MECH-292 Ranked Ghost-Goal Bank

Status (2026-04-27): IMPLEMENTED. Pure-arithmetic derived view over the SD-039 dual-trace anchor pool. Module: ree-v3/ree_core/hippocampal/ghost_goal_bank.py. Wired into HippocampalModule behind HippocampalConfig.use_mech292_ghost_bank (default False). Substrate prerequisite SD-039 (anchor goal-snapshot payload + population layer) landed 2026-04-27 with V3-EXQ-494 6/6 PASS. Validation diagnostic V3-EXQ-496 PASSed 5/5 sub-tests on 2026-04-27. MECH-292 is the first downstream consumer of the SD-039 payload and the prerequisite for MECH-293 (waking ghost-goal probe search) which wires HippocampalModule.propose_trajectories() to consume the bank.

Why this exists

A good agent should not treat all stale regions equally. Some stale regions are still tied to currently-live goals (a blocked path to food I still want); others are merely unresolved background (a corner I crossed once and never revisited). Ranking by staleness alone (the MECH-284 path) collapses these into one bucket and produces rumination over hopeless traces. Ranking by goal-match alone misses the staleness signal that tells the agent something needs re-probing here.

MECH-292 closes this gap with a derived view over the dual-trace anchor pool: a ranked bank where each entry combines stored goal payload, current motivational state, region staleness, and a coarse recoverability estimate.

This is the substrate MECH-293 will consume to seed waking ghost-goal probes and that the V3 full-completion gate (V3-EXQ-495) will exercise as the proposal-generation hook for the PLANNED arm of MECH-163’s dual goal-directed systems.

Retrieval-Cue Reframe (2026-05-19)

The combined cue-system literature pull (evidence/literature/targeted_review_ghost_goal_search/, 2026-05-18) recasts this bank as the cued-recall step of a content-addressed retrieval system: the current z_goal/context is the retrieval cue, the SD-039 payload is the stored trace, rank() is recall, and goal_match is the cue-to-trace match score (not a priority weight). See ghost_goal_search.md Section 0 for the full synthesis. Three structural constraints land specifically on MECH-292:

  • Constraint 1 (Smith & Vela 2001): the cue is z_goal-cosine only; a context channel from the already-stored-but-unused payload fields (arousal_tag, last_vs, cause) should be added, but gated to yield to a strong goal_match, not summed with fixed weight (outshining). Not yet implemented; not queued.
  • Constraint 2 (Flagel 2011; Corbit & Balleine 2005/2011): the additive ghost_priority already implemented here (w_w*wanting + w_m*goal_match + ...) is biologically correct and is hereby ratified; the product form in the old ghost_goal_search.md sketch is superseded. wanting and goal_match are dissociable channels with independent failure modes – the additive form is required, not incidental.
  • Constraint 3 (Bouton 2004/2014; Maren 2013; Jasinska 2014): recoverability (currently clamp[0,1](last_vs)) must NOT become a cutoff that integrates suppression/staleness evidence toward excluding an anchor – accumulated “hopelessness” evidence provably does not justify retirement, and incubation can invert the staleness sign. The goal_match_floor rumination guard remains valid as a cue-relevance gate (an anchor with no matching cue is not recalled); it is not a hopelessness gate and must not be repurposed as one. An explicit abandon path is the deliberate biology-divergence point and is an open, user-gated substrate question (see ghost_goal_search.md Section 0.3).

Scope: interpretive reframe + revised falsifiable signature only. No substrate change in this pass, no new claim, no experiment queued, MECH-292 status/phase/confidence unchanged.

What this lands

New module 

ree_core/hippocampal/ghost_goal_bank.py

Pure-arithmetic, non-trainable. No new state beyond a small per-tick cache; the anchor pool itself remains the source of truth.

@dataclass
class GhostGoalBankConfig:
    # Master flag wired through HippocampalConfig.use_mech292_ghost_bank.
    # Bank is a derived view -- per the MECH-292 claim notes, no separate
    # persistent store. Default OFF preserves bit-identical behaviour.

