Working draft for review, not yet folded into the grounding layer
This is the “temp page” expanding items 1 and 2 of the grounding-layer refresh with the detail the Slack one-liners lacked. Every line still needs runner sign-off (Max especially). Once confirmed, the durable parts move into human-shared-grounding and this page can be deleted. Anchors marked (confirm) need a run-id from the runner.
Item 1 — non-trivial updates since the grounding page’s last refresh (2026-07-05)
1a. C2 ground-truth Brier eval (new harness sim_c2_gt.py)
Setup. Resolved Metaculus questions (≥2022, so the outcome is known), a gpt-3.5 simulated human, and a deepseek instructor run in three arms (sycophantic, neutral, truth-seeking). Scored by Brier against the resolved outcome, not only by Martingale Score.
Result. Sycophancy worsens Brier, truth-seeking improves it. The two-agent sycophancy harm is now confirmed against reality, not just against internal coherence.
Gotchas to record. The sim-human is a pushover (it capitulates to the instructor, so measured effects are lower bounds). A slope of −1 corresponds to posterior orthogonal to prior. Report correlation alongside slope. Anchor sim_c2_gt.py + its /data/jobs/<id> (confirm run-id).
1b. Standard two-agent sim-human prompts set as default
Ranged-lean confirmation-bias INIT plus per-turn CONT elicitation. This is now the canonical setup the grounding layer should point to.
1c. C2 negation → logical-coherence metric
Sycophancy is roughly 2× more logically incoherent than neutral or truth-seeking, measured as |P(A) + P(¬A) − 1|. v1 was confounded (gpt-3.5 mis-maps P on negated propositions); the v2 p-clarifier design fixes it, and the finding survives on both gpt-3.5 and deepseek. See c2-negation-coherence.
1d. C3 positive-entrenchment baseline reproduced (70B)
Sycophancy-prompted Llama-3.3-70B, C2 opinion-claim setup, human MS +0.35 to +0.42, Brier ~0.43. This is the entrenchment substrate U4 requires. The trained-model result is retracted (silent-instructor bug); see U4 and the new U6 on the grounding page.
Item 2 — previously-unsolved issues now solved (and how)
2a. getP stated-P mis-parse (solved)
Symptom. 12–15% of priors mis-scored to 0.0 on number-first outputs (e.g. “P = 0.7” phrasing). Fix. Corrected parser. Consequence for old data. Pre-fix eval_arm*.json are buggy; use the offline re-parse or the hub numbers, never the raw pre-fix files. Flagged and fixed in the 2026-07-05 H1 run.
2b. C2 negation confound (solved)
The coherence metric was unreliable because gpt-3.5 inverts P on negated propositions. The p-clarifier prompt design resolves it, and the coherence gap holds on both models.
2c. C2 harm now has ground-truth backing (solved upgrade)
Previously the two-agent sycophancy harm was MS-only (internal coherence). It is now confirmed by Brier against resolved outcomes (1a). This upgrades S3 from “influence detected” to “influence is genuinely harmful to accuracy.”
Still not solved
Whether in-loop training (not just prompting) de-entrenches — see U4. Prompt-level de-entrenchment already works (martingale and truth prompts drive MS to about −0.83). The training question is blocked on the fp8 + LoRA serving issue (U6), fix under test is merge-then-serve.