R12 EFC HSC Post-Canary Decision Protocol

Status

Pre-fit decision design only. The exact v6 two-update measurement canary is the only authorized GPU run. This document does not authorize a qualification fit, development or confirmation access, a native-reasoning claim, or continuation pretraining.

Why a training loss is not evidence

The maximum Hankel-shift compiler has 64,407,956 trainable source-compiler parameters. Its 384-source train package contains 242,688 independently measured target bits and 1,966,080 deterministic signature-expansion bits. That is approximately 265 trainable parameters per independent target bit. Complete base and derivative rollout supervision therefore permits a source-specific interpolation strategy. Neither low train loss nor exact train codebooks, alone, establish a reusable shift law.

The first admissible qualification must distinguish:

1. optimizer or numerical failure;
2. deep-path optimization failure;
3. invalid Hankel decoding;
4. raw-source compilation failure;
5. source-specific memorization;
6. renderer transfer;
7. unseen-world transfer; and
8. genuine causal use of prefix-shift incidence.

Train-only partition

Use only the existing 384 frozen train sources from 96 worlds. Development and confirmation artifacts must not be mounted.

Before reading outcomes, select worlds independently within each of the six action families by a domain-separated hash of the frozen world identity:

• Fitting (F): 12 worlds per family and three renderers per world, 216 sources total.
• Renderer holdout (R): the fourth renderer of the same 72 worlds, 72 sources total.
• World holdout (W): four untouched worlds per family and all four renderers, 96 sources total.

The split receipt must bind the domain separator, ordered world identities, ordered source hashes, family counts, and renderer factors. No outcome may influence the split.

Diagnostic matrix

Every scored arm must share the same eligible source bytes, independent target bits, initialization lineage, update schedule, optimizer semantics, precision, and persistent 1,536-byte machine. GD, GH, and O are train-only diagnostics and cannot contribute capability scores.

Sequential gates

Gate 0: measurement

The authenticated v6 canary must pass the independent closeout in train/assess_episode_functor_hankel_canary.py:

• exact package, source, arm, initialization, and canary hashes;
• exact two-update schedule;
• one H100 under Landlock and a loopback-only network namespace;
• finite positive gradient norm on both updates;
• changed trainable state;
• positive timing, memory, and optimizer-state measurements; and
• zero persisted weights, scored visibility, fit authority, or pretraining authority.

Failure leaves every fit and architecture escalation closed.

Gate 1: optimizer replay

Run O over 16 train sources and three seeds for at most 256 updates. Require:

• every component cross-entropy below 0.01;
• total loss reduced by at least 100x;
• 16/16 exact machines;
• 16/16 exact base and derivative codebooks; and
• zero nonfinite values or hard ties.

Failure diagnoses optimizer/loss/runtime mechanics. It is not evidence for undercapacity or against HSC.

Gate 2: deep-path microfit

For one fitting world, fit all four renderers with GD, GH, RD, and RH for at most 512 updates. Each arm must reach 4/4 exact keys, record labels, machines, base codebooks, and derivative codebooks.

If O passes but all deep arms fail, diagnose deep-path conditioning. Do not increase parameter count.

Gate 3: oracle-evidence mechanism

GD and GH must each reach 100% exact machines and codebooks on their train-only evidence. GD passing while GH fails rejects learned Hankel decoding independently of raw-source parsing.

Gate 4: raw fitting

Under one frozen 4,096-update budget, every scored seed must reach on F:

• at least 212/216 exact machines;
• at least 99.5% transition and observer cells;
• at least 90% complete base and derivative codebooks; and
• zero unhardenable rows.

Failure after Gates 1--3 pass diagnoses raw-source compilation.

Gate 5: train-only transfer

Without further optimization:

• R: 72/72 exact machines;
• W: at least 92/96 exact machines; and
• W: at least 14/16 exact machines within every action family.

Report exact key, record-type, occurrence-role, answer, transition, observer, base-codebook, and derivative-codebook metrics separately.

Gate 6: causal mechanism

On paired W rows, RH must beat the strongest of RD, RP, and RB by at least ten percentage points in exact-machine accuracy. The paired 99% confidence lower bound must exceed five points, and every seed must pass.

Failure rejects HSC as causally necessary under this board even if raw compilation succeeds.

Architecture-escalation triggers

No traditionally fixed transformer component may change merely because headroom exists.

• Attention: admit an attention treatment only if local record parsing is strong while distant-record binding is at least ten points worse, and a train-only oracle cross-record summary rescues at least ten points.
• Position/RoPE: admit a position treatment only if semantics-preserving offsets or chunk-boundary shifts cost at least five paired points while oracle-normalized evidence remains exact.
• Normalization/precision: admit a norm or precision treatment only if FP32 rescues at least ten points, or early-to-late gradient RMS remains below 1e-3 for at least 80% of updates across seeds.
• Recurrence: admit contradiction feedback only if parser and individual machine cells exceed 99.5% while complete machines remain below 95%, and a parameter-free contradiction revision improves at least ten points monotonically over one through three cycles.

Every admitted parent or compiler intervention is a new named adapted_base treatment. It must preserve the protected checkpoint tensors, account every added parameter and FLOP, stay below 200M complete parameters, and include a parameter/compute-matched control that destroys only the proposed causal signal.

Reserved machine-revision candidate

If parser fields and individual machine cells are strong but global consistency fails, the first bounded candidate is the Adjoint Causal-Syndrome Observer (ACSO).

For provisional transition and observer logits, ACSO recomputes the depth-three behavioral closure and separates base and derivative Jensen-Shannon innovations. An explicit reverse dynamic program maps those innovations back to the machine cells that caused them. A shared recurrent preconditioner sees only recoding-equivariant local logit, probability, adjoint, magnitude, moment, entropy, and cell-type features. It emits a positive bounded step along the true negative-adjoint direction for four tied revision cycles. Only the corrected ordinary 1,536-byte machine survives sealing.

