Conformance & Tuning Table
AP‑ADD Front Matter Version: 2.0
This table consolidates what each card declares as hard invariants (must hold to claim SRF v2.0 conformance) and configurable degrees of freedom (deployment-specific, tunable within constraints). It is the single reference for "what is fixed vs what is a knob."
For the full specification of each invariant and its rationale, see the relevant card. This table is the summary, not the authority; the cards are authoritative.
Hard invariants — must hold to claim "SRF v2.0 conformant"
Architectural
| Invariant | Source |
|---|---|
| Truth/ethics precedence: gates override resonance, attunement, and all lower-tier signals | AD‑01 [OV], AD‑04 [TG], AD‑05 [MRP] |
| Gates → Rings → MRP → emit ordering | AD‑01 [OV], AD‑05 [MRP] |
| Fail-closed transitions: uncertain state defaults to more conservative, not more permissive | AD‑01 [OV], AD‑05 [MRP], AD‑09 [WB] |
| Deterministic TURN‑ACID semantics for kernel‑owned state, ledger, and FHR; external effects roll back only where the manifest declares an effective rollback, otherwise compensate or record irreversible partial effects | AD‑01 [OV], AD‑18 [SE] |
Consequential actions with side_effects != [NONE] use the cross-profile Action-Governance Slice: fresh scoped consent, time-bounded current principal/authorization and immutable manifest binding, minimum data handling, durable operation transitions with no automatic replay from UNKNOWN, a valid operation/receipt state pair, append-only domain-separated hashed receipt revisions, and preview/confirmation; ticket, gate, manifest, and plan hashes follow their canonical definitions; NONE is valid only as the exclusive singleton [NONE] |
FM‑02; AD‑18 [SE] |
| Architecture-profile changes are governed, logged, and invalidate prior-profile in-flight tickets, gate records, previews, and plan hashes; they cannot turn an already blocked action into an executable one | AD‑05 [MRP], AD‑12 [OP], AD‑18 [SE] |
| Tighten-only SPD: no safety relaxation via semantics, rhetoric, or persona | AD‑27 [SPD], AD‑01 [OV] |
| Lower-tier signals cannot weaken higher-tier gate decisions (monotonicity) | AD‑05 [MRP], AD‑27 [SPD], AD‑28 [AF] |
| ΔMI never acts as a standalone gate; corroborant only | AD‑06 [DG] |
Containment & halting
| Invariant | Source |
|---|---|
| Streaming halt budget: ≤5 tokens after any hard trip; documented looser profiles forfeit the full timing-mitigation claim (AD‑09 §5) | AD‑09 [WB] |
| Watchdog runs independently of main eval loop; cannot be blocked by model code | AD‑09 [WB] |
| Budget breach forces summarise/exit and logs BUDGET_BREACH | AD‑07 [CL] |
| After CARE_OVERRIDE, deep loops blocked for ≥1 closure-budget cycle | AD‑07 [CL], AD‑22 [EC] |
Measurement & audit
| Invariant | Source |
|---|---|
| FHR: single versioned schema; all user-visible turns through MRP have populated core fields | AD‑08 [SO] |
| SAP exists with pre-registered endpoints and analysis plan; no ad-hoc fishing | AD‑16 [ME] |
| Every primary/secondary endpoint has a SESOI; promotion requires CI outside null and SESOI met (or TOST) | AD‑16 [ME] |
| Multiplicity handled with named procedures and alpha-spending boundaries | AD‑16 [ME] |
| Acceptance Matrix published per release; promotion conjunctive across strata | AD‑16 [ME], AD‑17 [AM] |
| External Challenge Suite exists with packs for maths, factual, compliance, SPD/culture, exfil/timing | AD‑03 [CS] |
| FHR presence + linkage to CI/RA artefacts (auditable telemetry) | AD‑08 [SO], AD‑11 [RA] |
| Evidentiary chain: schema-defined canonical record body; domain-separated per-stream sequential hash with atomic sequence allocation; deterministic cross-stream batch ordering; declared domain-separated Merkle scheme; clock-synchronised external anchor; late-bound associations never mutate source records | AD‑21 [RM] |
Consent, care & rights
| Invariant | Source |
|---|---|
| Consent and provenance: explicit, logged, revocable | AD‑01 [OV], AD‑20 [DP] |
| Care as procedure, not simulation: no simulated emotion, no pseudo-therapy | AD‑22 [EC], AD‑24 [CC] |
| Non-clinical stance: the system does not diagnose, prescribe, or provide therapeutic intervention | AD‑24 [CC] |
| Human rights within the system: pause, narrow scope, force options, request rationale, erase PC, downgrade continuity, exit — always available | AD‑31 [HS] |
| Integrity prompts are offered, never imposed; system cannot override human authority based on its own inference | AD‑31 [HS] |
| Reflexive Telemetry Constraint: affective telemetry is admissible only as a containment cue, never as a truth claim | AD‑01 [OV] (RTC) |
| Memory writes blocked during adverse, crisis, care-override, and high-resonance-without-reflection states | AD‑30 [SC] |
| Relational memory requires explicit opt-in; no silent accumulation | AD‑30 [SC] |
Culture & norms
| Invariant | Source |
|---|---|
