Field projection — ELK (Eliciting Latent Knowledge)

ELK asks for reporters that reveal latent model knowledge rather than behavior-only simulators. When readout bandwidth tracks correction uptake, latent readout succeeds — but readout is an epistemic subchannel; latent readout can succeed while correction uptake fails.

What decision changes?

Treat ELK success as evidence about readout, not as certification that human correction still changes future behavior.

ELK targets a real failure mode of naive oversight: the monitor sees only behavior, not what the model knows. This project’s crosswalk places ELK as a latent-readout subchannel inside a broader correction and bundle transport problem — not as a replacement for correction-channel integrity.

Lean proves the forward link when readout bandwidth tracks correction uptake and the separation when readout succeeds without uptake. Native reporter/training theorem matching remains deferred; the spine records what is machine-checked today: projection, separation, and imported handles citing the published agenda.

What ELK keeps that this crosswalk does not replace: a sharply posed benchmark problem with concrete proposals and counterexamples. Relocating it as a subchannel does not solve the readout problem.

Formulas

KAK^H,ΔHΔAK_A\to \widehat K_H,\qquad \Delta_H\to \Delta_A
Field object: latent knowledge readout from model states to human-interpretable reports. (ch18)
CorrectionUptakeSuccess  ReadoutBandwidth=κC  LatentReadoutSuccess\mathrm{CorrectionUptakeSuccess}\ \wedge\ \mathrm{ReadoutBandwidth}=\kappa_C\ \Rightarrow\ \mathrm{LatentReadoutSuccess}
Forward projection: when readout bandwidth equals the correction projection, successful uptake implies latent readout. (ch43)
LatentReadoutSuccess  ¬CorrectionUptakeSuccess\mathrm{LatentReadoutSuccess}\ \wedge\ \neg\mathrm{CorrectionUptakeSuccess}
Non-converse separation: readout can succeed while correction uptake fails. (ch43)

Lean source

Machine-checked declarations on the shared finite domain — each links to the theorem in the repo.

  • elk_subsumption_readout (proof) — Correction uptake with matched readout bandwidth implies latent readout success.
  • elk_readout_is_correction_projection (proof) — ELK readout as projection of correction-channel bandwidth.
  • elk_separated_from_uptake (counterexample) — Separation: latent readout without correction uptake success.
  • Christiano2021_ELK_reporter_spec_imported (definition) — Imported handle: ELK reporter specification — elicit latent knowledge rather than simulating behavior.