Validation and scope#

TensorCode provides operation, tool, persistence and training paths described below. Semantics still come from explicitly supplied models, training data, algorithms and policies; importing an operation does not install a cognitive ontology.

Compiled-program validation#

tests/test_program.py and TypeScript's test/program exercise bounded integer branch equivalence, loop expansion, state rollback and snapshots, native optimizer updates, artifact validation and cross-language execution. These are authored regression cases, not evidence of learned reasoning or general program synthesis. TypeScript's browser smoke check covers CPU, WASM and optional software WebGPU; it is an execution check, not a GPU speed benchmark. See programs for the commands and the exact numerical and gradient contracts.

Implemented surfaces#

Start with the documentation index for current API guides and runnable examples.

0.4 alpha pretrained vector checks#

The latent foundation report records bounded GB10 checks using pinned public weights. These validate integration and restoration, not a shared cognitive space or generalization:

Those measurements retain the pre-0.4.0a2 source revision. The separate a2 artifact recreation record documents new artifacts and old/new weight, output and training-continuation checks on the GB10. This is restoration validation, not a new pretrained quality evaluation.

The generated image depicts a red ceramic vessel but has malformed teapot geometry; native parity does not establish image quality. Arbitrary latent-to-language and latent-to-image bridges remain untrained until supplied paired supervision. The runnable example demonstrates initialization, collection, training, artifact saving and restoration; the operation guide specifies supported model families and inputs.

Pretrained tool measurements#

These are small, fixed-split experiments on public data, not general cognitive benchmarks. Complete model configurations, weights and evaluation records are hosted on Hugging Face. Each tool reconstructs its owned components without requiring a caller-supplied model.

The first Chatbot row is the original unbounded-workspace revision (d74b401, TensorCode 6607a8b). The second is its replacement on the bounded workspace update, trained with the same script defaults and seed on TensorCode 0.4.0a4. Seeds 8, 9 and 10 of the same recipe reach 31/64, 31/64 and 33/64 exact match; in all four seeds, bypassing the workspace gives lower cross-entropy, and in seed 10 bypass also has higher token F1. CUDA training is not bitwise reproducible (same-command weights differ by up to 2e-4 with identical predictions).

Chatbot inherits FLAN-T5-small language/instruction weights and uses oracle supporting passages supplied from annotations. This evaluates answering given relevant evidence, not retrieval, unrestricted conversation, or autonomous investigation. A bootstrap weight-alias bug was caught and corrected before this run; corrected foundation and workspace-bypass outputs were checked for exact parity before fine-tuning. Earlier invalid artifacts are not distributed.

The rankers inherit a frozen Electra-small encoder. All annotated supporting titles are treated as valid hypotheses. An authored token-overlap baseline scores 53.13% hit@1 and 42.97% recall@2. Decision exposes the same trained weights through the decision interface; it is not independently evaluated. Planner's labels measure document relevance, not observed causal utility of executed plans. A failed random-embedding pilot motivated this architecture; its evaluation subset was excluded from the final held-out subset.

These results establish owned pretrained model loading, parameter learning and fresh-process restoration on narrow tasks. They do not establish consistent benefit from the recurrent slot workspace, autonomous hypothesis discovery, calibrated uncertainty, or general reasoning. The workspace is an implemented, trainable research mechanism; schemas and gradient tests alone do not prove those capabilities.

Visual experiment: negative result#

Scene VSR results use 512 training and 128 held-out real COCO photographs with spatial-caption annotations. Image IDs are disjoint across splits, and the held-out photos exclude the earlier random-encoder pilot's test photos. The model owns frozen CLIP perception, positional image/text encodings, the shared workspace and a learned candidate scorer.

Accuracy falls from 55.47% to 50.78% after the prescribed ten training epochs. Blank-image accuracy is 53.91%, and zero-workspace accuracy is 52.34%. These results do not demonstrate useful visual grounding or scene reasoning. The checkpoint is an explicitly experimental negative result, not a recommended pretrained visual assistant. A prior random-patch model also failed (56.25% to 43.75% on a separate 64-photo evaluation).

