# Learning DiscoSCM Experiment Ledger > Status: evidence ledger; the canonical E01 confirmatory matrix is complete > (`1,400/1,400`, zero failed fits) and its frozen-gate verdict is > `not_confirmed`. E06 passed an interface-utility gate but did not support > semantic-specific attribution. The latest E07 perception/conditional-response > diagnostic is complete with status `semantic_not_ready`; it is architectural > motivation only. > Canonical implementation: paper-local runner plus hash-pinned Unit Mechanism > Learning dependency in `implementation/dependency-lock.json`. ## Evidence classes | Class | Purpose | Promotion level | |---|---|---| | Analytic | verify formula, gradients, interface, and truth firewall | mechanics/proposition only | | Primary synthetic | test hidden same-token alternative-query laws | structural claim candidate | | Secondary tabular | factual prediction and contamination stress | predictive-objective context only | | Causal/semi-synthetic | test identified target under declared design | causal claim only within that design | ## E00. Contract and smoke checks **Status:** implementation tests exist; not paper-level performance evidence. Checks include: - point/Gaussian/Cauchy families are explicit; - alternative queries do not re-enter the encoder; - observation indices remain bookkeeping; - oracle token truth cannot enter learner payloads; - the comparison config includes a compatible DSCM-style baseline; - paper-local dependency hashes match the canonical UML implementation. Promotion: report only as reproducibility and mechanics verification. ## E01. Y-shuffle hidden response-curve robustness **Status:** canonical confirmatory run complete (`1,400/1,400`, zero failed fits); primary robustness was **not confirmed** by the frozen gate. Question: as learner-visible factual `(W,X) <-> Y` links are progressively broken without changing the outcome marginal, does Unit Abduction factorization retain an evaluator-hidden, evidence-indexed composed response law better than a capacity-matched direct predictor? Frozen design: ```text 6,000 tokens per world; train/validation/test = 60/20/20 one learner-visible factual event per token 21 evaluator-only queries on [-1,1] worlds = matched, no_heterogeneity, weak_evidence, narrow_support rho = 0,.1,.2,.3,.4 combined train+validation outcome-pool Sattolo subset derangement test and oracle remain clean and byte-fingerprinted ``` The seven-model matrix is `cauchy_unit`, `gaussian_unit`, `deterministic_unit`, `direct_cauchy_matched`, `direct_gaussian_matched`, `varying_cauchy`, and `cauchy_moe4`. The primary comparison is 5,267 versus 5,216 parameters, a gap below 1%. All inferential fits use AdamW with learning rate `1e-3`, weight decay `1e-5`, batch 256, at most 300 epochs, and patience 30. Oracle data are excluded from training, validation, early stopping, and selection. Execution stages: | Stage | Seeds | Rows | State | Promotion | |---|---:|---:|---|---| | mechanics smoke | 11 | 8 | complete | mechanics only | | development | 11,22,33,44,55 | 700 | complete | diagnostic only | | confirmatory | 101--110 | 1,400 | complete, zero failed fits | frozen verdict | Development must be complete and analyzed before a checkpoint can be created. The live implementation admits one canonical checkpoint and one absent output namespace; a failed or completed confirmatory attempt cannot be overwritten or reopened under another official namespace. Primary endpoint: hidden curve-location error in `matched, rho=.4`, with ```text delta_s = log(max(E_direct,s,1e-12) / max(E_unit,s,1e-12)). ``` The primary robustness gate requires mean delta above zero, the lower endpoint of a two-sided 95% t interval above zero, and a one-sided exact sign-test `p <= .05`; ties are discarded. The no-heterogeneity attribution guardrail requires the 95% upper bound of `1 - geometric_mean(E_unit/E_direct)` not to exceed 10%. The report separates a confirmed robustness result from permission to attribute that result to token heterogeneity; overall acceptance requires both. Secondary outputs include rho=.3 effects, rho=.4-versus-clean degradation, factual NLL/median error, pinball losses, PIT-ECE, coverage/width, CDF error, and narrow-support inside/outside curve error. Geometry, loss-family, and capacity comparisons receive separate within-family Holm correction over all predeclared secondary hypotheses. ### Confirmatory verdict The frozen confirmatory analysis returned `status=not_confirmed` and `success=false`. At `matched, rho=.4`, ```text mean delta = -0.00714785 