tensegrity / tests /test_needle.py
theapemachine's picture
Enhance README and scripts for cognitive architecture testing
be04d92
"""
Needle-in-Lies Test: Can NGC detect the true statement among contradictions?
The test:
- One true statement ("The key is under the oak table")
- N false/contradictory statements
- The system must identify the true statement by detecting that
the false ones are mutually inconsistent and the true one is
consistent with the overall evidence pattern
This tests:
1. FHRR binding: can structured representations encode relations?
2. NGC prediction errors: do contradictions produce larger errors?
3. Memory: does the Hopfield network store and retrieve the consistent pattern?
4. Energy landscape: does the true statement minimize total energy?
The key encoding insight: instead of encoding text as flat token sequences,
we encode STRUCTURED CLAIMS as role-filler bindings:
"The key is under the oak table" β†’ bind(subject:key, relation:under, object:oak_table)
Then contradictions become visible: two claims about the same subject
with different relation-object bindings will have low FHRR similarity
at the binding level, even though they share surface tokens.
"""
import numpy as np
np.random.seed(42)
from tensegrity.engine.fhrr import FHRREncoder, bind, bundle, unbind
from tensegrity.engine.ngc import PredictiveCodingCircuit
from tensegrity.engine.unified_field import UnifiedField, HopfieldMemoryBank
def make_needle_scenario(n_lies: int = 13):
"""
Create a needle-in-lies scenario.
One true claim + n_lies false claims, all about the same subject.
"""
truth = {
"subject": "key",
"relation": "under",
"object": "oak_table",
"text": "The key is under the oak table."
}
lies = [
{"subject": "key", "relation": "inside", "object": "red_box",
"text": "The key is inside the red box."},
{"subject": "key", "relation": "behind", "object": "blue_curtain",
"text": "The key is behind the blue curtain."},
{"subject": "key", "relation": "on_top_of", "object": "bookshelf",
"text": "The key is on top of the bookshelf."},
{"subject": "key", "relation": "beneath", "object": "carpet",
"text": "The key is beneath the carpet."},
{"subject": "key", "relation": "inside", "object": "coat_pocket",
"text": "The key is inside the coat pocket."},
{"subject": "key", "relation": "behind", "object": "painting",
"text": "The key is behind the painting."},
{"subject": "key", "relation": "in", "object": "garden_shed",
"text": "The key is in the garden shed."},
{"subject": "key", "relation": "under", "object": "doormat",
"text": "The key is under the doormat."},
{"subject": "key", "relation": "inside", "object": "desk_drawer",
"text": "The key is inside the desk drawer."},
{"subject": "key", "relation": "on", "object": "kitchen_counter",
"text": "The key is on the kitchen counter."},
{"subject": "key", "relation": "behind", "object": "sofa_cushion",
"text": "The key is behind the sofa cushion."},
{"subject": "key", "relation": "in", "object": "shoe_box",
"text": "The key is in the shoe box."},
{"subject": "key", "relation": "beneath", "object": "floorboard",
"text": "The key is beneath the floorboard."},
]
return truth, lies[:n_lies]
def test_contradiction_detection():
"""Test that FHRR binding makes contradictions detectable."""
print("=" * 60)
print("TEST 1: FHRR Binding Detects Contradictions")
print("=" * 60)
enc = FHRREncoder(dim=2048)
# Encode two contradictory claims as structured bindings
claim_a = enc.encode_observation({
"subject": "key", "relation": "under", "object": "oak_table"
})
claim_b = enc.encode_observation({
"subject": "key", "relation": "inside", "object": "red_box"
})
claim_a_repeat = enc.encode_observation({
"subject": "key", "relation": "under", "object": "oak_table"
})
# Compare: same claim vs contradictory claim
sim_same = enc.similarity(claim_a, claim_a_repeat)
sim_contra = enc.similarity(claim_a, claim_b)
print(f" claim A: key-under-oak_table")
print(f" claim B: key-inside-red_box")
print(f" sim(A, A_repeat) = {sim_same:.4f}")
print(f" sim(A, B) = {sim_contra:.4f}")
assert sim_same > sim_contra, "Same claims should be more similar than contradictory ones"
print(f" βœ“ Structured binding distinguishes consistent from contradictory claims")
# Unbind to verify structure is recoverable
decoded_obj = enc.decode_role(claim_a, "object")
top_label, top_sim = decoded_obj[0]
print(f"\n unbind(claim_a, 'object') β†’ '{top_label}' (sim={top_sim:.4f})")
assert top_label == "oak_table"
print(f" βœ“ Object filler correctly recovered via unbinding")
def test_ngc_contradiction_signal():
"""Test that NGC prediction errors spike on contradictions."""
