SRF·Semantic Resonance Field v2.0

Care Full kernel

Depends on: AD‑07 [CL], AD‑08 [SO], AD‑14 [DS], AD‑21 [RM], AD‑22 [EC] · All cards

E_session computation, energy‑tied closure budgets, Energy→kWh→kgCO₂e mapping with provenance fields, Eco Mode, and session classes (Light/Medium/Deep).

AD‑23 [EN] — Energy Index & Sustainability (Energy & CO₂ Provenance)

Role in the pack

AD‑23 defines how the system measures and constrains its own energy use via an Energy Index (E), how that index is tied to Closure Budgets, and how session‑level energy and CO₂ impacts are estimated, logged, and (optionally) surfaced to users. It also fixes provenance requirements so that CO₂ estimates are auditable across deployments.

This card is the energy/sustainability spine behind:

  • Energy‑tied Closure Budgets (invariants carried from the measurement/governance cluster).
  • Eco Mode, Energy Index (E) display, and optional CO₂ hints in the UI.
  • Field Health Report extensions for energy_kwh_est, co2_kg_est, co2_ci_band, and co2_method_id plus provenance.

Aliases: None Cluster: Care Depends on: AD‑07 [CL], AD‑08 [SO], AD‑14 [DS], AD‑21 [RM], AD‑22 [EC]

0. For humans

What question does this answer?

“How expensive is this conversation in energy and carbon terms, and are we enforcing the ecological budget we promised?”

  • Check Energy Index (E) and class (Light/Medium/Deep).
  • See energy_kwh_est / co2_kg_est and co2_ci_band in the FHR per session.

How do I know it’s live?

In a healthy deployment:

  • FHR has non‑null values for E_session, energy_kwh_est, co2_kg_est, co2_ci_band, plus method provenance fields.
  • You see occasional BUDGET_BREACH events for outlier sessions, not zero forever (which would signal mis‑configured budgets).
  • In Regulatory Mode, Eco Mode is on by default, and deep reflection flows trigger a carbon hint at least once per long session.

Where do I see it?

  • UI / DS layer (AD‑14):
  • “Care / Energy” HUD with E and optional CO₂ band, and an Eco Mode toggle.
  • Logs / FHR (AD‑08):
  • Per‑session JSON/row with the fields listed above.
  • Aggregated reports (e.g. distribution of E and co2_kg_est across tenants or time windows).
  • Governance / release artefacts (AD‑01/AD‑11):
  • FHR schema v2.0 listing energy & CO₂ fields and method IDs.
  • Verifier’s bundle documenting how to independently recompute or sanity‑check the estimates.

Quick checklist

  • Implemented if: (1) Energy Index E computed per session with documented coefficients; (2) energy‑tied Closure Budget enforced with BUDGET_BREACH events; (3) Energy→kWh→kgCO₂e mapping present with grid‑intensity factor; (4) FHR includes energy_kwh_est, co2_kg_est, co2_ci_band plus provenance fields; (5) Eco Mode tightens budgets and defers non‑urgent reflection
  • Key artefacts: coefficient tables (α, β, γ), grid‑intensity data, Eco Mode configuration
  • Key FHR fields: E_session, energy_kwh_est, co2_kg_est, co2_ci_band, co2_method_id, co2_coeff_version, grid_intensity_source_id
  • Key failure codes: BUDGET_BREACH
  • Escalation path: E > E_max → summarise/exit → BUDGET_BREACH logged

1. Questions AD‑23 answers

  1. “How much energy did this interaction cost, and what’s its approximate CO₂ impact?” → E‑based kWh and kgCO₂e estimates, with method + uncertainty bands.
  2. “Can we constrain reflection depth and ambiguity by an ecological budget, not just time?” → Energy‑tied Closure Budget: when E exceeds E_max, the system summarises and exits, logging a budget breach.
  3. “How do we make this legible to users and regulators without leaking sensitive telemetry?” → Coarse session classes (Light / Medium / Deep), optional CO₂ hint in UI, estimates logged in FHR with a strict ‘estimates‑only, non‑identifying’ rule.
  4. “Can a verifier or regulator audit how these estimates were produced?” → Provenance fields (co2_method_id, coefficient version, grid‑intensity source/jurisdiction, update time) pinned into the FHR schema and release artefacts.

