MECH-293 Waking Ghost-Goal Probe Search

Status (2026-04-27): IMPLEMENTED 2026-04-27. Substrate prerequisite SD-039 landed 2026-04-27; MECH-292 (ranked ghost-goal bank) landed 2026-04-27 and is the upstream substrate this consumer reads from. V3-EXQ-497 substrate-validation diagnostic 5/5 PASS (UC1 module surface; UC2 master-OFF bit-identical; UC3 ghost branch fires with n_ghost_admitted=4 and mean_goal_match_at_seed=0.998; UC4 hypothesis_tag preserved on ghosts and value-flat candidates remain default-clean; UC5 budget arithmetic clamp + bank-size cap + min-floor all hold). Behavioural validation (V3 full-completion gate, MECH-163 dual systems test) lives in V3-EXQ-495 and is gated on this substrate. This doc specifies the second consumer in the ghost-goal cluster: the proposal-generation hook that turns a ranked ghost bank into actual probe trajectories.

This is the substrate the V3 full-completion gate (V3-EXQ-495) calls as the PLANNED-arm proposal hook in MECH-163’s dual goal-directed systems test. Without MECH-293, the gate’s PLANNED arm has no mechanism distinct from the HABIT arm and the experiment is uninformative.

Retrieval-Cue Reframe (2026-05-19)

Under the cue-system reframe (ghost_goal_search.md Section 0), MECH-293 is the internally-triggered retrieval step: an internal goal/context cue reinstates a stored trace and re-pursues it – biologically licensed by Gelbard-Sagiv 2008 (internal cue reinstates the encoding-time trace ahead of report). Constraint-3 implication for the probe budget: the “how large a waking probe fraction before recovery turns into rumination” question is not a free knob – inhibition-only recoverability makes rumination the predicted equilibrium (Bouton 2004/2014; Maren 2013), so the probe fraction interacts with the open abandon-path question (Section 0.3) and must be validated against the Constraint-3 rumination signature, not tuned in isolation. Interpretive reframe only; status/phase unchanged; nothing queued.

Why this exists

The current waking proposer (HippocampalModule.propose_trajectories) generates value-flat candidates: CEM around the current terrain prior plus E1 world-prior conditioning. ARC-007 strict says this is correct — the hippocampus does not own value computation. But “value-flat over the current terrain” is a strong default that is biased toward CURRENT context. When the agent has unresolved goals attached to non-current regions (a blocked path, a deferred target, a remembered resource cache), the value-flat proposer will not propose probes there; the only mechanism that re-surfaces those regions is random exploration.

MECH-293 adds a second proposal branch: a minority budget of candidates seeded around the highest-priority entries in the MECH-292 ghost-goal bank. These probes are hypothesis-tagged (MECH-094) until they yield realized outcomes — they explore, they don’t commit the viability map.

In the dual-system framing of MECH-163: the value-flat HABIT arm scores trajectories already on the table; the PLANNED arm uses MECH-293 to put goal-relevant trajectories on the table that wouldn’t otherwise be there.

What this lands

Module-level extension to HippocampalModule 

ree_core/hippocampal/module.py:
  HippocampalModule.propose_trajectories(...) extended:

    candidates = self._propose_value_flat(z_world, e1_prior, ...)  # existing path

    if self.config.use_mech293_ghost_probes and self.ghost_goal_bank is not None:
      ghost_candidates = self._propose_ghost_seeded(
          current_z_goal=current_z_goal,
          n_total=self.config.n_candidates,
          ghost_fraction=self.config.mech293_ghost_fraction,
          z_world=z_world,
          e1_prior=e1_prior,
      )
      # Mix: keep n_total candidates total. Replace the bottom-ranked
      # value-flat candidates with ghost-seeded ones rather than expanding
      # the candidate pool. Preserves downstream E3 selection cost.
      candidates = self._mix_value_flat_with_ghost(candidates, ghost_candidates)

    return candidates


  HippocampalModule._propose_ghost_seeded(...) -> List[Trajectory]:
    """Generate ghost-seeded probe trajectories from the MECH-292 bank.