The complete learned weights reproduce metrics after a fresh process restart. The full binary training checkpoint also restores optimizer state, RNG state and 5,120 training steps. This verifies the training/software lifecycle independently of the failed capability evaluation.

Learning from executed outcomes#

The service-recovery example collects 18 actual simulator transitions. Each training trace labels only the candidate that was executed, using its observed reward. After 324 small supervised updates, success changes from 0/6 to 6/6 evaluation scenarios and mean reward from -0.5 to 1.5. A fixed-action baseline and a policy that ignores changed state both score 0/6. Model, experience, optimizer continuation, session and trajectory reloads reproduce the recorded behavior. Full result.

The six scenario IDs are disjoint from training, but use the same three authored status classes and deterministic transition rules. This is a small mechanism test of feedback, learning and replanning, not a real-world benchmark, causal discovery result or test of novel action semantics. The production library does not import the example's environment or domain policy.

Source-wise verification and calibration#

The verifier run uses a pinned DeBERTa NLI foundation, 1,024 SNLI training pairs, 256 separate validation pairs for temperature fitting, and 256 test pairs. Training and real-model evaluation ran on the connected GB10. The foundation was already trained on SNLI/MultiNLI, so these splits establish isolation for this fine-tune, not previously unseen foundation data.

Fine-tuning lost one correct test prediction. Temperature fitting improved the reported probability metrics while preserving predictions; its value is 1.9768, fit on validation scores only. Full checkpoint reload reproduces logits exactly. Five explicitly authored diagnostic pairs also exercise entailment direction, contradiction and unrelated evidence. They are mechanism checks, not an additional statistical benchmark. NLI support is model inference, not a guarantee that a source is true, complete, current or trustworthy. Recalibrate after weight changes.

Owned visual language: inherited baseline#

Scene language results exercise an owned SmolVLM-256M model on 32 real VSR images. Spatial-caption yes/no accuracy is 20/32 (62.5%). With blank images, agreement with the original image labels is 15/32 (46.88%); with different images it is 13/32 (40.63%). Those controls measure original-label retention, not correctness against newly annotated altered images. The majority-label baseline is 17/32 (53.13%), just three fewer correct answers. The images do not overlap the earlier TensorCode visual training images; exposure during foundation pretraining is unknown.

This model was not fine-tuned. Its workspace residual starts at zero, so these results describe inherited VLM behavior and complete TensorCode ownership, not learned workspace improvement. Four additional scene descriptions are preserved for inspection without assigning truth scores. Three hit the 64-token evaluation limit, and blank inputs elicited hallucinated content. Outputs therefore remain explicitly unverified interpretations with full-image anchors and unknown confidence; they do not create accepted scene facts or symbolic graph structure.

Hypothesis generation and faithful realization#

Hypothesis training fine-tunes FLAN-T5-small on 1,024 human QA2D declarations, with 128 development and 128 test examples grouped by disjoint source articles. Inputs contain only the question and original SQuAD paragraph; answer annotations and target declarations are excluded. After three fixed epochs, test declaration exact match rises from 1/128 to 30/128 and token F1 from 19.97% to 82.92%. This is improved declaration generation given relevant evidence, not reliable factual inference: inspected outputs still change numbers and other facts. Foundation exposure to these datasets is unknown.

A post-hoc NLI audit accepts only 35/128 human references and 29/128 generated statements under the configured support policy. Even among untruncated pairs, only 26/94 human references pass. This exposes a verification/domain limitation; NLI acceptance cannot serve as answer accuracy. The report preserves raw outputs, truncation counts and the distinction between model support and truth.

Realization training addresses a separate failure: the decoder shortened accepted statements into fragments that failed subsequent verification. FLAN-T5-base learns to preserve an already selected statement using the production realization prompt. The target is deliberately present in that input; this is faithful rendering, not target-blind question answering. Three fixed epochs on 256 statements raise normalized statement exact match from 4.69% to 100% on 64 article-disjoint development examples. Verbatim match reaches 96.875%; the remaining differences remove backtick quotation marks. No separate test score is claimed for this component.

