in any tracker's status column, or in a model's context window — and each turn of each conversation about it looks healthy.

The mission plane reports it as: mission 78 days old, 47 days since progress, task lifespan expired, park 8 days overdue, currently blocked on operator for 21 days and counting.

Vocabulary

Everything is an item in a tree under the root horizon. There are eight kinds:

Two more terms appear in the outputs: a sojourn is one recorded stay in a stage (enter → exit), and the incumbent is the way you are working today, as opposed to a probe of some alternative.

Quickstart

export HORIZON_MEMENTO_STORE_PATH=~/.horizon/missions.db   # the default: one local file

A file-backed store is the right default, but it is the wrong choice on any host whose filesystem resets between deploys — the plane would look correct and silently forget everything, which is worse than not running at all. For those, point it at MySQL 8:

pip install "horizon-monitor[mysql]"
export HORIZON_MEMENTO_STORE_DSN='mysql://user:pass@host:3306/horizon'  # wins over _PATH
export HORIZON_MYSQL_SSL_CA=/path/to/server-ca.pem   # or ..._CA_B64 for a PEM in an env var

TLS verification is mandatory — the backend refuses to connect without a CA. Each API key maps to an assigned tenant id (scripts/provision_tenant.py), so rotating a key keeps that tenant's history; unknown or revoked keys get no mission access at all.

A clock is only as good as what reaches it, and a record that depends on remembering to write is worth nothing on the day you forget. So the plane can derive events from append-only sources you already produce:

python scripts/ingest_artifacts.py --store ~/.horizon/missions.db \
    --repo /path/to/repo --item-id <mission-id>

Each commit becomes an ARTIFACT event carrying the source's own provenance, and the event's valid_time is the commit's timestamp — not the moment you ingested it. Safe to run from cron: it dedupes on the source's native id and asks only for what is new.

Two things it will not do. It will not guess which mission an artifact belongs to — --item-id is required, and the adapter interface has no parameter capable of attaching one. And it will not judge what counts as progress. Those are yours.

Not sure what to register in the first place? Ask what your history suggests:

python scripts/ingest_artifacts.py --store ~/.horizon/missions.db \
    --repo /path/to/repo --propose

It reports the shape — how many artifacts, over what span, starting when — and proposes a created_valid equal to the earliest one. It proposes no title, because what the work is cannot be read off a commit log. Nothing is written; registering the mission is your call.

GitLocalAdapter is the reference implementation; trackers and mail metadata fit the same ArtifactAdapter interface.

from datetime import date, datetime, timezone

from horizon_monitor.memento import (
    EventKind, ItemKind, MementoConfig, MementoStore, evaluate,
)

store = MementoStore("missions.db")   # a real file; set this up once

root = store.register_item(
    kind=ItemKind.HORIZON, title="engagement horizon",
    created_valid=datetime(2026, 1, 1, tzinfo=timezone.utc),
    end_date=date(2030, 1, 1),
)
mission = store.register_item(
    kind=ItemKind.MISSION, title="ship-the-thing", parent_id=root,
    stall_days=14,                    # silence longer than this is a stall
    created_valid=datetime(2026, 6, 1, tzinfo=timezone.utc),
)
store.record_event(                   # progress: a side-effect of the work
    item_id=mission, kind=EventKind.PROGRESS,
    valid_time=datetime(2026, 7, 2, tzinfo=timezone.utc),
)

# The evaluation instant is always a parameter — the engine never reads a clock,
# so the same store at the same instant always yields the same report.
report = evaluate(
    store.snapshot(), datetime(2026, 8, 18, 12, tzinfo=timezone.utc), MementoConfig()
)

row = next(r for r in report.items if r.item_id == mission)
print(f"age:             {row.age_days} days")            # 78 days
print(f"since progress:  {row.days_since_progress} days") # 47 days
print(f"recording path:  {row.recording_path}")           # no recent work
print(f"horizon share:   {row.horizon_share:.4f}")        # 0.0595

The store is a real database, so run the setup once — registering a second root raises DuplicateRootError by design, which is the one-finite-root guarantee working, not a bug. For the full picture — an expired task, an overdue park, the blocking entity, a refused write and a fired signal — run examples/memento_mori_mission_clock.py (no arguments, no network, no API key; it uses a fresh temporary store each time).

