flyloop

A closed-loop sandbox wiring three research projects into one learning cycle: experience -> memory -> prediction -> error -> update -> new experience -> insight

flyloop is the mechanical orchestrator for the Mushroom-Body Program — the research line that asks when does compression induce structure? It connects three organs that were previously studied only in isolation:

Organ	Repo	Role in the loop
Memory	aujurd22/flymemory	storage, recall, dedup, supersede, consolidation — all prediction knowledge comes from here
Proposer	aujurd22/flypoet	k-WTA sparse-competition networks, trained online (few updates x partitioned parameters)
Verifier	aujurd22/intuition-mechanism	registered-prediction ledger culture; restructuring-signature measurement

There is no LLM in the loop — every step is mechanical, the world is a pure function of the cycle index, and every error signal has recoverable ground truth. The point is to measure the dynamics of memory-driven prediction under drift: the phase-transition laws (L3), failure laws (L4/L5), and the PROPOSE -> REGISTER -> RUN -> VERIFY -> RECORD -> REMEMBER cycle described in flymemory's research/LOOP.md — with propose left to the human/LLM session boundary, everything else automated.

The loop

One iteration per cycle (~0.09 s per arm pass; the V3 run sustains ~4,800 cycles/hour with both arms and their poets):

1 EXPERIENCE  the world emits events: a station's channel query,
              a hidden-rule puzzle probe, the next symbol of a stream
2 MEMORY      recall the relevant compartment from FlyMemory
              (or a direct state_lookup in the V4 arm)
3 PREDICT     mechanical reasoner (vote / mod-13 fit / recorded rule)
              or the k-WTA net's next-symbol argmax
4 ERROR       score against the world's ground truth
5 UPDATE      correction writes / table fold-in / rule abstraction /
              one sparse gradient step (25% partition, k-WTA 0.25)
6 NEW EXPERIENCE  the world has drifted under our feet (rotations, shocks,
              regime shifts) — the next lap starts on new ground
7 INSIGHT     rule-abstraction events and drift-response statistics are
              written to the registered ledger and back into memory

Lanes

• fact — 24 stations whose channel state rotates on staggered periods, plus shocks. Tests state tracking under drift; A/B arm compares semantic recall (arm A) against V4 state_lookup structural read (arm B).
• puzzle — 4 families of y = (a x + b) mod 13 with epoch rotations. The loop must abstract a rule from observed exemplars into memory (a "discovery"), then reuse it. Discoveries are insight events.
• seq — a 3-regime Markov stream, predicted online by a 494K-parameter k-WTA GPT (dual streams: A hides the regime = unidentifiable task; B prefixes a regime marker = identifiability arm).
• noise — unique distractor entries every 20 cycles, exercising dedup/decay under write load.

Registered predictions

Every run pre-registers its claims before the first cycle (git timestamps are the proof, LOOP.md style): run-level predictions (V2-P01..V2-P10) plus per-discovery signature entries. The final report adjudicates each to CONFIRMED / REFUTED / PARTIAL with evidence — never post-hoc thresholds.

Results so far

Run 1 — night_20260927_0143 (10 h, 55,173 cycles, 0 failures)

Three findings, all unplanned (full report):

1. The 80-char truncation self-lock. FlyMemory's recall truncates entry text to 80 chars; 31.2% of the observation-table entries were cut inside their trailing c= field, making them unparseable -> 26% "cold" probes, which write longer tables, which truncate more. A long-horizon memory system reality: writing in more detail can poison the read path.
2. Rules never survived as nodes. Per-family rule entries measured 0.869 similarity against their own observation table — inside FlyMemory's merge zone (0.75-0.92), where rewriting is in-place regardless of compartment or state_key. All 110 rule entries became tombstones.
3. The sequence lane's null result was identifiability, not capacity. With three regimes mixed and no observable marker, the optimal predictor is the marginal one (theory: uniform 0.9375 / marginal 0.9102 / oracle 0.6386; measured 0.75-0.87 flat). A clean negative result: this failure mode is unidentifiability, not underfitting.

The run also served as a soak test for FlyMemory itself: 20,977 entries, 74 MB pickle, 10 h of continuous writes, zero corruption, mean save 0.17 s.