Both runs execute on the connected GB10 and restore complete model and optimizer checkpoints. Hypothesis continuation is bit-exact only in the explicitly recorded deterministic CUDA probe; the default CUDA continuation showed numerical variation despite equal initial state, batch and loss. None of these component measurements establishes an end-to-end reasoning improvement by itself. Training scripts received reporting-only corrections after these runs; recorded script hashes identify the versions actually executed, while current sources add configurable card metadata and truncation audits.

Complete cognitive pipeline: experimental result#

The final cognitive evaluation freezes the full configuration and component hashes before selecting 32 new HotpotQA validation questions. It combines the trained hypothesis generator and realizer above, Electra ranking, calibrated source-wise NLI, and an owned MiniLM sentence encoder. Questions receive oracle supporting passages. The prior 16-question run was used for diagnosis and component changes and is recorded as development evidence, not a final test.

On the final 32 questions, the chatbot returns 30 abstentions, one correct answer and one circular non-answer. Both non-abstained responses were manually checked against the original sources and gold answers. The correct response identifies James Franco; the other repeats “uppermost age range” without supplying an age. Passing NLI therefore does not establish relevance or answer completeness. Literal short-answer exact match is zero because the one correct response is a full sentence; it must not be confused with the source-reviewed 1/32 correct rate. There were no failed calls or generation truncations. All eight omission and all eight replacement controls abstained, but high abstention on original evidence prevents interpreting this alone as successful counterfactual reasoning.

Dedicated MiniLM retrieval finds a supporting passage at rank 1 for all 32 questions on this small oracle-passage corpus; a lexical baseline gets 31/32. Both reach 32/32 at rank 5. This inherits a pretrained sentence embedding space; no retrieval fine-tune or open-domain retrieval result is claimed. In a separate smoke test using the actual developer guide, evidence survived episode boundaries and exact session save/load and was retrieved again, but the chatbot still abstained on the question. That verifies memory behavior, not developer-document answering competence.

These complete artifacts demonstrate owned component execution, revisable source state, persistent retrieval and response screening. They remain experimental: source-wise verification rejects many useful statements, generation can invent facts, and screening admits non-answers. Reliable multi-source inference, answer completeness, useful response coverage and a demonstrated cognitive workspace advantage remain unsolved. The symbolic graph path is still a stub.

Bounded-workspace replacement#

The published checkpoint was replaced on main by a retrain on the bounded workspace update (revision 6fc386f; the result above remains at 8836ba5 for source 6607a8b). The generator and realizer were retrained; the verifier, Electra ranker and MiniLM encoder are byte-identical. It was re-run on the same 32 questions, which are now known cases rather than an untouched test.

It answers 3 and abstains on 29. Source review finds two correct answers (Catwoman/Pitof, and James Franco with a disfluent repetition) and one wrong factual answer: asked for a pilot's birth month, it gives the crash date, and the verifier supported it at 0.836. All omission and replacement controls abstained, retrieval is unchanged at 32/32, and complete receipts replay exactly after reload. On the developer-guide smoke it answered with a corrupted "nugging Face"; given the earlier guide excerpt it abstained, as the previous revision did.

The generator reaches 36/128 declaration exact match and 84.14% token F1 on its article-disjoint test (previously 30/128 and 82.92%). The recipe learning rate missed the F1 bar in three seeds; learning rate 5e-5 was selected by a validation rule written after all variants' test scores were known, then cleared the bar in two further seeds. The realizer again preserves 64/64 selected statements. None of this establishes a workspace advantage or reliable multi-source inference.

Cognitive coverage: development experiments#

The controlled verifier comparison reuses the previous 32-question final set as development data after inspection. It is not a new held-out improvement result. The proposal generator, ranker, workspace, realizer, token budget and screening thresholds remain fixed. Only the verifier foundation changes, with temperature fitted on 256 separate SNLI validation pairs. No classifier weights are trained in this comparison.