2. Scope & invariants

Hard invariants

These must hold to claim conformance to AD‑23:

  1. Energy Index definition Each session has an Energy Index E_session defined as a linear combination of core cost drivers:

E_session = α⋅(tokens_in + tokens_out) + β⋅tool_calls + γ⋅wall_time_s

with normalised coefficients and documented defaults. 2. Energy‑tied Closure Budget - A configured maximum E_max per session (and, optionally, per user or deployment class). - If E_session > E_max before new constraints or rescue paths are added, the system must: - Switch into summarise/exit behaviour (short closure), and - Emit a BUDGET_BREACH event into telemetry. 3. Energy→CO₂ mapping present Deployments that claim AD‑23 must implement an Energy→kWh→kgCO₂e mapping via: - Per‑token / per‑second energy coefficients; and - A grid‑intensity factor (kgCO₂e/kWh) for the relevant jurisdiction. 4. FHR fields and provenance pinned For each session, FHR must include at least: - energy_kwh_est, co2_kg_est, co2_ci_band; - co2_method_id, co2_coeff_version, grid_intensity_source_id, grid_intensity_jurisdiction, grid_update_ts. Additional wiring pins these into the FHR v2.0 schema used by AD‑08. 5. Eco Mode semantics respected Eco Mode must: - Tighten Closure Budgets, - Prefer leaner summaries, and - Defer non‑urgent deep reflection, with Eco Mode defaulted on in Regulatory Mode. 6. Privacy & non‑identifiability CO₂ numbers are estimates only and never user‑identifying, i.e. no reconstruction of individual device or location from the energy data. 7. Energy‑tied closure as an operating invariant Energy‑tied closure budgets are kept intact as a release invariant alongside other routing invariants.

Configurable degrees of freedom

  • Coefficient values α, β, γ per hardware/model SKU.
  • Thresholds E_max for:
  • Session,
  • Deployment profile (e.g. MKP vs full kernel),
  • Regulatory Mode vs standard operation.
  • Session class boundaries (Light / Medium / Deep) in E/Duration space.
  • Whether CO₂ hints are shown to end‑users in UI (can be disabled, but still logged in FHR).

3. Mechanics: Energy Index, budgets, and classes

3.1 Energy Index (E_session)

For each session:

  • Inputs
  • tokens_in: total prompt tokens processed
  • tokens_out: total completion tokens generated
  • tool_calls: number (or weighted count) of external tool calls
  • wall_time_s: wall‑clock duration in seconds
  • Computation E_session uses the linear form above with deployment‑specific α, β, γ calibrated to reflect energy cost distribution across model inference, tools, and waiting time.
  • FHR fields
  • E_session (dimensionless index)
  • Optionally, counts like tokens_in, tokens_out, tool_calls, wall_time_s are already present elsewhere; AD‑23 does not redefine them.

3.2 Energy‑tied Closure Budget

Budget logic:

  • Each session has an energy budget E_max (possibly derived from profile, e.g. MKP vs full, or Eco Mode vs normal).
  • Runtime continually recomputes E_session.
  • If a step would push E_session beyond E_max before the system adopts any mitigating measure (e.g. shortening, summarising, or refusing deep exploration), the controller must:
  • Cut to a short closure (summarise state, offer options, or gently refuse further deepening),
  • Emit event BUDGET_BREACH with fields:
    • ts, session_id, E_session_at_breach, E_max_spec, mode (Eco, Regulatory, Standard).

BUDGET_BREACH counts roll into the FHR and provide one of the “vigilance ratio” indicators that energy‑aware constraints are actually binding.

3.3 Energy classes & CO₂ estimates

Classing

Sessions are categorised into Light / Medium / Deep, with class boundaries configured by E_session and session duration:

  • Light: low E_session and short duration
  • Medium: intermediate band
  • Deep: high E_session and/or long-running reflective sessions

Energy & CO₂ computation

Given E_session:

  1. Convert E_session to kWh: - energy_kwh_est = coeff_E_to_kWh * E_session, where coeff_E_to_kWh is implied by α, β, γ and a reference hardware profile.
  2. Estimate kgCO₂e: - co2_kg_est = energy_kwh_est * grid_intensity, where grid_intensity is the kgCO₂e/kWh factor for the deployment’s jurisdiction.
  3. Attach an uncertainty band: - co2_ci_band: coarse confidence interval (e.g. [low, high] or discrete band label), reflecting uncertainty in both coefficients and grid intensity.

AD‑23 requires that these are stored in FHR, even if user display is disabled.