What it measures

12 signal types (mission plane)

Separate from the conversation plane's 16 event types, not an extension of them. Each fires once on an edge — when its predicate becomes true — never again while the condition persists, and at most one new signal per turn, so a bad week cannot flood you. Tiers order that cap: P1 is time-critical, P2 structural, P3 informational.

These ride the existing process_turn contract for sessions bound with associate_mission. Every event carries plane: "mission", and the contract is deliberately loud — mission signals are surfaced to the operator with their numbers, unlike conversation signals, which apply silently. See agent rules for the block to paste into your host.

What it refuses

Accounting, never estimation. The engine never invents a duration, date, or amount:

• no forecasts, no completion predictions, no counterfactual "what the other path would have cost"

• no NPV/IRR/DCF, no discount rates, no currency conversion — money only ever multiplies measured time

• no p-values, confidence intervals, or sequential tests on path latencies: at single-operator sample sizes no dominance claim survives audit, so comparison is descriptive only

• no people analytics — entity latency is reported on functional slots; a person's wait is measured but never becomes a score, a ranking, or a resolvable identifier

• missing inputs degrade by omission with an explanatory field, never by substitution

Every row carries a derivation string spelling out the arithmetic it came from, and any summary statistic additionally carries the n it summarised. Identical store plus identical evaluation instant produces a byte-identical report.

Docs: product requirements · technical spec · agent rules · acceptance test plan

Configure

# Per-session override
monitor.configure(
    session_id=session_id,
    clarification_threshold=0.25,           # tighter D_JS gate
    event_modes={"alert.drift": "active"},  # activate one event
)

# Compound weight override
monitor.configure(
    fidelity_weights={"alpha": 0.35, "lambda_r": 0.12, "lambda_i": 0.28, "beta": 0.25},
    temporal_weights={"gamma": 0.08, "delta": 0.04},
    spacetime_coefficients={"alpha": 1.0, "beta": 1.0, "gamma": 0.8, "delta_st": 0.5},
)

Export

# JSON
result = monitor.export_to(session_id, target="json")

# LangSmith / Langfuse / OpenTelemetry / Arize
result = monitor.export_to(session_id, target="langsmith",
    connection={"api_key": "ls__..."})

Not yet published to PyPI — see Path 3 for a source install in the meantime.

pip install horizon-monitor[langsmith]   # or langfuse, otel, arize

Architecture

Input: plain strings (human_message, agent_response, optional timestamp, optional client_context)

Core pipeline (< 50ms on CPU):
  1. Embed both turns (local sentence-transformers, lazy-loaded)
  2–6.  IGT · D_JS · TWR · Bipredictability · Epsilon
  7. Temporal signals  — gap, retention, circadian, deictic
  8. Fidelity dynamics — composite score
  9. Health classification
 10. Pace signals       — velocity, acceleration
 11. Spacetime interval — ds² and interval class
 12. Causal reachability — light-cone membership
 13. Spatial signals    — device, location, frame shift
 14. Mode detection     — auto-classify conversation type
 15. Event evaluation   — 16 threshold checks
 16. Optional: SQLite persistence

Output: TurnResult dataclass (32 fields)

Design constraints (test-enforced):

• Zero LLM calls — pure arithmetic and local embeddings

• Zero external network calls by default — fully local

• Zero transitive framework dependencies in core

• < 50ms core pipeline on CPU — soft target (CI flags regressions past 50ms and hard-fails at 150ms)

• < 100MB memory for 100-turn conversations — hard-enforced at the claimed value

• All events observe-by-default — never interferes unless explicitly configured

Validation

What is proven, and what is not. Horizon's signals are correlational, in-domain measurements that track human quality ratings well. They are observability, not a proven outcome guarantee. Here is the honest status of each claim:

The four gates below pass on a labelled 5,602-record corpus (not bundled — see the evidence pack; scripts/build_validation_corpus.py regenerates a synthetic corpus that exercises the gate logic, not these exact numbers):

Cross-embedding stability: ρ_conv spread 0.026, ρ_turn spread 0.018 across three sentence-transformer backends (22M / 33M / 110M params). The fidelity signal lives in conversational structure, not in the embedding manifold. (Note: cross-embedding stability on the same corpus is distinct from cross-corpus OOD — first third-party run on MT-Bench pairwise labels gave ρ = 0.039; see evidence pack §Fix 4.)

