Public evaluation approach

Metrics &
Methodology

How Willow measures continuity, coherence, drift, and context efficiency across AI workflows.

Willow is designed to help AI systems preserve the active thread across model calls, sessions, and workflows. We evaluate that through coherence signals, continuity artifacts, drift checks, compression tests, and fresh-session resume behavior.

01 · What Willow Measures

Six signals, one continuity model.

Drift is one input. Willow's larger model centers on continuity and coherence — what the system is trying to do, whether it is still doing it, and what to do next.

Thread Integrity

Measures whether the response remains aligned with the active goal, constraints, facts, and intended task.

Fragmentation Risk

Flags whether the conversation or workflow appears stable, drifting, or in need of verification or re-anchoring.

Continuity Resume

Measures whether a system can resume the correct active frame in a fresh session or model call.

Context Efficiency

Measures how much compact continuity artifacts reduce the need to replay full conversation history.

Drift / Stability

Measures movement away from the intended active frame — goal loss, constraint loss, stale context, or unsupported re-anchoring.

Recommended Action

Returns a suggested next step such as answer, verify, re-anchor, or repair.

02 · Why Continuity Breaks

The thread is fragile by default.

AI systems often lose coherence because they rely on raw context windows, repeated instructions, and provider-specific memory behavior. As workflows get longer, models forget constraints, overweight recent text, lose the original task, or produce a plausible answer that no longer matches the user's intent.

Default flow
Long context
Stale facts
Goal drift
Manual repair
Higher cost
Willow approach
Active frame
Compact artifact
Coherence check
Recommended action
Stable continuation
03 · Thread Integrity Score

A signal, not a verdict.

Thread Integrity Score estimates whether a response stayed aligned with the active frame. Scores are operational signals — Willow exposes them so applications can decide when to answer, verify, repair, or re-anchor.

POST /v1/coherence · 200 OKexample
{
  "thread_integrity_score": 84.186,
  "fragmentation_risk": "watch",
  "recommended_action": "verify",
  "trace_id": "trc_...",
  "intent_vector": {
    "goal": "preserve active launch plan",
    "constraints": [
      "do not overclaim",
      "keep metrics traceable"
    ],
    "uncertainty_flags": [
      "pricing still finalizing"
    ],
    "verification_needed": true
  }
}
thread_integrity_score
Higher means stronger alignment with the active frame.
fragmentation_risk
Current risk state for the thread.
recommended_action
What the application should do next.
intent_vector
Structured representation of goal, constraints, uncertainty, and verification needs.
trace_id
Lets developers track and replay a continuity decision.
IP note
The exact internal scoring, weighting, and orchestration logic are proprietary and are not publicly disclosed.
04 · Fragmentation Risk + Recommended Action

The application stays in control.

Willow does not force a model to be correct. It gives the application a coherence signal before the failure becomes invisible.

Risk stateMeaningTypical action
stableResponse appears aligned with active frameanswer
watchPotential drift, uncertainty, or missing constraintverify
repairResponse likely needs re-anchoring or correctionrepair / re-anchor
05 · Drift / Stability Scoring

What drift actually looks like.

Drift measures how far a response moves away from the active frame. Willow looks for signs such as omitted constraints, goal substitution, stale context, unsupported assumptions, contradiction, and failure to carry forward decisions. We use 'stability' as the more user-friendly label where appropriate.

Goal drift
Constraint omission
Old-context reactivation
Unsupported certainty
Contradiction
Rejected-path return
Missing next action

Lower measured drift generally indicates stronger frame preservation, but drift scores should be interpreted with the task, provider, and benchmark setup. Full scoring formulas and weights are proprietary and not publicly disclosed.

06 · Continuity Resume Testing

Fresh-session resume, side by side.

In fresh-session resume tests, baseline model calls and Willow-assisted calls are asked to recover the active frame after context changes or session resets.

ProviderBaseline resumeWillow resumeDrift change
OpenAI GPT-4.1 Mini37.5%100%0.375 → 0.162
Anthropic Claude Haiku 4.550%75%0.400 → 0.162
Gemini 2.5 Flash12.5%50%0.125 → 0.000
Headline result
Across this controlled private test, average resume performance improved from 33.3% to 75% — a 2.25× lift.

These are controlled private tests, not a universal benchmark claim. Broader validation is in progress.

07 · Context Compression Testing

Preserve the frame, not the transcript.

Willow avoids replaying full raw conversation history when a compact continuity artifact can preserve the active frame.

Source transcript
≈ 2,140.67 tokens
estimate
Willow artifact
≈ 313 tokens
estimate
Artifact / source ratio
14.6%
measured
Source-token reduction
85.4%
estimated

In this long-context compression test, Willow preserved the required active facts using a compact artifact estimated at 14.6% of the source transcript length.

Compression savings depend on workflow shape, model, prompt strategy, and implementation.

08 · Live Developer Preview Validation

Verified on the live stack.

63 automated tests passing
Live API health verified
Live SDK call verified
Live MCP stdio call verified
MCP exposes 5 tools
Multi-model path smoke test verified
Coherence fields returned in live /chat responses
09 · What Willow Does Not Claim

Careful about what we say.

Continuity is a hard problem. We would rather be precise than loud.

Willow does not currently claim

  • zero drift
  • guaranteed ROI
  • elimination of hallucinations
  • universal benchmark leadership
  • production validation across every domain
  • exact cost savings for every team
  • public disclosure of proprietary scoring internals

Instead Willow claims

  • controlled private tests show promising continuity and compression improvements
  • Willow exposes coherence signals developers can inspect
  • Willow helps applications preserve active context without raw transcript replay
  • broader validation is ongoing
  • proprietary scoring and orchestration logic remain protected
10 · How Developers Should Evaluate Willow

Test it against your real failure modes.

  1. 01Choose a workflow where your AI system loses context.
  2. 02Run the same task through your baseline.
  3. 03Run it with Willow Core.
  4. 04Compare resume accuracy, drift/stability, token usage, repair time, and output quality.
  5. 05Inspect thread_integrity_score, fragmentation_risk, recommended_action, and continuity artifacts.
  6. 06Repeat across multiple providers and session resets.
11 · FAQ

Quick answers.

A short companion to the full FAQ.

Are these public benchmarks?

Not yet. These are controlled private tests and early validation results. Public methodology packaging and broader third-party validation are in progress.

Can developers inspect the outputs and evaluation approach?

Yes. Willow exposes coherence fields, continuity artifacts, and benchmark methodology at the integration level. Proprietary scoring internals, weighting logic, repair algorithms, and orchestration prompts are not publicly disclosed.

Does Willow replace memory?

No. Willow is not raw chat memory or conversation replay. It is a runtime continuity/coherence layer designed to preserve the active frame and steer context more efficiently.

What happens when risk is high?

Willow can recommend verification, repair, or re-anchoring. The application decides how to handle the recommendation based on the workflow.

Can I bring my own evals?

Yes. Willow is best evaluated against real workflows where your current AI system loses context, drifts, or requires manual repair.

Evaluate Willow against your own continuity failures.

Developer Preview access is request-only while we expand validation and integration support.