Assistant review against the supplied sources classified the responses as follows:

The replacement failed the predeclared promotion gate: incorrect/incomplete outputs increased from one to four. Omission, replacement and the authored conflicting-source controls abstained, but those controls do not establish answer completeness. Every non-verifier tensor stayed identical; complete artifact and probe-logit reloads were exact. The experimental artifact remains on GB10 and was not promoted or published as an improved checkpoint.

A separate question-answerability model rejected two non-answers but still gave high scores to the wrong surfing location and a magazine-name response where the question asked for its type. Those are post-hoc development scores, not a tested combined pipeline. Earlier source segmentation lost the correct James Franco proposal and admitted a repetitive non-answer. Neither diagnostic became a production fallback or relaxed threshold.

The next training target is evidence-conditioned completeness and constraint satisfaction. Source support, topical relevance and fluent realization each leave important failure modes. Planned fresh final cases remain untested because the candidate intervention failed development acceptance. All factual outcome labels here are source-grounded assistant review, not independent human annotation.

Actual learning across process restarts#

Banking77 restart results were produced by the executable example. Four subprocesses terminate in sequence: capture/save, baseline evaluation, reload/train/checkpoint, and final checkpoint evaluation. The run captures 79 experiences containing 9,997 explicitly labeled training rows. Six literal train/test overlaps are excluded. Vocabulary is built from training data only.

On 3,080 official held-out examples across 77 labels, accuracy changes from 0.97% to 89.48% and cross-entropy from 4.3646 to 0.4566. The model is a trainable 96-dimensional mean embedding and linear classifier, with fixed seed 7, Adam at 0.01 and 20 epochs. This demonstrates that persisted operation dependencies support actual supervised parameter updates after restart. It does not demonstrate autonomous target discovery, calibrated confidence, or arbitrary-program differentiation.

Dataset source: PolyAI Banking77. The result file records source hashes, process IDs/exit status, settings and command. Data, experiences and weights stay outside the repository.

Historical graph experiments#

The graph neural adapter and graph applications have been retired. The current graph API reserves symbolic operations and does not implement them. The measurements below describe the earlier implementation, not current capability.

The historical MUTAG example read the official TU dataset with a pinned archive hash. A fixed seed-7 split assigns 150 whole molecule graphs to training and 38 to evaluation. Training-only atom categories become supplied node features; adjacency drives two trainable message-passing steps, followed by mean pooling and classification. No node or graph appears in both splits.

Held-out cross-entropy changes from 0.68665 to 0.29464. Accuracy changes from 33/38 to 34/38, against a training-majority baseline of 26/38. This measured real parameter learning through the now-removed graph adapter, but the random initialization already classifies most test graphs correctly. One additional correct graph on this small split is weak accuracy evidence, not a benchmark result or evidence of inferred chemistry. That neural adapter preserved relation labels but did not use them. Full results and split IDs.

The retired dependency-impact example used Python AST import extraction and explicit graph callbacks. It performed static dependency analysis, not learned symbolic interpretation; there is no replacement graph application while symbolic operations remain unimplemented.

Actual multimodal model behavior#

The local evaluation sends real image bytes and text through ImageEncoder, Message, Transform, and structured message operations. Models are loaded explicitly from locally downloaded, pinned revisions. The public candy photograph has SHA256 fc417c899e94f8df465b7541c5a70f0eebb85c414d06345f0b290c061eccc84c. It contains five candies and visible fish-like printed symbols. This published model-card example is a smoke test, not unseen benchmark data.

SmolVLM results and Qwen results, including the follow-up preserve the actual outputs and failures. Final runs used the available NVIDIA GB10 device through the same optional local adapter. No semantic answer repair, confidence fabrication, parser relaxation or hidden fallback was applied. Token-budget exhaustion also raises an explicit error rather than returning a partial answer as complete.