95% t interval = [-0.0495877, 0.0352920] Unit wins = 4/10 seeds one-sided exact sign p = 0.828125 ``` All three primary requirements failed: the mean was not positive, the interval lower endpoint was not above zero, and the sign-test p-value was not at most `.05`. The point estimate corresponds to a 0.72% geometric disadvantage for the Unit model, not a robustness gain. The no-heterogeneity guardrail did pass: the 95% upper bound on apparent Unit advantage was 6.01%, below the predeclared 10% ceiling. This says that the run did not detect a material manufactured advantage when heterogeneity was removed. It cannot rescue the failed primary endpoint or provide evidence that token heterogeneity caused an advantage. Overall acceptance therefore remains false. ## E02. Gaussian-to-Cauchy pressure test **Status:** executed as supportive secondary analysis within E01; it does not confirm Cauchy geometry. Questions: - Does Cauchy retain calibration when factual token evidence is weak? - Does Gaussian become overconfident in remote-candidate worlds? - Does Cauchy pay an efficiency cost under matched local Gaussian uncertainty? - Are conclusions stable to invertible or normalized latent reparameterization? In the matched world at `rho=.3`, the Cauchy Unit model was favored over the Gaussian Unit model on the predeclared absolute factual, curve, and calibration endpoints after within-geometry-family Holm adjustment. For example, hidden curve-location error had mean log ratio `1.21548` and Holm-adjusted `p=0.02539`. The relative clean-to-40%-shuffle degradation contrast did not support Cauchy (`mean delta=-0.56276`, candidate-direction Holm `p=1`). The DGP in this world is itself Cauchy-matched, the findings are secondary, and no secondary endpoint can change the failed primary gate. They are therefore supportive model-family diagnostics, not a general confirmation of Cauchy epistemic geometry, latent semantics, or causal validity. A Gaussian win remains a valid result rather than a protocol failure. ## E03. Broader mechanism-robustness extension **Status:** deferred beyond the bounded Y-shuffle experiment. E01 already reports heterogeneous/no-heterogeneity, informative/weak evidence, and supported/unsupported query regions. A later extension may add: - matched versus nonlinear misspecified mechanisms; - outcome-contamination families beyond marginal-preserving Y-shuffle; - compatible DSCM-style fixed-event-noise targets. No pooled average may hide a failure in no-heterogeneity or unsupported-query cells. ## E04. Secondary factual robustness **Status:** existing evidence may be referenced only after provenance review; no historical tabular result is imported into the V3 manuscript. Allowed interpretation: ```text Heavy-tailed likelihoods can exhibit useful factual contamination behavior. ``` Forbidden interpretation: ```text Tabular corruption proves token abduction, factorization, or causal validity. ``` ## E05. Selected-token Y-shuffle target correction **Status:** semantic gate complete; `development_ready=false`; development and confirmatory were not run. E05 is independent of the frozen E01 evidence. It uses the clarified single- selector ontology, sets factual `O=X=x`, keeps `selected_code` evaluator-only, and scores a 21-point response slice with the simulator-selected code fixed. Train corruption is action-stratified and nested, so it primarily breaks the token-cue/outcome link while preserving each action stratum's outcome multiset. The canonical semantic matrix completed `72/72` formal cells and `18/18` oracle-isolated diagnostics. Gold evaluator error was zero. With the true shared mechanism frozen, the learned abductor approached the evidence-only floor (maximum normalized excess `0.00729`). The both-learned Unit model failed the clean recovery gates: matched normalized excess was `0.5307` and the no-heterogeneity normalized error was `0.6808`, both above the frozen `0.20` ceiling. Factual NLL did not reliably distinguish token-recovering and non-token initialization basins. Therefore E05 stops before a robustness claim. Its supported diagnostic result is that the selected-token inference path is learnable when the mechanism is fixed, while the jointly learned one-shot diagonal factorization is not yet stable enough to evaluate Y-shuffle superiority. See `selected-token-experiment-v2-zh.md` and `implementation/evidence/yshuffle-selected-token-v2/semantic/analysis/verdict.json`. ## E06. Selected-token perception/abduction interface repair V3 **Status:** sealed development diagnostic