    1. n_ghost = max(1, int(round(n_total * ghost_fraction))) bounded by
       the number of available bank entries; if bank is empty, return [].
    2. Pull top n_ghost GhostGoalBankEntry items from rank(current_z_goal).
    3. For each entry, seed a CEM probe around the anchor's z_world rather
       than the agent's current z_world: action_object_mean is conditioned
       on (entry.anchor.z_world, e1_prior, residue_val_at_entry_z_world).
    4. Each generated trajectory is tagged hypothesis_tag=True (MECH-094)
       and labelled with provenance:
         trajectory.metadata = {
             "source": "mech293_ghost_probe",
             "anchor_key": entry.anchor.key,
             "ghost_priority": entry.ghost_priority,
             "goal_match": entry.components["goal_match"],
         }
    5. The trajectory is otherwise an ordinary CEM-refined sequence: E2
       rolls forward in action-object space, terrain scoring proceeds
       normally. The only differences are the seed z_world (anchor's,
       not current) and the hypothesis tag.
    """

Ghost fraction and budget

The minority budget is the gate against rumination. Bench too high and the proposer re-probes blocked goals at the expense of current-context trajectories. Bench too low and ghost probes never escape the random-exploration baseline.

Default: mech293_ghost_fraction = 0.2. So if n_candidates = 16, the proposer emits 13 value-flat plus 3 ghost-seeded. The fraction is exposed as a knob on HippocampalConfig so EXQ sweeps can probe the recovery-vs-rumination tradeoff.

Hypothesis tag preservation (MECH-094) 

ghost trajectories carry hypothesis_tag=True throughout:
  - E2 rollout: hypothesis_tag forwarded to predictions
  - Terrain scoring: residue updates suppressed during ghost-probe
    rollouts (existing MECH-094 path)
  - Anchor writes: any tick_anchor_set call triggered during the rollout
    inherits hypothesis_tag=True so the anchor write skips waking-stream
    payload (already enforced by SD-039 population layer's
    simulation_mode argument)
  - Viability-map updates: blocked

The ghost probe is a plan, not a commitment. If E3 selects a ghost probe and the agent acts on it, the actually-executed trajectory loses the hypothesis tag at the commit boundary and behaves like any other realized trajectory from that point forward (committed-trajectory recording, backward credit sweep, etc., all proceed normally per existing MECH-090 / MECH-290 paths).

Config wiring 

HippocampalConfig:
  use_mech293_ghost_probes: bool = False
  mech293_ghost_fraction: float = 0.2
  mech293_min_ghost_candidates: int = 1   # floor when bank has entries
  mech293_max_ghost_candidates: int = 8   # cap regardless of fraction
  mech293_replace_lowest_ranked: bool = True
    # True: ghost candidates replace bottom value-flat candidates,
    # preserving total candidate count.
    # False: ghost candidates are appended (raises n_candidates by
    # n_ghost). Used only for diagnostics / EXQ-487-style ablations.

REEConfig.from_dims():
  use_mech293_ghost_probes: bool = False  # master flag
  mech293_ghost_fraction: float = 0.2

Precondition (raised on __init__): use_mech293_ghost_probes=True requires use_mech292_ghost_bank=True AND use_sd039_anchor_payload=True AND use_anchor_sets=True. Without those upstream flags, the bank is empty and ghost probes are silent — the precondition makes the misconfiguration loud rather than silent.

Diagnostics

Per-tick propose_trajectories extension records on HippocampalModule._last_propose_diagnostics:

{
  "mech293_n_ghost_proposed": int,
  "mech293_n_ghost_admitted": int,    # post-mix
  "mech293_max_ghost_priority": float,
  "mech293_mean_goal_match_at_seed": float,
}

Captured in run-pack manifests by experiments that opt in.