The local adapter establishes a working supplied-model vision/language path. These runs do not establish reliable visual understanding or reliable local structured decisions. The current local adapter uses prompted JSON plus strict validation; it does not implement grammar-constrained decoding. A stronger or constrained backend is still needed for dependable use of these particular structured tasks. The tiny-model failures remain part of the evidence rather than being replaced by the stronger model's partial successes.

Verification and operational limits#

CI runs the complete suite and package build on Python 3.11, 3.12 and 3.13, plus a separate dependency-free wheel job. Core operations, training, provider adapters and the lazy tools namespace import without PyTorch or Transformers. Importing concrete owned tool classes requires the tools extra. See the test guide for local verification commands.

The HTTP adapters are exercised through local servers, including real request bytes, refusal/truncation, redirects, timeouts and invalid outputs. No OpenAI or TypeSafe credentials were configured; live hosted-provider quality and account-specific behavior remain unverified. The Jev adapter supports only its documented typed choice/score operations, not chat, images or retrieval.

Saved tensor artifacts restore on CPU. Complete trainer checkpoints cover supported module state and SGD/Adam/AdamW optimizer state, plus Python and PyTorch RNG state, module modes and explicit caller progress; external data-loader, NumPy RNG and scheduler state require separate handling. Data-only JSON avoids executable deserialization, but supplied codecs and configuration declarations are trusted application code, not a sandbox for arbitrary untrusted inputs. Replaying recorded external outputs holds them constant and does not rerun their effects.

Memory transactions and turn serialization cover one process; distributed locking is not implemented. Cancellation during commit settles that commit before releasing the turn lock. External effects from supplied callbacks cannot be undone by rolling back tool state. Symbolic graph operations remain unimplemented; graph records preserve caller-supplied structure without inferring semantics from free-form evidence.

Tools can own pretrained models and generate proposals internally. Training targets still require explicit evidence or feedback. Selection/retention policies and action authority remain explicit; symbolic graph operations remain stubs.

Earlier measurement#

The earlier in-process Banking77 run measured held-out accuracy increasing from 0.97% to 89.38%, with cross-entropy decreasing from 4.3646 to 0.4628. It used the same official split and overlap exclusions. Its older examples/banking77.py entrypoint is available in source history at 7827ac0; the current runnable path is the restart example. The newer restart measurement above is a separate run, not a replacement of the earlier evidence.

Response-quality training pilot#

The response-quality report records an experimental owned transformer with separate support, completeness and constraint heads. It trained on 55 assistant-reviewed natural candidates from 20 questions, with source-disjoint calibration (17 candidates / 6 questions) and development (16 / 6). These previously inspected examples are adaptation data, not a fresh final benchmark. Reference answers and reviewer explanations never enter model inputs. The five-epoch GB10 run made 70 optimizer updates; encoder and head weights changed, exact optimizer continuation passed, and a fresh process reproduced all 88 calibrated prediction receipts.

The checkpoint was not promoted. Support and constraints accept every development candidate; completeness accuracy is 0.50 versus 0.8125 for an all-positive baseline. Joint screening accepts five known-good development candidates and no known failures, but accepts two bad calibration candidates. It acts as a completeness filter on those partitions, with no demonstrated source-sensitive support or constraint rejection. This does not establish an improvement to the complete chatbot. No final questions were accessed and no tool defaults changed.

examples/cognition/train_response_quality.py reproduces preparation, explicit training, calibration and persistence checks. It is an experimental training runner using an internal assessor, not a supported pretrained response-quality tool. Its local foundation must have pinned Hugging Face download metadata and matching asset hashes. Further work needs broader reviewed supervision and evidence-use ablations before integrating a checkpoint into active tools.

Later quality comparisons expand to 550 reviewed candidates across 192 source-disjoint questions. Neither native paired-input encoding nor reviewed evidence interventions satisfies the admission gates. A frozen XL generative assessor also accepts known failures; workspace-only adaptation worsens rejection while reproducing exactly after reload. These experiments establish reproducible failures, not qualified screening components.