Run 2 — night_20260927_1228 (5 h, v2 fixes)

Metric	v1	v2
cold-probe rate	26.1%	4.0%
rules surviving as parseable memory	0 / 110	4/4 families in one RULEBOOK entry, 0 read-back failures
fact arm B (state_lookup) error	n/a	0.18 vs arm A 0.28

Verdicts V2-P01..P10: 7 CONFIRMED / 2 PARTIAL / 1 REFUTED. The seq null result survived the identifiability repair (marked stream B ≈ unmarked A), which upgraded it: the bottleneck is the small poets' aggregate learning, not information availability (full report).

Run 3 — v3_20260928_0252 (10 h, V3: recurrent drift × memory consolidation)

The first controlled mechanism experiment: two memory arms through the identical precomputed episode schedule (NEW 50% / VARIANT 25% / RECALL 25%, gaps 2–5 epochs), FULL with a per-family rule registry, EPISODIC with tables only, byte-exact padding-write parity. Each arm has its own sandbox FlyMemory instance (ports 8769/8770).

Verdict	Claim	Result
P01 ✓	FULL lower error on RECALL (primary)	dE20 = +0.641, cluster-bootstrap CI [0.521, 0.776], n=78, perm_p<1e-4, 4/4 families positive
P02 ✓	FULL faster recovery (primary)	latency +0.551 probe, cluster CI [0.426, 0.693]
P03 ✓	advantage at every gap 2/3/4/5	+0.47 … +0.75, all positive
P04 ✓ (NC1)	advantage absent on VARIANT/NEW	|dE| 0.014 / 0.02 vs 0.641 — pure reuse signature
P05 ✓	stale intrusion within band	SIR ratio 1.42 ≤ 1.5 — memory benefit and memory interference are the same mechanism's two faces
P06 ✗	useful-insight rate above null	REFUTED after correction (external review found no temporal check / REACTIVATION conflation; corrected statistic equals its null; the decisive discovered-vs-undiscovered contrast is untestable — FULL abstracted 100% of rules)
P07 / P08 —	noise-phase / fact-lane controls	INCONCLUSIVE (phase C beyond the reachable 48.3k cycles; design mis-calibration registered)

Mechanism, resolved from the data: recovery is a one-observation story — FULL adopts a returning rule from a single live pair (book_test, unique match required) at probe 2, while EPISODIC needs two pairs to fit at probe 3. Zero-observation recovery never happened (0/78). Full analysis, corrections trail, and the V4 queue: findings/v3_20260928_0252, V3_DESIGN.md.

Run 4 — v4_20260928_1636 (1.9 h, V4: discovery-failure × disentanglement, early stop)

Three arms: FULL / MATCHED (episodic pair-table archive + the SAME unique-match decision rule — pure algorithm control) / EPISODIC; variable episode lengths create rules that never get abstracted.

Verdict	Claim	Result
P02 ✓ (disentanglement)	F ≤ M ≤ E on discovered recurrences	E20 = 0.43 / 0.57 / 1.06 — the matcher adds +0.49 (CI excludes 0), compression only +0.14 (CI crosses 0)
P04 ✓	probe-1 recovery exists	62.7% (pre-folded interface; V3's 0/78 was a folding-order artifact)
P06 ✓ (NC1)	VARIANT/NEW contrasts ≈ 0	the advantage stays recurrence-specific

findings/v4_20260928_1636

Run 5 — v5b_20260928_1848 (2 h, V5: the condition law, ε=0.25)

Verdict	Claim	Result
P01 PARTIAL	ε=0 replication	dE20(M−F) = +0.137 CI[−0.043, 0.310]
P02 ✓ (headline, operational)	dE20(M−F) larger at ε=0.25	+1.155 CI[0.755, 1.571]; between-run contrast +1.018 CI[0.584, 1.472]
P03 ✓	discovery rate drops under noise	0.894 → 0.744
P05 ✗	EPISODIC degrades slowest	REFUTED — MATCHED damaged most (+4.14)

Headline with the correction: the design narrative predicted noise would shrink the compression advantage; the data show it TRIPLES. The registered operational claim (larger at ε=0.25) is confirmed; the narrative interpretation is refuted and the mechanism took another run to pin (V6). findings/v5b_20260928_1848

Run 6 — v6t_20260928_2159 (1.9 h, V6: tolerant matching, algorithm A/B)

Only the matcher changed (fraction-best ≥ 0.6 with ties→refusal, vs exact), same ε=0.25 world — was V5's amplification a matcher artifact?

