# ACSO Deep-Fault Oracle Recovery Protocol **Status:** v2 consumed; **NO-GO**. ## Retired v1 The uncommitted v1 draft is void. Its fixture accidentally read one confirmation world, its cyclic objective was not oracle-fixed, its receipt was too aggregated, and its tie/recoding gates were incomplete. No v1 official artifact was produced. V2 makes no unseen-board claim and uses every eligible deep fault on the already frozen 200-world mechanics board. ## Question Before ACSO is connected to HSC or fitted, does depth-three causal closure repair transition faults that an oracle-fixed one-step objective cannot distinguish? This is a targeted oracle upper-bound mechanics test. Failure kills the present multi-step revision rule. Success does not establish learnability, raw-source compilation, transfer, or reasoning. ## Frozen board and eligibility - Generate the existing 200-world identifiable board with exact seed `efc-identifiable-pilot-v1` and frozen 96/48/32/24 counts. - Consume every structurally unique world across all four split labels. Split labels have no inferential role because this is an exhaustive mechanics audit, not a fit or held-out capability claim. - Build target base signatures through action-word depth three and action-prefix derivative signatures through suffix depth three, hence total action depth four, independently from each exact finite machine. - A transition fault is eligible only when the wrong destination and correct destination have the same answer under **both** observers. Such a fault is invisible to all depth-zero/one observer behavior but can be separated by deeper future behavior. - Exclude observer-row faults and immediately distinguishable transition faults. They do not test the proposed multi-step causal signal. The frozen board contains 672 eligible faults across 88 worlds: 112 worlds have zero, 64 have six, and 24 have twelve. The implementation must recompute and verify this exact distribution and independently confirm that every eligible destination pair differs within suffix depth three before any official decision. ## Fault construction For every eligible fault and each margin in `{0.05, 0.10, 0.20}`: - initialize each correct transition/observer logit to `+margin` and every incorrect logit to `-margin`; and - swap the correct and selected wrong transition logits in exactly one row. Every initial hard machine must differ from its oracle in exactly one untied row. There are 2,016 primary cases and 2,016 recoded pairs. ## Revision arms Both arms start from byte-identical corrupted logits, compute the full depth-three closure, and run four cycles. - **Causal treatment:** optimize all base and action-prefix derivative signatures through depth three. - **One-step control:** optimize only base words of length zero/one and action-prefix derivatives of total length one. Deeper signatures are still computed but masked before the reverse dynamic program. The exact oracle is a fixed point of both objectives. For each categorical row, divide its logit adjoint by its maximum absolute entry; zero rows remain zero. Subtract `0.1` times that row-normalized direction. This positive diagonal preconditioner preserves descent for each arm's objective and is the strongest parameter-free member of the preregistered `[0.001, 0.1]` ACSO step range. There is no line search, outcome-dependent stopping, learned controller, or margin-specific tuning. ## Evidence For every fault, arm, recoding, and cycle zero through four, serialize: - base, derivative, and total innovation for that arm's objective; - exact full-machine recovery; - intended fault-row recovery; and - whether **any** transition or observer row is tied. The report also records per-world, per-margin, and global aggregates plus a SHA-256 over the ordered per-fault evidence. ## Recoding and execution controls Derive deterministic nonidentity state, action, observer, and answer permutations from each world identifier. Recode the exact machine and map each fault descriptor through those permutations. Require identical exact, fault-row, and all-row-tie decisions at every cycle, with innovation curves equal within `1e-6`. Run every fault independently with batch size one. Vectorized fault batches are prohibited in the official audit, eliminating batch-reduction dependence. ## Source and custody binding An official result is eligible only when: - seed, counts, margins, cycles, step, and thresholds equal this document; - all 200 worlds, 88 eligible worlds, and 672 faults are present; - the protocol, runner, ACSO implementation, board generator, and Hankel codebook match their Git `HEAD` blobs; - their file SHA-256 values, the Git commit, and a canonical board-manifest SHA-256 appear in the report; and - an exclusive reservation is durably created at a previously absent output path before evaluation, remains unchanged throughout evaluation, and the report is published with no-clobber hard-link semantics plus directory `fsync`; the inode-bound reservation remains beside the final report as a permanent custody receipt. Any nondefault seed, subset, source drift, dirty bound file, or altered threshold is fixture-only and cannot emit GO. ## Gates The current multi-step ACSO rule is **deep-fault oracle GO** only if: 1. every primary and recoded treatment total-innovation curve is nonincreasing within `1e-7`; 2. treatment exact and intended-row recovery are 100% at every margin and in every represented world; 3. treatment exact recovery exceeds one-step-control exact recovery by at least 80 percentage points at every margin and in every represented world; 4. no treatment tie appears in any row after the final cycle; 5. all primary/recoded decisions and innovation curves satisfy their gates; 6. every execution uses batch size one; 7. no nonfinite value appears; and 8. all source, board, count, and custody bindings pass. Any violation is **NO-GO** for integrating or fitting the current multi-step revision rule. Thresholds, margins, cycles, normalization, eligibility, and controls cannot change after source freeze without a new named protocol. ## Claim boundary A pass proves only that an explicit target-informed multi-step correction field repairs these bounded synthetic deep faults better than an oracle-fixed one-step ablation. It does not prove that Shohin infers target signatures, that HSC generalizes from source bytes, that the 3,995,137-parameter preconditioner learns useful scaling, that a sealed machine transfers to unseen task families, or that Shohin reasons natively. Pretraining remains prohibited. ## Consumed result The exact source freeze at Git commit `27d5c4bd00591fbafa3dffe68a4c209bda0e8099` produced `deep_fault_oracle_no_go`. All source, count, evidence, recoding, and output custody bindings passed. Across every margin, causal treatment recovered `0/672` exact machines and `0/672` intended rows after four cycles; the one-step control also recovered zero. Both primary and recoded treatment innovation decreased monotonically, no final tie appeared, no recoding decision differed, and maximum recoding innovation delta was `1.4901161193847656e-07`. Full artifact payload SHA-256 is `5b874fdabde074a2d29e142419291a2f24e84093c085f3d7b47870d550a33cb4`; ordered fault-evidence SHA-256 is `97b9d2f15daa7d83060db84dabe7c3a3edbfe85750ee38949017416187deef33`; full artifact file SHA-256 is `99572912b620959c4a635bbedef96eba9172824c7dad95463630ee4112754455`. Post-hoc localization found that the cycle-zero causal gradient favored the wrong destination over the correct destination on all 672 eligible faults. Even an oracle per-cell positive gate recovered 100% at margins `0.05` and `0.10` but 0% at margin `0.20` under the same four-cycle/`0.1` bound. These post-hoc results are diagnostic, not preregistered evidence. The consumed NO-GO kills the current four-cycle positive-gradient preconditioner. It does not kill HSC or a separately preregistered signed causal-retraction mechanism.