Attune Context-Layer Spec — 6 New Structural Models

Purpose: define exactly how the 6 new models (Framework Meta-Model, 3 Horizon Model, ISO 31000 Risk Process, Change Integration Model, I→T→E→Comb-Shaped Growth Model, Understanding Learners / Honey & Mumford + Integrated L&D) plug into Attune’s existing scoring pipeline — as context/enrichment layers injected before the 19 utterance-level frameworks score, not as competing scorers.

This matches the real architecture shown on the Model card & arch page: a 3-encoder fusion transformer (text/audio/video) → cross-modal fusion → framework-grounded heads (Attune · 8 competencies, Attune Classified, psych-safety framework, NVC · GROW, Custom). The 6 new models become an additional pre-fusion context head whose output is passed alongside the transcript into the framework-grounded heads, the same way a “Custom” head already works today.

1. Where it sits in the pipeline

TEXT / AUDIO / VIDEO  →  [existing 3 encoders]  →  Cross-modal fusion transformer
                                                          │
                     ┌────────────────────────────────────┴───────────────────────┐
                     │                                                            │
        [NEW] Structural Context Head                          Framework-grounded heads (existing)
        ────────────────────────────                           ──────────────────────────────────
        • Framework Meta-Model (schema validator)                Attune · 8 competencies
        • 3 Horizon Model (H0–H3 classifier)          ──context──▶  Attune Classified
        • ISO 31000 (risk-stage classifier)            injection    psych-safety framework
        • Change Integration Model (context tagger)                NVC · GROW
        • I→T→E→Comb Growth Model (maturity classifier)            Custom
        • Understanding Learners (style + L&D-stage classifier)
                     │                                                            │
                     └────────────────────────────────────┬───────────────────────┘
                                                            ▼
                                              OUTPUTS (existing, per utterance & session)
                                              + 4 NEW session/team-level tags (below)

The structural context head does not produce its own 0–100 utterance score. Instead it emits tags that get concatenated into the framework-grounded heads’ input context — the same mechanism the model card already documents for “Framework citation” (e.g. “Attune Comp #7 · Evokes”). This is additive, not a retrain of the existing 355M-param text encoder.

2. New session-level output fields (added to the existing output list)

Per the model card’s current output list:

· 10 dimension scores (0–100)
· Intuition confidence (0–1)
· Utterance tags w/ framework citation
· Intuition read (natural language)
· Intent-impact gap %
· Next-best-action suggestion
· Bias flags with severity
· Modality-agreement breakdown

New fields appended:

· Horizon tag (H0/H1/H2/H3) — 3 Horizon Model
· Risk-stage tag (identify/analyze/evaluate/treat/monitor, or "none") — ISO 31000
· Change-context tag (swift/disruptive vs. planned/deliberate) — Change Integration Model
· Growth-shape tag (I/T/E/Comb) with trajectory delta vs. last session — I→T→E→Comb model
· Framework-schema trace (technique→practice→model→principle chain) — Meta-Model, powers "Framework citation" links
· Learning-style mix (Activist/Reflector/Theorist/Pragmatist %) + L&D-stage tag — Understanding Learners model

3. Concretely, how each improves an existing framework’s accuracy

3b. Known interaction requiring reconciliation: Growth Stage vs. Learning Style

The I→T→E→Comb Growth Model (#24) and Understanding Learners (#25) both write to the same “Next-best-action suggestion” output field, but along independent axes: - #24 answers: what depth/breadth stage is this person at (I/T/E/Comb)? - #25 answers: how does this person learn best (Activist/Reflector/Theorist/Pragmatist)?

These can genuinely conflict — e.g. a Comb-shaped senior leader can still be a strong Activist learner who wants hands-on piloting, not a theory briefing, even though their growth stage says “senior.” The next-best-action generator must treat #24 and #25 as two inputs to reconcile (e.g. “Comb-shaped + Activist → sponsor them to pilot a cross-domain initiative hands-on” rather than a generic Comb-shaped suggestion), not let one silently override the other. This is called out explicitly so it isn’t built as an implicit precedence rule during implementation without a product decision.

4. What this does NOT claim

• This spec does not retrain the 355M-param text encoder, the 90M-param audio encoder, or the 24M-param vision encoder — those require real annotated training data (2.4M sessions, 1,240 annotators) that only EQCompanion’s ML team holds.
• The claimed F1/Cohen’s κ improvements above are directional hypotheses based on the failure modes EQCompanion’s own model card already documents — they must be validated against the same held-out 12,000-session eval set the model card references, not assumed.
• Every framework here is being added the same way the model card labels its own numbers: “illustrative model card … not an audited or independently verified specification.” Recommend running this through the same 12,000-session eval + external senior-credentialed coach review before claiming production-ready lift.

5. Rollout decision (updated)

Stakeholder decision: all 6 models ship Live · citable from day one. This overrides the earlier staged-rollout recommendation (Research preview → Live). Implementation steps:

1. Encode the 6 models as structured JSON (see model_studio_extension.json) — schema, signals, weight defaults. Done — all 6 now flagged Live · citable.
2. Add them to Model Studio as a separate “Structural & Strategic Lenses” section, not merged into the 19-item weighted list (see UI recommendation in the JSON file).
3. All 6 go live simultaneously — no staged Research-preview phase.
4. Validate against the 12,000-session held-out eval set as a fast-follow, not a pre-launch gate.
5. Resolve the #24/#25 reconciliation logic (section 3b) before the next-best-action generator ships either model’s output to end users — this is an implementation decision, not a data-validation one, so it isn’t blocked by the eval timeline.

⚠️ Flagged risk, accepted by stakeholder decision

Framework Meta-Model and I→T→E→Comb Growth Model are low-risk to ship live — they map onto data Attune already collects (framework citations, trust trajectory) and don’t introduce new unvalidated inference.

3 Horizon Model, ISO 31000 Risk Process, Change Integration Model, and Understanding Learners are going live without the held-out-eval validation the model card holds the other 19 frameworks to. Concretely this means: their horizon/risk/change-context/learning-style tags will influence Cohere’s team-health aggregation, Meridian’s session-prep view, and Attune’s shadow-coach whispers and next-best-action suggestions in production before anyone has measured their F1/Cohen’s κ against the 12,000-session external-coach-annotated eval set. Recommend: - Tag these 4 outputs internally as “provisional” in logs/telemetry so a fast-follow eval can be diffed against real usage data. - Set a concrete eval checkpoint (e.g. first 2–4 weeks of production data) rather than letting “fast-follow” drift indefinitely. - Do not publish F1/κ numbers for these 4 in the public Model card until that eval completes — the model card’s existing credibility rests on every published number being externally verified; publishing unvalidated numbers for the new 4 would undermine that standard even though the shipping decision itself is fine.

Full validation plan: see fast_follow_eval_plan.md for the complete fast-follow eval design covering these 4 models — scope/out-of-scope, why the existing 12k held-out set can’t be reused as-is per model, per-model ground-truth annotation design (including the flagged Change Integration Model data gap), metrics & pass thresholds (including the new MAE-based metric for Understanding Learners’ style-mix output), sample size & annotator plan, a 5-week timeline anchored to the “2–4 weeks” checkpoint above, the per-model (not all-or-nothing) go/no-go decision rule, live-production-data handling during the gap, and the reporting format matching the Model card page’s Evaluation table.