Oczy Research Agenda

A prioritized set of research projects for the Plastic World Model Agent (../experiments.txt). Each project is a problem, falsifiable hypothesis, approach, and discriminating success criterion. The original projects 01–07 have matching designs under ../experiments/; later numbered specs are pre-registered remediation and successor experiments, with concrete experiment directories linked where they exist.

The thesis in one line: experience → fast change → replay → compression → slow change → forgetting raw trace. "An agent should not store experiences; it should metabolize them." The missing art is the metabolism.

The organizing problem: the eval went blind

The single fact that shapes this whole agenda is that the headline benchmarks are saturated. Across recent runs the git log repeatedly reports "All benchmark metrics at 1.0"; autoresearch.sh prints code_qa_accuracy=1.0; and architecture variants that are mechanically different (e.g. scalar vs hybrid consolidation — consolidation_strength 10.0 vs 36.0, cold_drift 0.10 vs 0.64) tie on every behavioral metric. When the measuring instrument can no longer separate two architectures, no architecture work can be honestly claimed to win.

So the agenda is built as a spine and ribs:

• The spine is project 01 — rebuild the eval so a behavior difference is the only thing it can register (behavior_delta_per_byte, thesis §14). Everything else is masked until this lands.
• The ribs (02–07) are the architecture frontiers the un-blinded eval is meant to adjudicate: the steering ceiling, the perception depth, the scope-control failure, the unclosed metabolism loop, the memory-bloat bottleneck, and the predictive (world-model) foundation.

The projects

Current verdicts (Campaign 0d48130, 2026-07-11)

All reported completed runs used CPU-only; Campaign Exp03 was infrastructure-blocked before execution. Curated campaign log: ../experiments_logs/2026-07-11_campaign_0d48130.md. The S1.4 layer-L probe log is ../experiments_logs/2026-07-01_s1_4_hf_layer_probe.md. Nulls and refutations are reported alongside wins.

Post-remediation frontier

Sprints 0–3 invalidated the cheap activation-steering route, established retrieval as the honest baseline, and left a narrower question: can learned neural state outside a frozen LM acquire and express experience without an answer-time content store?

Dependency and verdict order:

18  LM-weight comparator ───────────────┐  TESTED-PARTIAL (gate passed,
                                        ├─► same reporting table         bimodal full run)
19  direct cortex: label vs latent ─────┘  PENDING
                    │
                    ▼
20  meta-trained cortex over frozen language organ  PENDING
                    │ ACCEPT + causal-state audit
                    ├──────────────────────► 21  cortex-routed frozen language +
                    │                           action organs  BLOCKED (needs 20)
                    │
                    └───────────────────────────────┐
                                                    ▼
22A standalone frozen-backbone LoRA-bank ────────► 22B cortex + PMD integration
    PMD comparator  PENDING (independent of 20)     BLOCKED (needs 20 + valid 22A)

Retrieval is mandatory in every comparison table but disabled in the primary conditions for 19B, 20, and 21. A retrieval win is a result; it cannot be relabelled as cortex metabolism.

Dependency graph

                01  Benchmark v2  (de-saturate — unblocks honest claims for all)
                 │                   TESTED-NULL: 5 de-saturation events,
                 │                   zero behavior delta
   ┌─────────────┼───────────────────────────┬───────────────┐
   │             │                            │               │
  02 KV-slot    03 Layer-L hiddens           06 Bounded      07 World model
  fact inject    (perception depth)           growth          (predict the
  REFUTED        REFUTED (S1.4)               ACCEPTED        correction)
  kv_slot=0      gap < +0.10                  m1=0.002        ACCEPTED-PARTIAL
   │             │                            │   5 seeds      uptake=1.0
   │             ├──────────────┐             │               critic_auc=0.0
   │             ▼              ▼             │
   │            04 Context-     05 Metabolism │
   │            scoped          loop closure ─┘
   │            attractors      (compounding drift)
   │            ACCEPTED        TESTED-NULL
   │            SSI=1.0         drift=0.0
   └────────────┴──────────────┘
        02/03/04 all chip at the same wall: a residual control vector
        carries posture; facts need a reserved slot (02) keyed by real
        context (03/04). 05 needs 03's real hiddens; 06 needs 01's metric.
        Soft-dep: 07 → 01 — the world-model can run independently, but its
        `accept_pred_auc ≥ 0.70` acceptance threshold only becomes
        meaningful once 01's de-saturated scorecard exists.

Reading order if you only do one thing: 01, then whichever rib has live momentum. 03 is refuted (S1.4: mid-layer hiddens do not beat final layer on two architectures); 04 and 05 both consumed 03's input and have their own verdicts (04 ACCEPTED, 05 TESTED-NULL). The highest-leverage open question is now 18 (TESTED-PARTIAL) → 19 (PENDING) → 20 (PENDING).

Conventions these proposals follow

• Matched-pair, single-variable controls — the repo standard (see the SVD-vs-non-SVD and S-vs-H runs). Every experiment isolates one axis.
• Mock vs real driver — the _MockDriver (deterministic hash embeddings, zero semantics) is a structural null; the real LFM2.5-1.2B-Instruct Q4_K_M driver is the semantic test. Results are always reported separately.
• Non-saturating metrics only — success criteria deliberately avoid the metrics already pinned at 1.0 (code_qa_accuracy, real-driver co_recall, capped forgetting/identity ratios). Where a 1.0 metric is kept, it is used only as a regression floor, labeled as such.
• Honest nulls are a result — several proposals (notably 01 and 05) pre-register the possibility that the answer is "no measurable difference," which is itself a de-saturation win over a saturated tie.

Provenance

Drafted 2026-06-28 by a multi-agent pass: six readers mapped the subsystems, one author drafted each project, and an adversarial grounding checker verified every claim about the existing code against real files (correcting line numbers, distinguishing semantics-free mock 0/0 from real-driver 1/1, and softening overstated capabilities). Two of the six map briefs (organs, benchmarks) degraded to placeholders, but the per-project grounding pass read those subsystems directly, so the cited eval_suite.py / autoencoder.py / hypernet.py / organism.py facts are first-hand. Treat any remaining unverified forward-looking claim as a hypothesis, not a finding.

2026-07-09 human-authorized extension: Research/19 was amended before implementation to close the label-prefix interpretation loophole; Research/20 and /21 plus Experiment/09 were added to test a meta-trained cortex controlling frozen language and action organs without retrieval in the primary path.

2026-07-11 status update: Campaign 0d48130 adjudicated projects 01–07, 14, and 18. Verdicts and outcome annotations added to each project spec; this README updated with a current-verdicts table, R18 result in the frontier table, and dependency-graph status labels. Projects 19 and 20 remain PENDING; 21 remains BLOCKED (depends on 20). No hypotheses, success criteria, or kill criteria were modified. Evidence: ../experiments_logs/2026-07-11_campaign_0d48130.md.

2026-07-16 human-authorized extension: Research/22 was added as a standalone strict-versus-calibrated PMD addressability comparator over a shared frozen backbone and LoRA memory bank. Its Stage A is independent of Research/20, while Stage B cortex integration requires both a valid Stage A route and Research/20 acceptance. Research/21 was not changed.