Remediation gaps source: DESIGN_FIXES_redteam_remediation.md

Full evidence pack: docs/reviews/V0_2_0_EVIDENCE.md

Deployment

Self-hosted Docker (MCP server on port 3847)

cd deploy/docker
docker compose up

Horizon serves the MCP API via SSE. Point .cursor/mcp.json to http://localhost:3847/sse. The Dockerfile pre-caches the all-MiniLM-L6-v2 weights at build time — zero cold start.

Hosted (DigitalOcean App Platform)

The official hosted endpoint is live at https://horizon.leocelis.com. It runs on DigitalOcean App Platform (single instance, in-process session state — sessions do not survive a restart) and requires a Bearer token, rate-limited and isolated per key. See Path 1 above.

Development

git clone https://github.com/leocelis/horizon.git
cd horizon
python -m venv .venv && source .venv/bin/activate
pip install -r requirements-dev.txt

pytest tests/ -v                         # full suite
pytest tests/unit tests/integration tests/e2e -v   # fast path (~6 min)
ruff check src/ tests/
black --check src/ tests/
./scripts/compliance/check.sh              # EU AI Act offline gate

ComplyEdge TrustLint — EU AI Act

Horizon integrates ComplyEdge TrustLint on LLM-facing artifacts — same offline + runtime + trust pattern as IVD.

Integration guide: docs/integrations/COMPLYEDGE.md. CE adoption guide: oss-trustlint-adoption-guide.md.

Repository layout

horizon/
├── src/horizon/         # package source (PEP 517/518 src/ layout)
│   ├── engines/         # IGT, D_JS, TWR, coherence, fidelity, epsilon, mode
│   ├── spacetime/       # temporal, circadian, deictic, velocity, interval, light cone, spatial
│   ├── events/          # 16-event evaluator
│   ├── integrations/    # OpenAI, Anthropic, LangChain, export targets
│   ├── mcp/             # MCP server + CLI
│   └── storage/         # optional SQLite persistence
├── tests/               # unit / integration / e2e / perf / validation
├── examples/            # runnable framework demos
├── deploy/              # Procfile, build.sh, runtime.txt, docker/
├── docs/
│   ├── product/         # public product overview
│   ├── content/         # published pieces on conversation dynamics monitoring
│   ├── integrations/    # Cursor / Claude Desktop / Copilot setup guides
│   ├── cursor-rules/    # horizon-monitor.mdc (canonical Cursor agent rule)
│   ├── spec/            # HORIZON_TECH_SPEC.md + intent.yaml
│   └── reviews/         # E2E reviews, validation evidence
└── pyproject.toml

Background

Horizon's design was inspired by the Trans-Horizon Communication Protocol (THCP), a speculative framework that maps human–AI communication onto general-relativity metaphors. The five THCP "conjectures" are design intuitions, not proven laws — each is useful only because it pointed at a concrete, computable signal:

THCP is design motivation only — see docs/product/THCP_FIDELITY_MONITOR_PRD.md for the full conjecture-to-signal mapping.

Community

• Request alpha access: Open a Discussion →

• Ask a question: GitHub Discussions

• Bug reports: GitHub Issues

• Contributing: CONTRIBUTING.md

License

MIT — see LICENSE.

Legal

Performance claims: The +15.7% quality lift and 87% fewer hallucination events figures in this README are from synthetic, scripted controlled A/B scenarios with hand-tuned reference controllers — not production traffic and not the in-domain validation corpus (V1–V5 gates use a separate labelled set). Results may vary by domain, model, and deployment configuration. Do not use these figures in external marketing without conducting your own domain-specific evaluation. See LEGAL.md §5 for full scope and evidentiary basis.

High-stakes domains: Do not enable event types in active mode in healthcare, legal, financial, or emergency service contexts without domain-specific validation and human oversight. See LEGAL.md §4.