complete; interface utility passed; semantic-specific attribution not supported; not confirmatory. E06 preserves E05 as immutable evidence and opens a new matched-capacity namespace. It tests one repaired boundary: the abduction encoder receives the token cue `z` but not the manipulable action coordinate `a`, while the shared mechanism continues to receive the complete query. The `full_x` and `z_only` arms have identical parameter counts, initial states, model seeds, epoch counts, and exact minibatch orders. Both reuse one factual belief object over the full query grid; `selected_code` remains evaluator-only. Clean execution completed `36/36` formal cells and `36/36` frozen-mechanism diagnostics. In `matched_identifiable`, fixed-realization response-curve MAE fell from `0.223607` to `0.030654`, an `86.291%` relative improvement, and `z_only` action sensitivity was exactly zero. The interface gate therefore passed. The gain was not specific to token information. The absolute `z_only` gain was `0.192953` in matched data but `0.234974` after removing token heterogeneity, so the matched-minus-control interaction was `-0.042021`. The frozen-mechanism belief diagnostic also found a strongly underdispersed Cauchy approximation: matched `z_only` 80% coverage was `0.3822` versus `0.7917` for the exact truncated-Uniform posterior, with half-IQR `0.01216` versus `0.04832`. Consequently E06 supports an action-invariant perception/abduction interface repair, not recovery or calibration of \(Q_i\approx P(U\mid O=o_i)\). The interface gate authorized a paired `rho=.4` stress diagnostic. There, matched `z_only` remained absolutely better (`0.162978` versus `0.234197`), but it degraded by `0.132324` from clean versus `0.010590` for `full_x`. This is not a robustness result and cannot upgrade the failed semantic- specificity guard. See `selected-token-semantic-repair-v3-results-zh.md` and `implementation/evidence/selected-token-semantic-repair-v3-sealed/`. ## E07. Perception + conditional response P0 **Status:** preconfirmatory descriptive semantic diagnostic complete; `semantic_not_ready`; architectural motivation only. E07 keeps one factual `(O,x,y)` record per token and one cached abduction object across a hidden same-unit query surface. It separates two architectural questions that the raw bilinear model confounded: - whether raw query coordinates need explicit perception before entering the response mechanism; - whether the response should be nonlinear in perceived query coordinates while remaining affine in the token candidate. The frozen matrix compares six globally parameter-matched arms across `warped_affine`, `warped_nonlinear`, and `no_unit_heterogeneity`. All `108/108` formal fits and `48/48` nonranking diagnostics completed. Explicit perception and conditional response forms improved hidden surfaces relative to the raw bilinear model in the tested warped worlds; `flow_conditional` also beat the capacity-matched direct predictor in one predeclared contrast. The semantic gates did not pass. The invertible-flow arm did not outperform a residual-MLP perception map, several predeclared `0.10` improvement gates were missed, learned-abductor/frozen-mechanism and gold-unit/learned-mechanism paths remained above their `0.05` ceilings, bad initialization basins persisted, and the no-unit control retained spurious evidence-dependent response variation. The run therefore stopped before shuffle, additional seeds, or confirmation. Allowed interpretation: the diagnostics motivate retaining explicit query perception and a conditional affine-in-unit response family, while locating response learning and unit-effect regularization as unresolved bottlenecks. Forbidden interpretation: E07 validates P--A--R, proves invertibility necessary, recovers `Q_i`, confirms a factorized-model advantage, or supports causal or counterfactual identification. Canonical report: `perception-conditional-p0-results-zh.md`; config and evidence are under `implementation/evidence/perception-conditional-p0-semantic-v1/` and `implementation/evidence/perception-conditional-p0-semantic-v1-analysis/`. ## Artifact requirements Every future run must preserve: - exact config and seed; - split and oracle-firewall fingerprints; - source/dependency hashes; - raw per-seed metrics; - aggregation and multiplicity policy; - development/confirmatory label; - runtime and environment; - explicit failure status. No result enters `claim-register.md` as empirical support until these fields are present and `validation-criteria.md` is satisfied.