Two engines. One evidence loop.
During R&D, MAL and 8zCoding deliberately remain separate. MAL creates intelligence-rich task packets and blind candidate sets; 8zCoding owns local execution, hidden verification, benchmarking and evidence. The separation is a measurement boundary, not a product limitation.
Intelligence before code
Reconstruct the problem, sharpen the behavioral contract, generate alternative implementations and keep provider identity private from the evaluator.
Reality after code
Compile, execute, verify, benchmark, measure memory and artifacts, bind results to exact candidate bytes, and preserve failures as evidence.
Allocation must earn its complexity
Equal allocation, greedy, bandit-style and DCC controllers compete. DCC is a candidate — never a privileged winner.
A wind tunnel before real software.
The first domain is intentionally narrow: count(n, prefix) must return the exact number of legal N‑Queens completions. That makes it possible to harden the judge before asking whether the search process itself is intelligent.
Correctness is a hard gate
Independent Python and C++ reference paths, strict I/O, canonical output and known totals make silent “almost correct” code unacceptable.
Known wrong programs must die
The A0/A1 calibration rejected all 11 seeded wrong candidates, including prefix, diagonal, symmetry, truncation, grammar and nondeterminism defects.
Platform bugs count
Actual Windows 11 + MSVC runs exposed and repaired compiler sidecar and CRLF/LF protocol defects before the verifier was accepted.
We have a strong local judge for the frozen N‑Queens behavioral domain. It does not establish that multi-LLM coding, MAL, DCC or 8zCoding live search is superior. Those are later experiments.
The machinery works. The favored mechanism has not won.
The current controller package is useful precisely because it does not force a success story. It separates engineering validity from scientific promotion.
| Layer | Status | What it means |
|---|---|---|
| A0/A1 verifier & oracle | PASS | Strict behavioral judge; actual Windows 11 + MSVC core and package-close confirmed. |
| A2 synthetic calibration | ENGINEERING PASS | Controller mechanics, replay legality and synthetic hard predicates operate as designed. |
| A3 N‑Queens TUNE replay | ENGINEERING PASS | Real candidate ecology and controller replay executed deterministically. |
| A4 opportunity / oracle | PROVISIONAL | Primary metric is AUC-U, but the current oracle optimizes final utility first. Alignment successor required. |
| A4 passive LZ prediction | NOT PROMOTED | lz_norm did not achieve the required out-of-bank predictive improvement. |
| A5 one-shot replay | ENGINEERING PASS | Internally precommitted same-generator holdout executed without post-holdout retuning. |
| P-DCC-UTILITY | NOT MET | C5-R did not materially beat the strongest ordinary controller. |
| P-DCC-COMPRESSION-MECHANISM | NOT MET | Incremental value of the LZ/process-compression channel is not established. |
| A6 live coding | NEXT | Allowed after bridge conformance and sandbox gates; can begin with equal or ordinary allocation. |
The A5 evidence is a one-shot internally precommitted same-generator holdout: useful internal validation, not independent replication. Two order variants of one content seed are nested measurements, not two independent banks.
An observed mean delta inside ±0.02 says the average difference is small relative to the predeclared practical margin. It does not, by itself, establish formal equivalence or non-inferiority.
The A4 oracle found a bug in our own claim boundary.
The replay protocol declares AUC-U as the primary utility, but the current static/adaptive oracle chooses schedules by final utility first and AUC second. That makes it possible for a deployable controller to score higher AUC-U than the object called the AUC oracle.
Negative “room”
HOLDOUT_570173 / O1 reports room_from_c0 = -0.093269... and room_from_c4 = -0.091226.... A true upper bound for the same primary metric should not do that.
Split the objectives
The planned v0.2.1 successor separates ORACLE-A-AUC, ORACLE-A-FINAL, ORACLE-S-AUC and ORACLE-S-FINAL, with a truth-known canary where AUC-optimal and final-optimal schedules differ.