Joint native-foundation/workspace adaptation likewise approves all 91 development candidates, including 28 known failures; the adapted foundation with workspace bypass rejects all 91. Its native weights changed and exact fixed-next-batch optimizer continuation passed. This joint experiment does not isolate native-only training and provides no qualified replacement for current screening.

The conditioning diagnosis found an unbounded workspace residual over 1,000 times native token magnitude. After normalizing and bounding that residual, workspace-only retraining preserves varied judgments but still fails the unchanged admission requirements:

These are the same 91 reviewed development candidates, not a new final set. Both bounded runs reproduce exactly after reload. Joint retraining reduces accepted known failures from 14 to seven while preserving 49 accepted good answers. All three individual quality axes pass their fixed development criteria, but the combined gate still fails. Active and bypass paths make the same threshold decisions on every eligible calibration/development candidate; a workspace benefit has not been demonstrated.

Reviewed evidence controls expose a further failure: the jointly trained model approves all 12 original positive anchors, all 12 evidence-free variants, and nine of 12 unrelated-source variants. All inputs fit the token budget. Two independent assistant reviews supply the support labels; these are authored development interventions, not human gold or untouched final examples. Better ordinary error counts therefore do not establish reliable use of supplied evidence. These weights are not promoted.

The matched native-only training control starts from exactly the same foundation and baseline receipts, freezes all adapter weights, and trains through explicit workspace bypass. It retains 50 good answers but accepts nine known failures and seven unresolved cases on development. On calibration it accepts 32 good, six bad and one unresolved, versus joint training's 36, twelve and three. This is a mixed comparison, not consistent joint superiority. All individual development axes pass, but combined admission fails. Active/bypass decisions agree on all 175 eligible cases; all 183 saved receipts and the fixed next optimizer update reproduce exactly. Evidence controls also fail: eleven evidence-free and ten swapped-source variants pass, while all twelve positive anchors are retained. No qualification or promotion follows.

TRAIN-augmented native-only adaptation starts from the same original foundation and exact baseline receipts, with adapters frozen. Reviewed TRAIN evidence interventions and near-correct contrasts expand the corpus to 527 candidates; 37 exceed the token budget, leaving 1,054 supervised axis labels. Three epochs now require 792 updates, so this comparison does not separate added-data effects from additional optimization. Development accepts 49 reviewed-good, five known-failure and five unresolved candidates; calibration accepts 33, five and one. The unchanged per-axis criteria pass, but combined admission still fails. All 183 receipts and the fixed next optimizer update reproduce exactly; active/bypass decisions agree on all 175 eligible candidates.

On the unchanged reviewed source controls, both support-only and combined three-axis thresholds retain all 12 positive anchors and reject all 12 evidence-free and all 12 source-swapped variants, on both paths. All 36 inputs fit. The evaluator checks disjointness from the actual augmented training corpus and unchanged calibration/development bytes. This is measured improvement in evidence sensitivity on these reused, assistant-reviewed development interventions; it does not establish general grounding or a learned workspace contribution. The next work is to inspect the remaining combined-gate failures under the existing rubric, without changing thresholds or relabelling failures to pass. Complete public-tool behavior and a frozen, untouched final evaluation remain required. These weights are not qualified or promoted.

A frozen-weight instruction comparison tested one development-derived replacement that explicitly asks about every assertion and question restriction. Original-prompt receipts reproduce exactly for all 183 candidates and 36 source controls; model state is unchanged. The replacement regresses native-bypass joint admission:

Longer instructions increase calibration exclusions from eight to 14 and newly exclude three development inputs. Two newly excluded development candidates were previously accepted known failures; exclusion is not corrected reasoning. Full coverage denominators retain all candidates: calibration has 40 known-good candidates, versus 34 in the common subset. Both combined gates still fail. All 36 source controls fit under both prompt sets: support and joint thresholds retain all 12 positive anchors and reject all 24 negative variants on both paths. These are unchanged-weight responses to authored instructions, not learning or untouched final results. The replacement is rejected; original instructions remain the incumbent unqualified experiment. No second prompt, new run or integration follows. Any later fine-grained training or representation change needs a separate protocol; this negative result does not establish a foundation capacity limit.

