tensorcode-investigator-cognitive-experimental-001

Experimental TensorCode cognitive Investigator

This revision replaces the earlier weights on main. It is the Investigator component extracted from the bounded-workspace replacement of jacob-valdez/tensorcode-chatbot-cognitive-experimental-001 (revision 6fc386fbfe992b8fb02c9049d4a761ed51c54132). The earlier weights were trained before the bounded workspace update (memory_update="relative_rms_bounded") and load only with TensorCode source commit 6607a8b. Revision a5ee35f6c644fa13d4bda5850ea37eb32ec3f6d2 remains available, unchanged, for that runtime.

This component checkpoint owns its hypothesis generator, source verifier, ranking operations, workspace and retrieval encoder. It was produced with bot.investigator.save_pretrained(...) on the frozen replacement Chatbot; no separate training was done. Its direct investigate call returns proposals and source-wise NLI scores; the Chatbot adds its final selection and realization policy and owns the realizer, which is not included here. Loading requires TensorCode; it does not call a hosted inference provider. Evidence is supplied explicitly by the caller. Generated statements remain hypotheses.

The proposal generator (flan-t5-small) was retrained on the bounded architecture on human QA2D declarative targets, with foundation learning rate 5e-5 (the previous revision used 3e-5; see the Chatbot card for how this was chosen). The SNLI verifier with held-out temperature calibration, the Electra ranker and the MiniLM retrieval encoder are reused byte-for-byte. Authored thresholds screen the verifier's judgments; SNLI calibration does not establish calibrated truth probabilities for evidence QA.

Use

from tensorcode.tools.investigator import Investigator
model = Investigator.from_pretrained("jacob-valdez/tensorcode-investigator-cognitive-experimental-001")
result = model.investigate({
    "question": "Your question",
    "evidence": [{"source_id": "document-name", "text": "Actual source text"}],
})
print(result)

Pin the commit hash from this repository's history for reproducible loading. Saving model weights excludes sessions and conversations.

Evaluation

There is no independent Investigator answer benchmark. The pipeline evaluation below applies to the enclosing Chatbot; it is not a claim that the Investigator alone realizes answers.

Component qualification (investigator-qualification.json), in fresh processes with local_files_only on an NVIDIA GB10:

• The state dict (647 tensors) is bitwise equal to the replacement Chatbot's Investigator component, and the configuration equals the Chatbot's nested cognition.investigator configuration (memory_update: relative_rms_bounded).
• investigate() receipts equal the owned Chatbot component's receipts on all 32 evaluation cases.
• save_pretrained then from_pretrained gives a bitwise-equal state dict, identical configuration bytes and equal receipts on the correct-answer cases.

Enclosing Chatbot, on the same 32 HotpotQA distractor-validation questions with oracle supporting passages as the previous revision (reused known cases, not a held-out test): 3 answered and 29 abstained. Source-grounded assistant review found 2 correct answers (Catwoman/Pitof; James Franco, disfluent) and 1 incorrect answer (EgyptAir 990: it gives the crash date, October 1999, instead of the relief first officer's birth month, February; the verifier falsely supported it). The previous revision gave 1 correct answer, 1 circular non-answer and 30 abstentions. All 8 source-omission and 8 source-replacement controls abstained. Owned MiniLM retrieval over the 64 oracle passages reached 100% top-1 and top-5.

Generator on the document-disjoint QA2D/SQuAD test set: 36/128 exact declarations, token F1 0.8414 (previous revision 30/128, 0.8292).

evaluation.json is the Chatbot's raw report; the other evaluation and training records match the Chatbot repository.

Limitations

This model does not establish general cognition, reliable multi-hop reasoning or factual guarantees. Source-wise NLI can reject valid cross-source conclusions and can accept wrong or irrelevant statements. The system's scores and authored screening are not ground truth. The evidence-QA and episodic retrieval evaluations use small oracle corpora. Reviews are by assistants, not independent human annotation.

Foundation licenses and dataset terms remain applicable. SQuAD is CC-BY-SA-4.0, and the QA2D mirror declares MIT.

Owned components and provenance

• Proposal foundation: google/flan-t5-small, revision 0fc9ddf78a1e988dac52e2dac162b0ede4fd74ab; adapted on 1024 human QA2D declarations joined to original SQuAD contexts (hypotheses-qa2d.json).
• Verifier: cross-encoder/nli-deberta-v3-small, revision fa2804872c3b4bd748f38c0185cc85775361e735; SNLI adaptation and held-out temperature calibration, unchanged (verifier-snli.json).
• Retrieval: sentence-transformers/all-MiniLM-L6-v2, revision 1110a243fdf4706b3f48f1d95db1a4f5529b4d41; masked-mean pooling, L2 normalization, unchanged.
• Ranker: jacob-valdez/tensorcode-investigator-hotpot-001, revision 1bc225917c3646fcb9702df91ff5e445846c1dc7, unchanged.
• Weights sha256 dce1d1b77d7eb2a0004cda91d7cedf64645e1ea727af65e55524a1367e18c697.

See final-freeze.json for component, configuration and source hashes, data-manifest.json for the evaluation cases, and manual-factual-review.json and fresh-process-replay.json for the review and replay. Source version 090ebc4. No parameters or policy thresholds changed after the final evaluation began.