What MECH-293 does NOT do

  • It does not commit. Ghost probes go into E3 the same way value-flat candidates do; if E3 ranks them low, they don’t leave the head.
  • It does not modify the bank ranking. MECH-292 is upstream.
  • It does not bypass MECH-094. Hypothesis tag stays True through the rollout; only commit / execution loses it.
  • It does not violate ARC-007 strict. The proposer is still proposing trajectories without a value head; the ghost-seeded ones are biased BY LOCATION (the anchor’s z_world) not by an internal value computation. Goal-match enters via MECH-292’s ranking, which consults stored payloads — that ranking lives outside HippocampalModule.

Falsifiable signature

Versus a random-probe control with the same total candidate count:

  • Ghost-seeded probes should show higher mean goal_match of the seed location than random probes (the bank is doing its job).
  • After a goal becomes blocked at a previously-pursued location, ghost-seeded recovery (mean steps to re-acquire benefit) should be faster than random-probe recovery in ≥ 4 / 7 seeds.
  • Revisitation rate of fully incoherent stale regions (high staleness, low goal-match, low recoverability — anchors that the bank’s goal_match_floor filters out) should be lower under MECH-293 than under uniform random probing. This is the rumination-suppression signature.

The behavioural test that exercises both is V3-EXQ-495 paradigm A (detour) — see v3_exq_495_mech163_planned_system_gate.py.

MECH-094

Three call-site checks already exist in the substrate:

  1. _propose_ghost_seeded sets hypothesis_tag=True on every emitted trajectory before returning.
  2. The CEM rollout path in HippocampalModule already forwards hypothesis_tag through E2 calls (existing MECH-094 wiring).
  3. SD-039 population layer’s build_goal_payload(..., simulation_mode=True) returns None, so any anchor write triggered during a ghost-probe rollout writes goal_payload=None (the rollout cannot retroactively author a populated payload from its hypothetical state).

No new MECH-094 plumbing required.

Phased training

None. Pure routing logic; ghost-probe generation reuses the existing CEM machinery, just with a different seed z_world.

Validation experiment

V3-EXQ-XXX (placeholder; assign next free number after MECH-293 lands) diagnostic with five sub-tests:

  • UC1 module_importable
  • UC2 master_off_no_op (propose_trajectories returns the same candidate count and provenance as pre-MECH-293 baseline)
  • UC3 ghost_branch_fires (with non-empty bank, at least one trajectory has metadata["source"] == "mech293_ghost_probe"; bank-empty -> zero ghost trajectories)
  • UC4 hypothesis_tag_preserved (every ghost trajectory carries hypothesis_tag=True; any anchor write triggered during the rollout records goal_payload=None)
  • UC5 budget_respected (n_ghost matches ceil(n_candidates * ghost_fraction) bounded by [min_ghost_candidates, max_ghost_candidates] and by the bank size)

Behavioural validation = V3-EXQ-495 V3 full-completion gate.

Dependencies

Claim Role
SD-039 provides per-anchor payload (transitive via MECH-292)
MECH-292 provides the ranked bank _propose_ghost_seeded consumes
ARC-007 strict constraint: no value head; MECH-293 stays compliant by routing through MECH-292’s external ranking
ARC-018 hippocampal trajectory proposal / evaluation loop being modified
ARC-032 goal-biased sequence generation (MECH-293 is one instantiation)
MECH-089 theta-packaged waking E3 updates (architectural context)
MECH-094 hypothesis-tag invariant (preserved by call-site scoping)
MECH-269 anchor / probe substrate (transitive via MECH-292)
MECH-291 mode-sensitive sequence generator framing (one generator across waking and offline modes; MECH-293 is the waking arm)

What’s next

  1. Land MECH-292 (mech_292_ghost_goal_bank.md).
  2. Land MECH-293 per this doc.
  3. Run V3-EXQ-495 V3 full-completion gate (ree-v3/experiments/v3_exq_495_mech163_planned_system_gate.py). The PLANNED arm uses use_mech293_ghost_probes=True; HABIT arm uses use_mech293_ghost_probes=False with goal_weight > 0 (EXQ-327 setup); ABLATED arm has neither. Paradigms A (detour) and B (novel context) run in a single script.

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