The complete XL proposal pipeline returns seven correct answers, one wrong source attribution and 24 abstentions with joint verification. Changing only to source-wise verification returns 18 correct, two incorrect, one incomplete, one ambiguous and ten abstained responses. These 32 historical questions have been reused for development; neither comparison is untouched final validation. Source entailment alone does not establish whether an answer satisfies the question. Attempting to repair fragments through claim verbalization also fails: 25 nonfaithful rewrites pass NLI. No default or checkpoint is promoted.

Typed-decision decoding#

Zero-shot FLAN-T5-base compares owned structured operations in the default generated-JSON mode against decoding='likelihood'. No training, threshold or prompt selection occurred.

Mean-normalized likelihood gives 31.4% Banking77 accuracy (ECE 0.26) and identical response-quality rankings. Argmax response-quality accuracy is at or near the majority rate (support 0.78 vs 0.76, completeness 0.73 vs 0.73, constraints 0.69 vs 0.66) because true is usually preferred; the ranking signal needs a separately calibrated threshold. The 88 candidates are reused development data with assistant-reviewed labels, and the model is much smaller than the XL foundation used by the cognitive tools. This establishes that likelihood decoding makes an untrained foundation usable as a typed decision operation. It does not qualify a verifier and has not been connected to any tool.

Algorithm substrates and simulated joint control#

The algorithm framework executes a shared typed tensor artifact through symbolic, native differentiable and scheduled-attention substrates. Conversion tests cover outputs, explicit state, gradients, trained parameter export, source export and cross-language execution. A structural-bias adapter also preserves a pinned pretrained CLIP attention layer's output at zero bias; that check does not demonstrate learned binding or execution of the algorithm by CLIP.

The G1 policy comparison plots recorded sensor-driven walking and manipulation sessions with paired fixed-replay controls. Runtime trajectory retargeting is an authored algorithm; frozen locomotion is supplied by Unitree. Ground-truth simulator poses are used for offline scoring and replay, not controller targets. Session evidence retains failures and identifies the source revision of each recording. The separate GR00T task is a reference demonstration, not a matched benchmark.

The lab also exposes a guarded Cartesian G1 controller: geometric approach and alignment, bounded continuous IK, sampled tool clearance, fresh-vision descent gating and bilateral tactile verification before lift. The +4 cm pedestal case, which trajectory retargeting failed, succeeded in the recorded development run with a 15.16 cm lift and 3.52 seconds held above 10 cm. A shifted-target development run aborted on unexpected contact; its trace is retained. The revised controller uses a tighter descent tracking gate. See the current protocol and all outcomes for the complete regression and single combined-condition evaluation. The frozen controller completed all six regression conditions and the one fresh combined condition, with 14.74–14.82 cm peak lift and 3.40–3.42 seconds held above 10 cm. The shifted case took 41.38 simulated seconds versus 31.18 seconds nominal.

The optional portable quadratic iteration compiles through ALG.0/1/2. Its completed ALG.0 raised-pedestal run includes 587 convergence-gated SciPy fallback calls. The initial full ALG.2 rollout was stopped for interpreter latency and is not a completed physical result. The controller assumes calibrated sensing and known cube/support geometry; four clearance samples are not a continuous-path or whole-body safety certificate. These results establish a bounded simulation improvement, not learned ICL, hardware readiness or superiority to GR00T.

The handling extension composes walking, RGB-D pickup and receiver localization, friction grasp, transport, tactile support transfer, release and withdrawal. Local fingertip reaction forces provide an online payload-weight estimate and Jacobian-transpose arm feedforward. Neither object mass nor receiving-surface world coordinates enter the controller. The receiver detector assumes a stationary, axis-aligned, magenta 20 cm surface; the gripper and orange 5 cm cube geometry are known. Scene decoration is presentation only.