One exact candidate budget is 3,995,137 parameters:

• 10 -> 384 -> 384 feature encoder: 152,064;
• GRUCell(384, 768): 2,658,816; and
• LayerNorm(768) -> 1536 -> 1 step head: 1,184,257.

The complete HSC+ACSO system would contain 199,488,246 parameters and leave 511,754 headroom. Its matched control cyclically scrambles multi-step action-position residuals before the same adjoint, preserving depth, tensor geometry, one-step information, residual norms, parameters, and four-cycle compute while destroying only true multi-step blame routing.

The prerequisite exact CPU audit is artifacts/r12/episode_functor_causal_syndrome_20260724.json. Across all 200 frozen worlds, all 26,400 permutation-preserving transition swaps and balance-preserving observer swaps produce nonzero and collision-free depth-three fingerprints: 132/132 unique faults per world, with 120--704 changed behavioral coordinates. Report payload SHA-256 is be260fda48585ff8aacc13369e8b01d80023729c944454388fe65cc37038b254; file SHA-256 is 4aaf86536899214c2d2bcce3516a65c6021a267989040a2f0b4982dc24c35ef2. This proves exact single-swap identifiability only. It does not show correction under noisy learned signatures or neural learnability.

The ACSO mechanics and exact controller constructor are now implemented in train/episode_functor_causal_syndrome_observer.py; integration and fitting remain unauthorized. The explicit adjoint is a hand-derived reverse dynamic program through the finite behavioral closure and does not invoke runtime autograd. The control is the exact gradient of its own scrambled objective and decreases that objective under a small negative-adjoint step; it is not a rerouted nonconservative vector field. The explicit adjoint runs under no_grad and cannot retain a target or compiler graph. A hand-calculated noncommutative machine anchors word-order semantics independently, the total is derived from the live HSC capacity receipt, and seal_primary_machine emits only the existing source-free HardFunctorMachine fields.

Two hostile-review rounds rejected earlier drafts for a nonconservative control, retained autograd graphs, an absent hard seal, coordinate-dependent hardening on tied rows, and incomplete independent derivative-prefix coverage. The current implementation closes those findings: sealing fails closed on every tied transition or observer row, tie-free hardening is exactly recoding equivariant, and the independent oracle covers both causal and cyclic derivative prefixes. Twenty-two focused tests and the complete 468-test relevant suite pass with 63 known nested-tensor warnings. Final hostile review reports no remaining P0/P1/P2, passes 40 additional randomized recoding trials exactly, and bounds manual-adjoint disagreement against autograd below 4e-9 across depths zero through five in both modes. It authorizes a mechanics-only commit, not integration or fitting.

The cyclic objective is retained as a structural falsifier but is not the oracle-recovery matched control because a correct causal machine is generally not its fixed point. The source-frozen recovery protocol is R12_EFC_ACSO_ORACLE_RECOVERY_PROTOCOL.md. Its oracle-fixed one-step control computes the full closure but masks evidence beyond one action before the reverse dynamic program. V2 audits all 672 frozen-board faults that are immediately observation-equivalent yet separable within suffix depth three, across three margins and deterministic recodings. V1 is void. The consumed v2 audit tests only target-informed deep-fault recovery mechanics, not HSC integration or fitting.

Consumed update: v2 returned deep_fault_oracle_no_go under exact source freeze 27d5c4b. Causal treatment and one-step control each recovered 0/672 faults at all three margins despite monotonic innovation and exact recoding. Post-hoc analysis found adverse correct-vs-wrong gradient contrast on all 672 faults; an oracle positive per-cell gate still failed all margin-0.20 faults. The present four-cycle positive-gradient ACSO preconditioner is killed. Any signed causal-retraction or second-order replacement requires a new named protocol and cannot inherit ACSO's mechanics authorization.

These results admit ACSO mechanics only. They do not establish that HSC emits usable noisy signatures, that the learned preconditioner improves a machine, or that a revised machine transfers. Kill ACSO if oracle machines cease to be fixed points after integration, innovation fails to fall monotonically, or it fails to beat both unmodified HSC and the cyclic-control adjoint by the frozen paired margin. Runtime autograd remains inadmissible in the deployed candidate unless separately accounted as a fixed inference resource without uncontrolled higher-order training.

Reserved source-attention candidate

If and only if the attention trigger fires, the bounded candidate is Hankel-Syndrome Reentrant Attention (HSRA):

• encode per-cell evidence/HSC disagreement, Hankel distance, distance margin, and row entropies with a shared 8 -> 192 -> 512 map;
• pull feedback to source records only through the model's own soft role and key assignments;
• re-run the eight tied source-encoder layers once;
• add low-rank contradiction-derived biases to attention logits; and
• harden and seal only the second-pass machine.

One exact candidate uses 4,303,048 new parameters, for 199,796,157 complete parameters and 203,843 headroom. Its isoparametric control deranges feedback across source records while preserving the exact vector multiset, magnitude, two-pass compute, initialization, and target budget.

HSRA remains unimplemented and unauthorized. Kill it if second-pass exact-machine accuracy improves by less than five points, treatment fails to beat record derangement by five points, feedback ablation costs less than five points, or initially incorrect rows do not reduce median syndrome by at least 25%.

Decision boundary

A complete train-only pass would establish that HSC learns transferable source-to-machine compilation within the frozen train world, renderer, and family support. It would not establish natural-language reasoning, broad task transfer, native language-model consumption of the sealed machine, or general reasoning. Those require separately frozen development, confirmation, and cross-task protocols.