The oracle defect makes A4 room/regret interpretation provisional. It does not retroactively erase raw controller traces or their direct AUC-U means. The v0.2 package remains historical evidence; the correction belongs in a new versioned successor.
Blind generation on one side. Untrusted execution on the other.
The next system-level experiment is not “ask many models and vote.” It is a controlled bridge in which MAL prepares the same canonical coder packet for multiple fresh providers, hides identity from evaluation, and 8zCoding returns only evidence that is safe to use for repair.
Stateless providers
Each first-round model receives the same canonical task, no peer answers and no MAL synthesis preference. MAL is locally stateful only for provenance, costs, hashes and resume.
Opaque IDs
The evaluator sees candidate bytes and opaque IDs — not provider prestige or route order. Private provenance remains separate.
One conformance kit
MAL_8ZCODING_BRIDGE_CONFORMANCE_KIT_V1 must define the exact schemas, canonicalization, path safety, ZIP/hash rules and feedback redaction both engines obey.
Evidence-guided repair must be compared against an equal-cost fresh cold sample. The first pilot allows at most one repair round, then evaluates on a separate final post-repair holdout.
Generated code is data until the sandbox says otherwise.
MAL never executes provider-generated code. A6 must not import or execute untrusted candidate code directly inside the normal MAL process or an unprotected host environment.
Disposable execution
Use a disposable Windows Sandbox, VM, isolated WSL/container or comparable environment. Direct host execution is outside the intended A6 security contract.
No network. No credentials.
Bounded mounts, read-only or mediated oracle access, process-tree control, timeouts, memory/disk/output limits and no secrets in the candidate environment.
Safe feedback only
Models may receive failure classes and redacted diagnostics, never hidden fixture bytes, expected hidden answers or enough repeated information to reconstruct the final holdout.
The shortest path to new evidence.
Several tasks can proceed in parallel, but the first claim-bearing live coding run stays behind explicit interface and sandbox gates.
Not more activity. Better verified software per total cost.
The stack earns its extra machinery
Across more than one programming task, the system produces a better verified Pareto frontier — correctness, runtime, size, robustness and discovery cost — than strong single-agent, Best-of-N and simple evolutionary baselines under comparable total budget.
The simpler system wins
If MAL prompt forging, Top‑5 diversity, evidence repair or DCC fails to repay its coordination overhead, those layers should be removed or demoted. The verifier, blind forge, evidence bridge and release pipeline can still remain valuable.
If equal allocation or a simpler ordinary controller matches DCC at lower cost, the simpler controller wins. ssMDL×DCC only means the governor is allowed to compete — not that it is guaranteed victory.
A measurable first work cell — if it earns the role.
If MAL × 8zCoding proves value, it can become the first operational cell of a future AI8 Work Environment: locally continuous project state, replaceable AI providers, task/evidence graphs, multiple workstreams and controllers that themselves remain subject to comparison.
State belongs to the project
Provider/model calls can stay fresh and replaceable while local continuity preserves goals, artifacts, evidence, decisions, budgets and open tensions.
MAL coordinates possibilities
Architect, researcher, alternative architect, falsifier and integrator roles can propose work without being mistaken for independent empirical evidence.
8zCoding closes the loop
Programming is the first domain because compiler, tests, hidden oracles and benchmarks make it unusually hard to substitute persuasive prose for a working artifact.
An effective AI8 work cell would be evidence about a workflow architecture. It would not, by itself, establish AGI, ASI, consciousness, personhood or ontological continuity.
Executable evidence exists — with a known open debt.
8zCoding_NQueens_Controller_Replay_A2_A5_v0_2_RND.zip
This is an executable research laboratory, not a deployment release. Its controller traces and engineering results remain useful; A4 oracle room/regret is provisional until the bounded oracle-alignment successor is built and rerun.
The package is referenced here by identity, not assumed to be a public download. If released publicly later, it should be published as a proper Code Capsule with exact commands, package-close receipt and the oracle-alignment limitation visible beside it.