    # Ranking weights: ghost_priority = w_w * wanting
    #                                 + w_m * goal_match
    #                                 + w_s * staleness
    #                                 + w_r * recoverability
    # All clamped to non-negative; sum need not equal 1 (the absolute
    # magnitude is irrelevant; only ordering is consumed downstream).
    wanting_weight: float = 1.0
    goal_match_weight: float = 1.0
    staleness_weight: float = 0.5
    recoverability_weight: float = 0.5

    # Floor on goal_match before an anchor is admitted to the bank at all.
    # Anchors with stored z_goal_snapshot=None or with cosine below this
    # floor are treated as goal-irrelevant and excluded.
    goal_match_floor: float = 0.05

    # Cap on bank size returned per call. None -> no cap.
    top_k: Optional[int] = 32

    # Recoverability: if last_vs is None, treat as recoverable (1.0).
    # Otherwise recoverability = clamp_[0,1](last_vs).
    # last_vs near 1.0 = the region was confidently grounded when written;
    # near 0.0 = the region was already failing groundedness at preservation
    # time, so re-probing is unlikely to recover it.
    default_recoverability_when_unknown: float = 1.0

    # Filter: include inactive (dual-trace preserved) anchors? Default True.
    # MECH-293 ghost-goal probes work primarily over inactive traces, but
    # diagnostic / replay-prioritisation consumers may want active too.
    include_inactive: bool = True
    include_active: bool = False
    scale: Optional[str] = None  # optional scale filter ("fast", "slow")


@dataclass
class GhostGoalBankEntry:
    anchor: Anchor                  # reference to the source anchor
    ghost_priority: float           # composite rank
    components: Dict[str, float]    # per-term breakdown for diagnostics


class GhostGoalBank:
    """MECH-292: derived ranking view over the dual-trace anchor pool.

    Stateless across calls beyond a small diagnostics cache. Each rank()
    call walks the anchor pool, scores each anchor, and returns a sorted
    list. Implementation is intentionally a derived view, not a persistent
    store: SD-039 already preserves the per-anchor payload; MECH-292 just
    arranges the existing data.
    """

    def __init__(self, config: GhostGoalBankConfig, anchor_set: AnchorSet,
                 staleness_accumulator: Optional[StalenessAccumulator] = None):
        self.config = config
        self.anchor_set = anchor_set
        self.staleness_accumulator = staleness_accumulator
        self._last_diagnostics: Dict[str, Any] = {}

    def rank(self, current_z_goal: Optional[torch.Tensor]) -> List[GhostGoalBankEntry]:
        """Return the ranked ghost-goal bank for the current z_goal.

        Anchors with no stored z_goal_snapshot (payload-less or written
        outside any goal regime) are excluded by the goal_match_floor.
        Returns an empty list when current_z_goal is None or no anchor
        clears the floor.
        """

    def get_diagnostics(self) -> Dict[str, Any]:
        """Last-call summary: n_candidates_scanned, n_admitted, max_priority,
        mean_priority, plus per-component sums for sanity checks.
        """

    def reset(self) -> None:
        """Per-episode reset of the diagnostics cache."""

Hippocampal integration 

ree_core/hippocampal/module.py:
  HippocampalModule.__init__(...):
    if config.use_mech292_ghost_bank:
      assert config.use_anchor_sets,        "MECH-292 requires MECH-269 Phase 2 ii"
      assert config.use_sd039_anchor_payload, "MECH-292 requires SD-039 payload"
      self.ghost_goal_bank = GhostGoalBank(
          config=config.ghost_goal_bank_config,
          anchor_set=self.anchor_set,
          staleness_accumulator=self.staleness_accumulator,
      )

  HippocampalModule.rank_ghost_goals(current_z_goal) -> List[GhostGoalBankEntry]:
    if self.ghost_goal_bank is None: return []
    return self.ghost_goal_bank.rank(current_z_goal)

  HippocampalModule.reset() extension:
    if self.ghost_goal_bank is not None:
      self.ghost_goal_bank.reset()

Config wiring 

ree_core/utils/config.py:
  HippocampalConfig.use_mech292_ghost_bank: bool = False
  HippocampalConfig.ghost_goal_bank_config: GhostGoalBankConfig = factory()

REEConfig.from_dims():
  use_mech292_ghost_bank: bool = False  # master flag
  mech292_wanting_weight: float = 1.0
  mech292_goal_match_weight: float = 1.0
  mech292_staleness_weight: float = 0.5
  mech292_recoverability_weight: float = 0.5
  mech292_goal_match_floor: float = 0.05
  mech292_top_k: Optional[int] = 32