The recorded nominal handling episode completes in 48.04 simulated seconds, with 1.59 mm final horizontal placement error, 4.24 seconds continuously settled after release and 18.44 cm hand clearance. Its force observer estimates an 80 g payload as 80.15 g during the latter half of transport. These are single-episode measurements, not precision or generalization guarantees. The same frozen source also completes the changed-receiver and 200 g payload checks: 2.84 mm and 1.41 mm placement error respectively. The 200 g payload is estimated as 200.27 g. A matched 80 g run with load compensation disabled also completes; compensation reduces full-transport median tracking error from 7.50 mm to 2.23 mm in that pair.

Handling evidence retains every completed attempt, including unreachable low-receiver targets and a withdrawal timeout. Placement scoring requires receiving-surface contact, a stationary released cube, no pad contact and over 10 cm hand clearance continuously for at least one second, as well as controller completion. A trace validator recomputes those conditions; a cube dropped onto the tray during an abort is not counted as success.

Language, observed scenes and policy references#

The warehouse example adds typed sequence/parallel instruction trees, revisioned sensor observations, grounded questions and a dispatcher compiled through ALG.0/1/2. Its physical executor supports one parcel-to-receiver operation per calibrated workcell. A grounded task tree is not proof of motion feasibility. The matched final two-robot recordings both complete: parallel execution takes 52.30 simulated seconds, versus 100.70 for the sequential instruction (48.1% less elapsed time). Action-source hashes and initial physical state match. The second sequential assignment occurs one 20 ms control tick after the first controller completes. This measures concurrency across independent workcells, not improved manipulation quality. Parallel placement errors are 1.10 and 6.69 mm; sequential errors are 1.53 and 1.19 mm, with independent contact/release/settling checks. The separate three-G1 run executes the dispatcher through ALG.2 and completes all three placements in 55.38 seconds, with 1.10–1.76 mm XY errors. This is a single scaling demonstration; no matched substrate performance advantage is claimed. All three physical robots share one simulation and retain their own sensor/control traces. The optional local HTTP service previews recorded observations and never actuates robots. Unknown, ambiguous, stale and low-confidence references remain explicit.

A pinned Qwen3-VL-2B-Instruct model proposes compact task trees. On the fresh 12-request challenge it handles 9 correctly versus 3 for the restricted authored grammar; supported actions/questions score 6/7 versus 0/7. This is a small language interface evaluation, not learned manipulation. A temporal paraphrase incorrectly becomes parallel and passes the limited semantic guard. Conditional and deadline omissions are rejected. The language evidence retains these failures and separates development reuse from fresh challenges.

The inverse-geometry diagnostic fits known colored boxes to RGB-D without true poses as inputs. Across nine recorded handling frames, parcel center error is 1.734 mm versus 3.493 mm for corrected centroid geometry, while receiver error regresses from 0.264 mm to 0.702 mm. These are offline perception diagnostics; the existing sensor belief drives manipulation. Final-scene QA uses fresh rendered observations and abstains where fit confidence is insufficient.

The lab includes two freshly recorded successful native public Ψ₀ G1/Dex3 trials, and five original local Ψ₀.1 videos including three failures. The local pilot's complete 80-row comparison records Ψ₀.1 at 19/20 and 18/20 across its two levels, versus behavioral cloning at 20/20 and 19/20. Both use the native AMO whole-body tracker. These native tasks are separate references, not matched comparisons against the warehouse controller or evidence of superiority.

Physical Intelligence's released π₀ ALOHA checkpoint also runs through its native 14-action contract. Both recorded seeds lift with the right hand (native maximum reward 2) but fail the transfer within the official 300-step horizon: 0/2 native successes. Executed actions are checked against saved model predictions, and rewards are independently recomputed from contact records. This is an ALOHA arms reference, not π₀ controlling a humanoid. Failures remain visible in the lab.