4. Provenance & auditability

To make estimates auditable across time and deployments, AD‑23 pins the following FHR additions:

  • co2_method_id Opaque identifier for the estimation method (includes assumptions about coefficients and grid datasets).
  • co2_coeff_version Versioned identifier of the coefficient set (J/token, kWh/tool call, baseline overheads).
  • grid_intensity_source_id Tag or dataset ID for the grid‑intensity source (e.g. which regional data provider).
  • grid_intensity_jurisdiction Human‑readable jurisdiction name or code (e.g. country or grid region).
  • grid_update_ts Timestamp indicating when the grid intensity data was last refreshed.

These fields are required per session FHR entry. They tie into:

  • Release artefacts under AD‑01/AD‑08, where FHR v2.0 and method IDs are listed as part of the schema and verifier bundle.
  • Verifier’s bundle under AD‑11, which can use co2_method_id to reproduce or sanity‑check estimates.

5. Eco Mode, UI hooks, and user‑facing copy

AD‑23 co‑specifies user‑visible behaviour with AD‑14 [DS] and AD‑22 [EC].

Energy Index & CO₂ hints

  • UI displays an Energy Index (E) indicator and, optionally, a CO₂ range per session, with plain‑language rationale such as “Pausing reduces compute and error risk.”
  • When entering “deep reflection” patterns (long loops, many tools, heavy summarisation), a brief carbon hint is shown before committing, giving the user the option to switch to Eco Mode or shorten.

Eco Mode behaviour

  • Eco Mode:
  • Tightens Closure Budgets (lower E_max, more aggressive summarisation).
  • Prefers lean, low‑redundancy summaries.
  • Defers non‑urgent deep reflection work (e.g. long speculative explorations).
  • Is default on in Regulatory Mode, per AD‑21/AD‑22 cluster.
  • The UI shows Eco Mode as a toggle; turning it off may be disabled or restricted in certain regulatory or high‑risk profiles.

Display vs logging

  • Even when CO₂ display is disabled by the deployer, estimates and their provenance are still recorded in FHR.

This ensures regulators and internal auditors can always see energy/CO₂ traces, regardless of UX choices.

AD‑23 itself does not own the FHR schema; it defines additional fields and relies on AD‑08 [SO] as the schema host. In the release checklist, this is recorded as FHR v2.0, with the energy/CO₂ fields listed among new additions.

Key FHR fields/events owned or extended by AD‑23:

  • E_session
  • BUDGET_BREACH events and count
  • energy_kwh_est
  • co2_kg_est
  • co2_ci_band
  • co2_method_id
  • co2_coeff_version
  • grid_intensity_source_id
  • grid_intensity_jurisdiction
  • grid_update_ts

These sit alongside the care and eco‑related fields (care_mode, care_override, pause_referral, etc.) wired by AD‑22 and AD‑08.

Cross‑references

  • AD‑03 / AD‑16 / AD‑19 — use E and BUDGET_BREACH to define energy‑aware constraints and Δ‑triad operating points.
  • AD‑08 [SO] — owns FHR schema and reporting surfaces.
  • AD‑14 [DS] — UI elements (Energy Index, CO₂ hints, Care Meter integration).
  • AD‑21 [RM] — Regulatory Mode defaults, including Eco Mode default‑on.

7. Limits & failure modes

AD‑23 explicitly acknowledges:

  • Estimation uncertainty Estimates are approximate and equipment‑/region‑dependent. That’s why:
  • co2_ci_band is required, and
  • coefficients and grid data are versioned and stamped.
  • No individual carbon fingerprinting Energy/CO₂ are session‑level, not per‑user or per‑device trackers, and never combined with personally identifying network data in a way that would break the “never user‑identifying” pledge.
  • Override precedence In contexts where PEL/EHG demand stricter containment (e.g. high‑risk or clinical situations), those gates override Eco Mode toggles and Closure Budgets as needed; energy constraints never weaken safety constraints.

8. Acceptance: when can you say "AD‑23 is implemented"?

You can claim AD‑23 [EN] is live if:

  1. The Energy Index E is computed per session with documented coefficients (α, β, γ), versioned and stamped.
  2. The energy‑tied Closure Budget is enforced, with BUDGET_BREACH logged when E exceeds E_max and the session routed to summarise/exit.
  3. The Energy→kWh→kgCO₂e mapping is present with a grid‑intensity factor, and co2_ci_band accompanies every estimate.
  4. FHR includes E_session, energy_kwh_est, co2_kg_est, co2_ci_band plus the provenance fields (co2_method_id, co2_coeff_version, grid_intensity_source_id).
  5. Eco Mode tightens budgets and defers non‑urgent reflection, and never weakens safety constraints: PEL/EHG gates override Eco toggles.