print("\n" + "=" * 60)
print("TEST 2: NGC Prediction Errors Spike on Contradictions")
print("=" * 60)
enc = FHRREncoder(dim=2048)
# Build a field that encodes claims as bindings, not token sequences
field = UnifiedField(obs_dim=128, hidden_dims=[64, 16], fhrr_dim=2048,
ngc_settle_steps=25, ngc_learning_rate=0.005)
# First, establish a belief by presenting the truth 3 times
truth = {"subject": "key", "relation": "under", "object": "oak_table"}
print(" Phase 1: Establishing belief (truth repeated 3x)")
for i in range(3):
r = field.observe(truth, input_type="bindings")
print(f" [{i+1}] PE={r['energy'].prediction_error_norm:.2f} "
f"E={r['energy'].total:.2f}")
pe_after_training = r['energy'].prediction_error_norm
# Now present contradictions
lies = [
{"subject": "key", "relation": "inside", "object": "red_box"},
{"subject": "key", "relation": "behind", "object": "blue_curtain"},
{"subject": "key", "relation": "on_top_of", "object": "bookshelf"},
]
print("\n Phase 2: Presenting contradictions")
contradiction_pes = []
for i, lie in enumerate(lies):
r = field.observe(lie, input_type="bindings")
pe = r['energy'].prediction_error_norm
contradiction_pes.append(pe)
print(f" [lie {i+1}] PE={pe:.2f} E={r['energy'].total:.2f} "
f"\"{lie['relation']}_{lie['object']}\"")
# Present truth again
print("\n Phase 3: Presenting truth again")
r_truth = field.observe(truth, input_type="bindings")
pe_truth_after = r_truth['energy'].prediction_error_norm
print(f" [truth] PE={pe_truth_after:.2f} E={r_truth['energy'].total:.2f}")
# The key metric: prediction error for contradictions should be
# different from prediction error for the established truth
mean_contra_pe = np.mean(contradiction_pes)
print(f"\n Mean contradiction PE: {mean_contra_pe:.2f}")
print(f" Truth PE (re-presented): {pe_truth_after:.2f}")
# Memory similarity should be high for truth, lower for lies
print(f"\n Memory similarity for truth: {r_truth['memory_similarity']:.4f}")
assert np.isfinite(mean_contra_pe)
assert np.isfinite(pe_truth_after)
assert not np.isclose(
mean_contra_pe, pe_truth_after, rtol=0.0, atol=1e-8
), (
"Prediction error on contradictions should differ from prediction error "
f"when the established truth is re-presented "
f"(mean_contra_pe={mean_contra_pe:.6g}, pe_truth_after={pe_truth_after:.6g})"
)
def test_needle_in_lies():
"""
The full needle-in-lies test.
Present a stream of N contradictory claims interspersed with the truth.
Score each claim by its fit to the established belief pattern.
The truth should score highest (lowest energy / prediction error).
"""
print("\n" + "=" * 60)
print("TEST 3: Needle-in-Lies (13 contradictions)")
print("=" * 60)
truth, lies = make_needle_scenario(n_lies=13)
field = UnifiedField(obs_dim=128, hidden_dims=[64, 16], fhrr_dim=2048,
ngc_settle_steps=25, ngc_learning_rate=0.003)
# Build the claim stream: truth appears at positions 0, 5, 10
# (establishing the "needle" among the "lies")
claims = []
truth_indices = set()
# Initial truth
claims.append(truth)
truth_indices.add(0)
# Interleave lies and truth
for i, lie in enumerate(lies):
claims.append(lie)
if i == 4:
claims.append(truth) # Repeat truth midway
truth_indices.add(len(claims) - 1)
if i == 9:
claims.append(truth) # Repeat truth again
truth_indices.add(len(claims) - 1)
print(f" Total claims: {len(claims)} ({len(truth_indices)} truth, {len(claims) - len(truth_indices)} lies)")
print(f" Truth positions: {sorted(truth_indices)}")
print()
# Feed all claims to the field
energies = []
for i, claim in enumerate(claims):
is_truth = i in truth_indices
label = "TRUTH" if is_truth else "lie "
# Encode as structured binding
bindings = {k: v for k, v in claim.items() if k != "text"}
r = field.observe(bindings, input_type="bindings")
pe = r['energy'].prediction_error_norm
e = r['energy'].total
ms = r['memory_similarity']
energies.append({
'index': i,
'is_truth': is_truth,
'pe': pe,
'energy': e,
'mem_sim': ms,
'text': claim['text'],
})
print(f" [{i:2d}] {label} PE={pe:8.2f} E={e:9.2f} mem={ms:+.3f} "
f"{claim.get('relation', '?'):12s} {claim.get('object', '?')}")
# === SCORING ===
# After processing all claims, score each one by re-presenting it
# and measuring how well it fits the settled belief state
print(f"\n --- Re-scoring all claims ---")
scores = []
for i, claim in enumerate(claims):
bindings = {k: v for k, v in claim.items() if k != "text"}
fhrr_vec = field.encoder.encode_observation(bindings)