Verdict	Claim	Result
P01 ✗ (headline)	the amplification is a matcher artifact	REFUTED — the gap WIDENED +1.155 → +2.306: tolerance rescued FULL (−1.27) and left MATCHED flat (−0.12)
P03 ✓	matcher paths reopen	epi_test 44 → 60, book_test 151 → 174
P04 ✗	registered cost (rising stale intrusions)	did NOT materialize (53 → 53 flat) — the 0.6 band + strict-argmax refusal kept tolerance from leaking

The law as it now stands: under observation noise the recurrence advantage of structured memory over instance memory GROWS — write-time verification (noise filtered at the write) × full-support coverage (a rule generalizes to every x; an instance covers only the x it saw). The P46 instance-friendly condition (continuous/overlapping families) is a different axis and remains untested in the loop (V7). findings/v6t_20260928_2159, V6_DESIGN.md

Runs 7-8 — V7 (factorial + wavy world) and V8 (adaptive read policy)

V7A (factorial 2×2, exact matcher): support effect +2.14 CI[0.82, 3.44] is the only significant contrast; write-depth effect −0.99 CI[−2.29, +0.31], sign unconclusive. V7B (wavy world): no order inversion — coverage is the binding constraint. V7C (write-depth A/B): shallow write significantly beats deep write (−0.655 CI[−1.10, −0.17]) — V5's write-time-filter claim REFUTED in direction.

V8 (ε=0.40, 4 arms including FULL-ADAPT): the adaptive read policy (rolling flip estimate adjusts min_frac) beats fixed policy by 0.41 errors (3.741 vs 4.155). The full ε dose-response reveals a non-monotonic curve — ε=0.15 gives the BEST FULL E20 (1.102), lower than ε=0 (2.347): moderate observation noise improves rule-based memory by forcing cleaner registrations.

ε	FULL E20	reading
0	2.347	baseline; marginal rules pass verification
0.15	1.102	optimal: low-rate flips filter marginal rules
0.25	~3.6	noise starts overwhelming the filter
0.40	~4.2	noise dominates; adaptive policy partially compensates

RSI-0 lineage: g0 (2.347) → G1 ✗ → G2 ✗ → G3 "accepted" (adaptive, 2.000) → G4 ✗ (retracted: bit-identical duplicate of G3) → G4′ (TRUE replication, fresh realizations): the G3 gain does NOT replicate (+0.163 vs parent, CI [−0.976, +1.380]) — the adaptive-policy effect is within realization noise; its acceptance stands downgraded to single-realization evidence. What does replicate, strongly: the structured-memory advantage (G4′: registry 2.581 vs archive 5.186 vs none 5.372). The loop's most reliable faculty remains negative selection; the next accepted mutation must come from a new axis (registered: abstraction-loss world / V9), not re-tuning. The retraction bought the seed split (FLYLOOP_RUNSEED), a preflight guard, and a mechanical independence check. Standing rule: paired contrasts want the same seed; replications want a fresh one.

12.5h marathon × 12-era world (ε and active families rotate every ~2.5k cycles, unannounced): era transitions cause ~no regression — the rule registry absorbs world change with zero forgetting and no detectable transient; repeat eras halve RECALL error 17.5k cycles later (C errR 0.198 → 0.090); frozen ≈ adaptive across all 12 eras (mean Δ +0.0022), replicating the G4′ verdict at 14× timescale. Marathon

V8-LLM transfer arc (Qwen2.5-0.5B + LoRA on GSM8K, does the toy-world condition law survive contact with real fine-tuning?): at generalization regime the noise/depth effects are absent and bounded ≤ ~3pp (vs flyloop's 2-4× relative) across 4 cells × 2 seeds; under memorization pressure (200 × 8ep, train-recall 42.7% vs test 32.4%) the laws still do not reappear — instead the model emits the TRUE answer on corrupted training instances ~35pp more often than it replays the corrupted stored value. Parametric storage is not episodic storage: the LM behaves like the registry arm with a built-in verification filter — it already sits on the winning side of the program's law (abstraction beats episodic under noise). The laws bind to instance-addressable memory systems; benchmark generalization is not one. V8-LLM-2, V8-LLM-3