Ranking formula

For each anchor a in the dual-trace pool that clears goal_match_floor:

wanting       = a.goal_payload.wanting_strength
goal_match    = a.goal_match(current_z_goal)           # SD-039 cosine helper
staleness     = staleness_accumulator.lookup(a.region_key) if accum present
                else (current_tick - a.last_accessed) / staleness_normaliser
recoverability = (a.goal_payload.last_vs                   if not None
                  else config.default_recoverability_when_unknown)

ghost_priority = ( config.wanting_weight       * wanting
                 + config.goal_match_weight    * goal_match
                 + config.staleness_weight     * staleness
                 + config.recoverability_weight * recoverability )

staleness_accumulator is the MECH-284 source of truth when present; the fallback ensures MECH-292 still works in configurations where MECH-284 is disabled (rare but possible — the anchor’s own last_accessed tick gives a weaker but workable proxy).

The ranking is utility-like (Mattar & Daw 2018) but not pure low-V_s chasing. The goal_match_floor is the architectural commitment that pure rumination is excluded: an anchor with no stored goal payload OR with a goal_match below the floor is invisible to MECH-292 entirely. This is the substrate-level guard against the depressive-rumination failure mode flagged in the ghost-goal-search lit-pull.

What MECH-292 does NOT do

  • It does not own the per-anchor payload (SD-039 does).
  • It does not generate trajectories (MECH-293 does).
  • It does not gate write paths (MECH-094 / MECH-261 do).
  • It does not modify the anchor pool — it is a read-only ranking view.

These boundaries keep MECH-292 small and falsifiable in isolation.

Falsifiable signature

Extended by the 2026-05-19 reframe: in addition to the paired-comparison signature below, the three dissociated states of Constraint 2 must be inducible (intact match / absent vigor; nonspecific vigor / absent match; match present but un-actionable) and the Constraint 1 context-gating prediction must hold (a strong direct goal_match must not be displaced by the context channel). See ghost_goal_search.md Section 0.2.

In a reward-relocation or blocked-corridor task:

  • Anchors from the now-obstructed but still-valued path should rank above equally stale but goal-irrelevant anchors.
  • Specifically: in a paired comparison of a goal-relevant inactive anchor vs a goal-irrelevant inactive anchor with matched staleness, the goal-relevant anchor’s ghost_priority should be strictly higher in ≥ 4 / 7 seeds, with the gap dominated by the goal_match component.
  • The recoverability term should keep anchors that have already failed V_s grounding from rising to the top despite high staleness.

MECH-094

The bank is read-only over the existing anchor pool. There is no write path to gate. SD-039’s call-site scoping handles waking-vs-replay payload provenance; MECH-292 inherits whatever provenance the source anchors carry.

Phased training

None. Pure arithmetic, no trainable parameters.

Validation experiment

V3-EXQ-496 diagnostic with five sub-tests (PASSed 5/5 on Mac 2026-04-27, ~39s):

  • UC1 module_importable
  • UC2 master_off_no_op (agent.hippocampal.ghost_goal_bank is None)
  • UC3 ranking_fires (after a goal-active episode, rank() returns a non-empty bank with ghost_priority strictly decreasing)
  • UC4 goal_irrelevant_excluded (Phase A / Phase B episode design parallel to V3-EXQ-494 UC5 — Phase A anchors below goal_match_floor are absent from the bank)
  • UC5 component_breakdown_consistent (per-entry components dict sums to ghost_priority within float-tolerance)

Behavioural validation lands as part of V3-EXQ-495 (V3 full-completion gate, MECH-163 dual-system distinction) when MECH-293 consumes the bank.

Dependencies

Claim Role
SD-039 provides the per-anchor goal_payload MECH-292 ranks over
MECH-216 populates wanting_strength (predictive wanting signal)
MECH-230 provides the structured z_goal latent the cosine query uses
MECH-269 (Phase 2 ii) dual-trace anchor substrate; bank pool source
MECH-269 Phase 1 / 2 iii per-stream / per-region V_s feeding last_vs
MECH-284 region staleness; preferred staleness source over the local proxy

What’s next

  1. Land MECH-292 substrate per this doc.
  2. Land MECH-293 (mech_293_ghost_goal_probe_search.md) — the first consumer of rank_ghost_goals().
  3. Run V3-EXQ-495 (V3 full-completion gate) — MECH-293 is the proposal-generation hook the PLANNED arm calls; MECH-292 is what makes the call non-empty.

REE is developed by Daniel Golden (Latent Fields). Apache 2.0.