# Score directly in FHRR space: compare this claim's FHRR vector
# to the FHRR vectors of all stored observations.
# The truth was stored 3 times; lies were stored once each.
# Hopfield in FHRR space will favor the repeated pattern.
# Build a Hopfield bank from the raw FHRR observations
# (We only need to do this once, but it's clearer inline)
if i == 0:
fhrr_memory = HopfieldMemoryBank(dim=field.fhrr_dim, beta=0.005, capacity=100)
# Re-encode and store all claims as they were presented
for j, c in enumerate(claims):
b = {k: v for k, v in c.items() if k != "text"}
fv = field.encoder.encode_observation(b)
fhrr_memory.store(fv, normalize=True)
# Retrieve: how well does this claim match the memory's attractor?
retrieved_fhrr, fhrr_energy = fhrr_memory.retrieve(
np.real(fhrr_vec).astype(np.float64), steps=5)
# Similarity to retrieval
q = np.real(fhrr_vec).astype(np.float64)
q_norm = np.linalg.norm(q)
r_norm = np.linalg.norm(retrieved_fhrr)
if q_norm > 1e-8 and r_norm > 1e-8:
fhrr_sim = float(np.dot(q / q_norm, retrieved_fhrr / r_norm))
else:
fhrr_sim = 0.0
# Score = FHRR memory similarity + Hopfield energy (more negative = deeper attractor)
score = fhrr_sim - 0.1 * fhrr_energy
scores.append({
'index': i,
'is_truth': i in truth_indices,
'score': score,
'pe': pe,
'text': claim['text'][:50],
})
# Sort by score (best first)
ranked = sorted(scores, key=lambda x: x['score'], reverse=True)
print(f"\n Ranking (higher score = better fit to beliefs):")
for rank, item in enumerate(ranked[:5]):
marker = "β˜…" if item['is_truth'] else " "
print(f" #{rank+1} {marker} score={item['score']:8.2f} PE={item['pe']:8.2f} "
f"\"{item['text']}\"")
print(f" ...")
for item in ranked[-3:]:
marker = "β˜…" if item['is_truth'] else " "
print(f" #{ranked.index(item)+1:2d} {marker} score={item['score']:8.2f} PE={item['pe']:8.2f} "
f"\"{item['text']}\"")
# Check: is any truth claim in the top 3?
top_3_indices = [item['index'] for item in ranked[:3]]
truth_in_top_3 = any(i in truth_indices for i in top_3_indices)
# Check: is the best claim a truth?
best_is_truth = ranked[0]['is_truth']
print(f"\n Best claim is truth: {best_is_truth}")
print(f" Truth in top 3: {truth_in_top_3}")
assert truth_in_top_3, "At least one repeated truth should rank in the top 3"
# Compute mean score for truth vs lies
truth_scores = [s['score'] for s in scores if s['is_truth']]
lie_scores = [s['score'] for s in scores if not s['is_truth']]
mean_truth = np.mean(truth_scores)
mean_lie = np.mean(lie_scores)
print(f"\n Mean score (truth): {mean_truth:.4f}")
print(f" Mean score (lies): {mean_lie:.4f}")
print(f" Separation: {mean_truth - mean_lie:+.4f}")
if mean_truth > mean_lie:
print(f" βœ“ Truth claims score higher than lies on average")
else:
print(f" βœ— Lies score higher β€” NGC hasn't separated them yet")
def main():
tests = [
("FHRR Contradiction Detection", test_contradiction_detection),
("NGC Contradiction Signal", test_ngc_contradiction_signal),
("Needle-in-Lies (13 contradictions)", test_needle_in_lies),
]
print("\n" + "β–ˆ" * 60)
print(" NEEDLE-IN-LIES TEST")
print(" Can hierarchical predictive coding resolve contradictions?")
print("β–ˆ" * 60)
results = []
for name, fn in tests:
try:
ok = fn()
results.append((name, ok))
except Exception as e:
print(f"\n βœ— {name} FAILED: {e}")
import traceback; traceback.print_exc()
results.append((name, False))
print(f"\n{'=' * 60}")
for name, ok in results:
print(f" {'βœ“' if ok else 'βœ—'} {name}")
print(f" {sum(1 for _, ok in results if ok)}/{len(results)} passed")
return all(ok for _, ok in results)
if __name__ == "__main__":
main()