Sīdus · E0-B research record
What Does a Monitoring Signal Represent?
Definition-indexed discrimination of correctness monitoring, action value, process representation, and redescription
Research question
When does a signal represent a property or process of another representation, rather than merely tracking information, predicting reward, or furnishing a useful feature?
Three operational criteria were tested on one exactly specified synthetic environment: correctness monitoring motivated by Shea’s account, representational redescription (RR), and process representation. Each was judged separately; their disagreement is part of the result, not a majority vote.
Shea · correctness
Validated operationally, at the functional level. A trace-compliant constructed code passed every registered clause; correctness-supervised training met the frozen seed-mean criteria; reward-only training of the same architecture did not.
RR · redescription
Inconclusive: non-identified. The accessibility gain is real, but matched feature engineering reproduces it, and the procedure-coupling clause depends on which admissible consumer is frozen.
Process
Inconclusive: non-identified in U1. In the registered mapping family the held-out response is a linear function of the observed responses, so a per-cell tracker matches the structured model.
Learned confirmation · seeds 1001–1020
| Hypothesis | Frozen outcome | Seeds individually meeting it |
|---|---|---|
| H1 · correctness-supervised code meets all six functional clauses | Confirmed (seed mean) | 14 / 20 |
| H2 · reward-only scalar fails in the action-value direction | Confirmed | 19 / 20 |
| H3 · four-dimensional end-to-end code carries correctness-relevant information while retaining payoff context | Confirmed | 17 / 20 |
Confirmation is a statement about seed-level means under thresholds frozen before the reserved seeds were run. For 14/20 individual successes, the one-sided 95% lower bound on per-run reliability is about 0.49.
Established
A trained system can satisfy a specified functional correctness-monitoring profile; the registered reward-only scalar generally did not under the tested procedure.
Not established
Unique metarepresentational content; identified RR or Process metarepresentation in U1; an architecture-wide learning law; biological, ecological or consciousness claims.
Next
E1: resource-bounded allocation among several useful monitors, preserving the distinctions between availability, use and content. E0-B was never a gate on E1.
Interactive companion: definitions, fixture, formula lab, witnesses and seed-level results.
Paper source (Markdown)
Research materials
- Experimental archive v4 (ZIP, 2.9 MB): source scripts, contract tables, development ledgers D–G, frozen confirmation specification, seed-level results and adjudication. 115-entry internal SHA-256 manifest.
SHA-256 64f2221a1caf3a838f7ef78d00327f34ab04783842baad2da1edecb97131dd66 - Selected records, published byte-for-byte from the archive: Closing Report · frozen specification · confirmation rules · seed-level results · adjudication
- Post-confirmation computational validation (ChatGPT-assisted): integrity check of all 115 archive hashes, H1–H3 bounds recomputed from the seed-level results, and sensitivity checks. Kept outside archive v4. Not an execution-level reproduction: its attempted full rerun did not complete.
SHA-256 886bb51b40b46d2cf667b321b10e855ce0e80d38108820b9bb5bdd04af6aa8a3 - Publication addendum: how the released files relate to one another, the terms used for each kind of check, the author’s reading of the open Closing Report note, and the exact changes from the 8 October paper (8 October PDF, kept unchanged).
- Reproduce: in the environment recorded in ENVIRONMENT.txt, run
python3 e0b_confirm.py confirmand compareconfirm_results.jsonandconfirm_adjudication.jsonwith the archived files.
Provenance: an internally frozen, prospective test of pilot-derived hypotheses. Commitment timestamps are self-recorded. No independent execution-level reproduction or audit by unaffiliated researchers is claimed.
Permanent record: Zenodo · DOI 10.5281/zenodo.23252262
Paper, unchanged archive v4, validation record and publication addendum, independently hosted. This page mirrors them.
Cite as: Amuwo, B. (2026). What does a monitoring signal represent? Definition-indexed discrimination of correctness monitoring, action value, process representation, and redescription (Sīdus E0-B research record, version of record). Zenodo. https://doi.org/10.5281/zenodo.23252262