Source artifacts from the multi-model landing package (README + Super-Blueprint + ADR-0238/0239/0240 + goldtether/dynamic_manifold/surprise) are now the contractual surface, implemented on the live algebra/* kernel: - GoldTetherMonitor.residual/update/may_relax_hitl/force_reset/autonomy - signature_aware_pca (5×K, null-safe), conformal_procrustes, cartan_iwasawa_extract - surprise_residual (Minkowski) + dual_procrustes_surprise audit dict Package stubs (core.algebra.backend placeholders, scipy, identity-only Procrustes) are replaced with dual-corrected Cl(4,1) operators. ADRs match package decision language; docs/adr remains canonical with decisions redirects. 34/34 Third-Door tests + 7/7 ADR-0199 arena regression green.
87 lines
4.5 KiB
Markdown
87 lines
4.5 KiB
Markdown
# ADR-0238: GoldTether-Modulated Supervised Autonomy
|
||
|
||
**Status**: Proposed (acceptance path: tests green + Josh review)
|
||
**Date**: 2026-07-11
|
||
**Deciders**: Joshua Shay (CORE lead) + multi-model R&D chain
|
||
**Traceability**: Issue #11, parent Issue #10
|
||
**Related**: ADR-0055 (inter-session memory), ADR-0056 (contemplation C1), ADR-0080, GoldTether physics, Practice/Serve boundary, ADR-0199 (Arena GoldTether — distinct contract)
|
||
**Canonical path**: `docs/adr/` (redirect stub under `docs/decisions/`)
|
||
|
||
## Context
|
||
|
||
CORE must grow a forever-lived intelligence through experiences while remaining under hard safety control. The current HITL gate is correct but static. We need a continuous, geometry-native modulator that:
|
||
|
||
1. Measures coherence residual of the lived trajectory (GoldTether).
|
||
2. Dynamically raises or lowers the autonomy floor (pseudoscalar floor).
|
||
3. Only allows progressive reduction of HITL once the system has proven, through articulate reason + contemplation + successful epistemic elevation, that it can self-review safely.
|
||
4. Never permits unsupervised mutation of identity, admissibility regions, or safety invariants.
|
||
|
||
This is the controlled path from supervised to self-supervised to eventual self-authoring (the "grow out of the need for HITL" trajectory).
|
||
|
||
### Dual ontology (mastery refinement)
|
||
|
||
| Name | ADR | Contract |
|
||
|---|---|---|
|
||
| **Arena GoldTether** | ADR-0199 | independent truth for practice scoring (`is_correct` / `gold_answer`) |
|
||
| **Coherence GoldTether** | **this ADR** | field residual + dynamic floor + `supervised_autonomy_level` |
|
||
|
||
Shared metaphor; different modules. Never shadow `core.learning_arena.protocols.GoldTether`.
|
||
|
||
## Decision
|
||
|
||
Introduce a first-class **GoldTetherMonitor** that:
|
||
|
||
- Continuously computes the coherence residual \( R_{\text{GoldTether}}(t) = \| F(t) \cdot \widetilde{F}(t) - 1 \|_F \) (and higher-grade projections). Implemented as `algebra.versor.versor_unit_residual` dual-checked with `reverse(F)`.
|
||
- Maintains a **dynamic pseudoscalar floor** that rises only on proven epistemic elevation (successful contemplation + dual-corrected claim that reaches VERIFIED).
|
||
- Exposes a `supervised_autonomy_level ∈ [0.0, 1.0]` that gates B1 / R0 / R1 sequencing and the practice-vs-serve boundary.
|
||
- Implements a hard decay schedule and a fail-closed reset if residual ever exceeds ε_drift.
|
||
- Exposes `may_relax_hitl()` — hard gate; HITL override can never be disabled by the monitor itself.
|
||
- Exposes `force_reset()` for emergency fail-closed (HITL / safety pack only).
|
||
|
||
The monitor is pure geometry, replay-deterministic, and dual-corrected. It never samples, never uses learned weights for the gate itself.
|
||
|
||
### Named configuration
|
||
|
||
`epsilon_drift`, `max_history`, floor step / decay step (documented in module), optional practice/serve envelope helpers.
|
||
|
||
### Supervised blend (dual-correction geodesic)
|
||
|
||
When a blended field state is required:
|
||
|
||
```text
|
||
R = word_transition_rotor(source, target)
|
||
out = rotor_power(R, α) * source # Spin left-composition
|
||
assert versor_condition(out) < 1e-6
|
||
```
|
||
|
||
Sandwich conjugation is not the state geodesic (maps identity→identity). Left composition lands α=0→source, α=1→target.
|
||
|
||
## Consequences
|
||
|
||
### Positive
|
||
- Single forever-lived trajectory can now safely accumulate wisdom.
|
||
- HITL phase-out becomes measurable and gated by geometry, not calendar or vibes.
|
||
- Full auditability of every autonomy increase.
|
||
- Compatible with existing vault/CRDT, contemplation loop, and risk-reward physics.
|
||
|
||
### Negative / Risks
|
||
- Incorrect floor calculation could stall progress or (worse) over-grant autonomy. Mitigated by dual-correction + explicit residual tests + HITL override that can never be disabled by the monitor itself.
|
||
- Additional telemetry surface. Mitigated by making the channel first-class (`telemetry()` schema `goldtether_coherence_v1`).
|
||
|
||
### Neutral
|
||
- New module `core/physics/goldtether.py`.
|
||
- New ADR chain dependency for any future self-authorship work (ADR-0240).
|
||
|
||
## Implementation Notes
|
||
|
||
See `core/physics/goldtether.py` for the exact operator suite.
|
||
All public methods must enforce algebraic closure before returning.
|
||
The monitor is the single source of truth for "how much HITL can be relaxed right now".
|
||
|
||
## Validation
|
||
|
||
- Property tests: residual is always ≥ 0 and dual-corrected under reverse.
|
||
- Replay tests: identical field sequences produce identical autonomy levels.
|
||
- Lifelong curve: epistemic elevation events correctly raise the floor; residual breach forces fail-closed.
|
||
- Never allows autonomy > 0 while residual > ε_drift.
|
||
- `may_relax_hitl()` false until floor and autonomy thresholds met.
|