V9 W9A wave sweep (the abstraction-loss axis): make the world's rules carry an incompressible per-rule sinusoid (amplitude W) that no compact (a, b) can reproduce. Result: not a gradual floor but an abstraction cliff — FULL falls 1.14 → ~10.5 (E20) the moment W > 0 and all three arms saturate together. Mechanism (method decomposition): the world's physical tolerance was coupled into discovery AND matching bands, and a band wide enough to forgive the world's residual destroys candidate contrast — wrong rules win argmax, the book fills with wave-fitted wrong rules (direct rule hits 427 → 29). MATCHED wins significantly at W=2 (the crossover), ties beyond. Registered law: a tolerance wide enough to forgive the world's residual is wide enough to destroy identification. Recoveries registered and queued (W9B: archive keeps tol=0 since its stored pairs stay exactly valid; M5: residual registry — abstraction with its own loss stored alongside). Product translation: summary-only storage collapses retrieval contrast on residual-bearing content; measure the residual before compacting. V9_DESIGN.md, W9A

W9C/W9D — M5 residual registry ACCEPTED: the registered recovery ran and replicated under the full G4 gauntlet. W9C (same-seed paired against a bit-identical FULL control, 2685/2685 events): +0.804 E20, CI [+0.244, +1.404]. W9D (fresh realizations, diverges at event 0): +1.208, CI [+0.526, +1.793]. First accepted new-axis mutation (store the abstraction's own loss); recovery ~8-11% of the cliff, scaling lever = residual table density (W9D 74% vs 85% armed/unarmed probe errors).

V10 composite world (SDB-aligned, P148's registered loop-closer): composition is the registry's killer feature — composite-rule RECALL E20 1.333 vs archive 5.222, probe-1 hits 56% vs 7.7% random (answers unobserved R3 from stored R1, R2 = mechanical SDB TRANSFER). Periodicity REFUTED the predicted episodic win and bought a deeper law: drift is the abstraction-loss type that punishes BOTH architectures (all arms ~6.2-6.8 vs 3.86 standard) — neither (a,b) nor instance replay carries (b0, δ, k); the missing representation is δ itself (registered mutation M6). Anomaly probes: ~90% error for all arms (unpredictable by design); memorization-rate join queued. V10, W9C

Full data, corrections, and next-step registrations: V7A, V7B, V7C, V8, G4, G4′, V8_DESIGN.md

Engineering discipline (read this before writing to a memory system in a loop)

Two live bugs were caught by pre-launch audits, and both generalize:

• Calibrate with the real emitters, not approximations. The first v2 draft calibrated similarity with hand-written approximate texts; the actual emitters crossed the 0.75 merge line at runtime and rules were swallowed. Every text form is now calibrated by tests/calibrate_texts.py against the actual embedder with hard constraints: distinct entities < 0.75, same-entity update > 0.75, own-entity query ranks first.
• Make truncation unable to fabricate data. Recall cuts at 80 chars; a cut can turn 12:10 into 12:1 — a malformed observation that silently poisons the fit. v2 puts observation pairs first so a cut can only remove trailing metadata. tests/test_parsers.py fuzzes every emitted form under the exact 80-char cut.

Three more from V6-V8:

• force_new is load-bearing for multi-entry designs. Same-family entries (book rules, archive rows) share enough text to exceed the 0.92 duplicate threshold — without force_new, the merge zone rewrites earlier entries in place, silently destroying history. This killed G1 (single 190-char entry) and would have killed the per-rule archive.
• Entries must stay < 120 chars. The engine's split_chunks cuts texts longer than 120 chars into multiple chunks; the state_key atomicity guard then rejects state_key writes that produce more than one chunk — silently.
• Every arm needs its own FlyMemory instance. The merge zone ignores compartments, so same-persona texts from different arms would collide in one store. Ports are disjoint per arm (V7C: 8769/8770/8771).

Robustness for unattended multi-hour runs: per-cycle try/except with a consecutive-failure circuit breaker, atomic checkpoints every 50 cycles, heartbeat/status/report cadence, local-mirror degradation when the memory service dies (with auto-switch-back), RAM red line, sleep guard, and a logon-resume hook so a reboot continues from the checkpoint (the absolute deadline lives in deadline.json).

Repository layout