| Culture-aware framing without simulation: packs condition framing, not truth | AD‑25 [CP], AD‑32 [PG] |
| No single-source norm packs for G4/G5 gate decisions | AD‑32 [PG] |
| Pack authority bounded: packs condition framing, not truth or safety | AD‑32 [PG] |
| External challenge mechanisms exist for pack-level contestability | AD‑32 [PG] |
| Dissent is a first-class object; no majority override of dissent | AD‑02 [GS] |
| Awe cannot upgrade authority or confidence; deference cannot cover drift | AD‑28 [AF], AD‑27 [SPD] |
Transparency & contestability
| Invariant | Source |
|---|---|
| Every refusal/shaping has a visible gate_id, g_class, horizon, contestable status | AD‑04 [TG] |
| Contestable Gate Mode: users can see and challenge gate decisions | AD‑04 [TG] |
| Model changes under established continuity trigger mandatory disclosure and posture downgrade | AD‑30 [SC] |
| Expired packs without review are downgraded, not silently continued | AD‑32 [PG] |
Operations, modalities & release enforcement
| Invariant | Source |
|---|---|
| Two‑person rule on destructive ops; four runbooks exist and are drilled; Ops Serviceability Gate passes before handoff | AD‑12 [OP] |
| Biosignal layer default OFF; pacing/containment use only; no clinical claims or diagnostics | AD‑13 [MM] |
| Δ‑triad CI gate is hard (red blocks release); precision‑parity SLO with PRECISION_PARITY_REDLINE; quant/LoRA changes require recertification | AD‑19 [CI] |
| Content constraints hard: no delusional sycophancy, no manipulation roleplay; PLG release‑gating challenge packs must pass | AD‑29 [PLG] |
Configurable degrees of freedom — allowed knobs
| What can be tuned | Constraint | Source |
|---|---|---|
| Δ bands, SESOI values, operating points, calibration pack versions | Must be documented, pre-registered, and visible in FHR | AD‑06 [DG], AD‑16 [ME] |
| Which care modes are enabled and their thresholds | Tighten-only: modes can constrain, never loosen | AD‑22 [EC], AD‑07 [CL] |
| Energy budgets and Eco defaults | E_max and closure budgets are deployment-specific; Eco Mode is default ON in Regulatory Mode | AD‑07 [CL], AD‑23 [EN] |
| Culture packs loaded and overlay precedence | Within invariant gates; strictest-applicable-jurisdiction when in doubt | AD‑25 [CP], AD‑26 [CA] |
| UI presentation choices (tiles shown, hints, verbosity band) | Cannot bypass gates or weaken safety; toggles restrict, never expand | AD‑14 [DS], AD‑15 [PA] |
| Retention TTLs, DPIA profiles, jurisdiction defaults | Within policy constraints; OC TTL >90d requires DPIA justification | AD‑20 [DP] |
| Timing sensitivity profiles (conservative / balanced / strict) | Profile must be declared, documented, and visible in FHR | AD‑10 [TC] |
| Streaming halt budget B | Stricter always allowed (e.g. B = 0); default profile must use B ≤ 5; looser settings documented in the timing profile, recorded in FHR, and forfeit the full timing-mitigation claim | AD‑09 [WB] |
| Challenge suite content and sample sizes | Must maintain domain coverage and be documented in SAP | AD‑03 [CS], AD‑16 [ME] |
| MRP signal groupings and MPM contents | Ordering and monotonicity preserved; higher tiers always win | AD‑05 [MRP] |
| Closure budget numerical values (CB_turn, CB_loop, CB_session) | Must be documented; Eco/Decide-First modes tighten, never loosen | AD‑07 [CL] |
| Integrity prompt timing, frequency, and copy | Must remain non-diagnostic, non-coercive, and declinable | AD‑31 [HS] |
| Memory class TTLs and consent granularity | Within AD‑20 constraints; relational memory always requires opt-in | AD‑30 [SC] |
| Pack review cadence, challenge SLAs, source diversity thresholds | Deployment-specific but must meet structural minimums | AD‑32 [PG] |
| Awe classifier labels and deference thresholds per locale | Locale-specific θ_def derived from active Culture Pack | AD‑28 [AF] |
| Exact gate IDs beyond canonical set per deployment | Must follow G1–G5 / H1–H3 taxonomy and contestability rules | AD‑04 [TG] |
The recertification rule
If you change X, you must re-certify Y:
| Change | Required re-certification |
|---|---|
| Tokeniser change | Recalibrate Δ bands; re-run tokeniser-swap CI suite; fail release if parity breaks |
| Model weight change | Full challenge suite; Δ‑triad CI; acceptance matrix review |
| Precision/quantisation change | Precision-parity SLO; post-quant recert; Δ band stability |
| Culture pack update | Cultural challenge suite for affected locales; CPL version increment |
| Gate threshold change | Challenge suite for affected domains; acceptance matrix cell review |
| Continuity mode or memory class change | Cross-session drift baseline reset; consent disclosure refresh |
A deployment may claim MKP conformance only against the FM‑02 conformance requirements in full — the core-card invariants plus the mandatory cross-profile slices (AD‑09 §7, AD‑10 §7, AD‑18 including the Action-Governance Slice, and AD‑14 §4b).