SD-039 Dual-Trace Anchor Goal-Snapshot Payload

Status (2026-04-27): IMPLEMENTED. Both the substrate-side surface (2026-04-26) and the module-level write-site population layer (2026-04-27) have landed. V3-EXQ-494 6/6 PASS confirms the falsifiable signature (goal-relevant vs goal-irrelevant dissociation) and dual-trace preservation of the populated payload.

Why this exists

MECH-269’s dual-trace anchor substrate (Phase 2 ii) preserves z_world and active/inactive status. MECH-284’s region-staleness scalar tells us which regions are unresolved. Neither tells us whether the unresolved region still corresponds to a live goal.

This is the gap the ghost-goal cluster (MECH-292 ranked ghost-goal bank, MECH-293 waking ghost-goal probes, ARC-060 hybrid field+bank architecture) needs closed before any of those mechanisms can rank or probe usefully. Without a stored goal payload, MECH-292 has only staleness to rank by, which collapses to “more stale = higher priority” and produces rumination over hopeless traces rather than intelligent re-probing of blocked goals.

SD-039 is the substrate change that lets the broader cluster discriminate “this stale region is still tied to my current goal” from “this stale region is just unresolved background.”

What this lands

Substrate (this session)

A compact motivational payload attached to each anchor at write, remap, or invalidation time. The payload survives mark_inactive, so inactive anchors remain queryable as ghost-goal traces.

ree_core/hippocampal/anchor_set.py:
  @dataclass class AnchorGoalPayload:
    z_goal_snapshot: Optional[torch.Tensor] = None
    wanting_strength: float = 0.0
    arousal_tag: float = 0.0
    last_vs: Optional[float] = None
    staleness_at_write: Optional[float] = None
    payload_written_step: int = 0

  Anchor.goal_payload: Optional[AnchorGoalPayload] = None
  Anchor.goal_match(current_z_goal) -> float  # clipped non-negative cosine

  AnchorSet.write_anchor(..., goal_payload=None)
  AnchorSet.mark_inactive(..., goal_payload=None)  # refresh on invalidate
  AnchorSet.reset_region(..., goal_payload=None)   # refresh on both traces
  AnchorSet.consume_boundary_events(..., goal_payload=None)
  AnchorSet.query_by_goal_match(current_z_goal, threshold=0.0,
                                 scale=None, active_only=False)
                                 -> List[Tuple[Anchor, float]]

Config flag: AnchorSetConfig.use_sd039_anchor_payload (default False). The flag is the substrate-side toggle. With False, callers omit goal_payload (or pass None), and behaviour is bit-identical to pre-SD-039. The flag becomes the module-level conditional when the write-site population layer lands.

Refresh-on-invalidate semantic

The design choice in the claims.yaml SD-039 entry: payload should reflect the most recent wanting/arousal at the point the trace is preserved. The substrate implements this by:

mark_inactive(..., goal_payload=p) writes p onto the anchor BEFORE inactivating it. Existing payload is replaced; nothing is cleared on inactivation (the entire point of dual-trace preservation).
reset_region(..., goal_payload=p) writes p onto BOTH the outgoing trace (cause-of-blockage payload) AND the new active anchor (new motivational identity).
consume_boundary_events(..., goal_payload=p) forwards p to each per-event write_anchor call so anchors installed in response to a boundary tick share the current motivational context.

Query helper

query_by_goal_match is the substrate hook MECH-292 will consume. It scans the active + inactive dual-trace pool, scores each anchor against the supplied current_z_goal via Anchor.goal_match, and returns non-zero matches sorted descending. SD-039 itself does NOT rank; the ranked-bank computation ghost_priority ~ wanting * goal_match * staleness * recoverability lives in MECH-292’s module.

active_only=True restricts to the active half (legacy behaviour). threshold=-1.0 includes every payload-bearing anchor regardless of match (diagnostic use).

Module-level write-site population (this session, 2026-04-27)

HippocampalModule.build_goal_payload(latent_state, goal_state, residue_field, bla_output, current_step, simulation_mode=False) constructs an AnchorGoalPayload once per agent.sense() tick from the current waking-stream signals:

z_goal_snapshot     <- goal_state.z_goal.detach().clone()  (None when inactive)
wanting_strength    <- ResidueField.evaluate_valence(z_world)[..., VALENCE_WANTING].mean()
arousal_tag         <- bla_output.arousal_tag              (0.0 when BLA off)
last_vs             <- mean(per_stream_vs.values())        (None when empty)
staleness_at_write  <- max(staleness_accumulator.snapshot().values())
                                                           (None when accumulator off)
payload_written_step <- agent._step_count                  (anchor_set._tick fallback)

build_goal_payload returns None (skipping population entirely) when:

the AnchorSet substrate is disabled,
AnchorSetConfig.use_sd039_anchor_payload is False (master flag OFF),
simulation_mode=True (MECH-094 gate – replay / DMN paths must NOT populate payloads from waking signals).

REEAgent.sense() calls build_goal_payload after the MECH-269 Phase 1 update_per_stream_vs step and forwards the result as goal_payload=... to both tick_anchor_set (the boundary-event-driven write/remap site) and apply_invalidation_broadcasts_to_regions (the MECH-287 broadcast-driven mark_inactive site). Hysteresis-fired mark_inactive inside tick_hysteresis does not refresh the payload – the prior payload is preserved as the cause-of-blockage trace per dual-trace semantics.

use_sd039_anchor_payload is wired through REEConfig.from_dims(...) and propagated to config.hippocampal.anchor_set.use_sd039_anchor_payload, so experiments can enable the substrate via the standard factory.

Items still deferred to follow-on sessions

MECH-292 ranked ghost-goal bank. Priority computation, bank size bound, decay over time. query_by_goal_match is the substrate input; the ranking lives in MECH-292.
MECH-293 waking ghost-goal probes. Probe-budget allocation in HippocampalModule.propose_trajectories() plus MECH-094 guardrail (ghost probes do not update the viability map until realized).
ARC-060 hybrid field+bank architectural framing. Documentation of the continuous-wanting-field plus discrete-ghost-bank duality.

Why the substrate landed before the population layer

The substrate is small, well-scoped, and testable in isolation. The contract suite (10 tests) covers the surface completely without touching agent state. A clean substrate landing meant MECH-292 / MECH-293 could be designed against a stable API while the population layer was being wired.
The population layer touches multiple consumer files. Each anchor write site (boundary events, broadcast-driven invalidation) calls into consume_boundary_events / mark_inactive; the population layer (this session) threaded a single shared goal_payload argument through both paths.

Falsifiable test (V3-EXQ-494, 6/6 PASS 2026-04-27)

From the SD-039 claim entry:

after reward relocation or path blockage, inactive anchors on the formerly valid approach path should retain non-zero goal-match with the current z_goal while unrelated stale anchors do not.

V3-EXQ-494 exercises this with six sub-tests:

UC1: substrate-side classes / methods accessible (PASS).
UC2: master OFF, anchors carry no payload, query returns [] (PASS).
UC3: master ON, anchors carry populated payload, query returns non-zero goal_match (PASS, 7/7 anchors match >= 0.99).
UC4: dual-trace preservation – inactive anchors retain payload across mark_inactive (PASS, 6 inactive + 1 active, all carry payload).
UC5: goal-relevant vs goal-irrelevant dissociation (PASS): mean goal_match on Phase A (goal-inactive) anchors = 0.0; mean on Phase B (goal-active) anchors = 0.998; 3/3 above 0.3.
UC6: MECH-094 simulation gate – replay-path writes do NOT populate payloads from waking signals (PASS).

UC5 is the substrate-level falsifiable signature. Behavioural validation (does ghost-goal querying improve recovery from blocked goals?) belongs to MECH-292 / MECH-293’s behavioural EXQ.

Biological grounding

Lit-pull synthesis: evidence/literature/targeted_review_ghost_goal_search/.

Berridge 1998, Barch 2010 — persistent wanting / goal representations exist as a discrete substrate and are not subsumed by the continuous valence field.
Mattar & Daw 2018 — utility-prioritised replay; gives the normative argument for goal-relevance ranking over pure recency.
Pfeiffer & Foster 2013 — goal-biased forward path search during waking quiescence.
Gillespie 2021 — broad, non-current, non-recent trace reactivation; closest neural match for ghost-goal bank prioritisation.
Muessig 2019 — one sequence generator across waking and offline modes; supports a single substrate that MECH-292 / MECH-293 will share.
Berkowitz 1989 — frustration / unresolved goal persistence; the behavioural phenotype the substrate makes computationally tractable.

Backward compatibility

AnchorSet.write_anchor / mark_inactive / reset_region / consume_boundary_events all accept goal_payload=None as default. Existing call sites that omit the kwarg are unchanged. Anchors carry goal_payload=None. query_by_goal_match returns [] when no payload-bearing anchors exist. The full contract suite (164/164) plus preflight (7/7) pass with the SD-039 surface in place but unused.