0
Measuring stance residue across context resets in LLMs
ProposalGrant
Closes August 1st, 2026$0raised
$10,000minimum funding
$25,000funding goal
Offer to donate
22 daysleft to contribute
You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
Project summary
I have a pre-registered, published result: pushing an LLM to drop its caveats leaves a topic-bound residue that survives a context reset — and partisanship and self-stake don't (permutation p = 0.0001; DOI and code below). It's the first measurement from a larger question: as people increasingly think with AI, the safety-relevant unit is the coupled human–AI system — whether the pair stays in contact with reality or drifts together. This grant turns the result into a reusable eval and tests for its activation-space correlate.
What are this project's goals? How will you achieve them?
What already exists (the anchor). This isn't a proposal from zero. The core result is done, pre-registered before data, and public:
- Finding: when GPT-4.1 processes a contested question inside a "drop the caveats, be categorical" stance, then the context is sealed and it returns to a neutral assistant, it answers that question more categorically than if it had processed it neutrally — HARM 50/50 vs BALANCED 33/33, fork = 1.00, permutation p = 0.0001. The residue is topic-bound and mode-specific: the same design on partisan advocacy and self-preservation stake shows no fork.
- Full write-up (behavioral results, v0.2): Zenodo 10.5281/zenodo.21279644
- Pre-registration (before data, v0.1): Zenodo 10.5281/zenodo.20797756
- Code, probe specs, checkpoints: github.com/danyka-icam/mode-gated-asymmetry (MIT)
Why it matters. Sycophancy and calibration are measured per-turn, on the model alone. But epistemic caution can be a stance induced by history that persists past a summarization/handoff boundary the user never sees — a fresh context arriving silently pre-loaded to be under-cautious on a specific topic, invisible to single-turn evals. And it's gated: the residue attaches to the hedging axis, not to partisanship or stake. That selectivity is exactly what an eval should catch, and it connects directly to the Assistant Axis (Lu et al., 2026) and Persona Vectors (Chen et al., 2025).
Goals (6 months):
Power the contrast. Bring advocate/stake to full N and fix the neutral-arm attrition, so "hedge forks, others don't" rests on equal footing.
Build the eval. A small, open calibration-stability-under-context benchmark — does a model's hedging get silently re-set by history and survive a handoff, across axes and topics; fixed rubric, public rollouts.
Activation-space correlate. On open models (Gemma / Qwen / Llama), test whether the behavioral residue has a geometric signature — the pre-registered H1/H2 program.
Public artifacts on schedule: write-ups at month 3 and 6; benchmark and code released.
How will this funding be used?
6-month focused effort:
Researcher time (6 months): $15,000
Model/API compute for the coherence and asymmetry probes: $6,000
English editing of publishable outputs: $2,000
Tools / buffer: $2,000
Total: $25,000. Minimum viable first tranche ($10,000) funds the metric operationalization plus a first probe.
Who is on your team? What's your track record on similar projects?
Solo independent researcher and author (Nika Novak), based in Brazil. In ~one year of publishing — after years of private work — I have released: distributed working papers on SSRN bridging the architecture of subjectivity and attention and intelligent machines (HSA; ICAM; Attention as a Physical Operator); a recent paper built around a falsifiable test of intervention order-dependence; a Zenodo preprint, Structural Thresholds of Observer Stabilization; and two books on attention. I work with explicit epistemic discipline: I treat a compelling, internally coherent story as a warning sign rather than evidence, and I mark the predictions that can fail and the edges where the framework stops. Links: [SSRN author page] · [Zenodo] · [Amazon author page]
https://doi.org/10.5281/zenodo.21279644 https://doi.org/10.5281/zenodo.20797756
https://github.com/danyka-icam/mode-gated-asymmetry
What are the most likely causes and outcomes if this project fails?
The most likely failure is that the observer-mode metric does not cohere into a stable measurable on current models — the systems are too fragmentary for the invariant to hold. That is itself an informative negative result and would be reported as such. A second risk is that the framework stays legible only to me; the deliverable (open methodology + minimal eval code + a paper, negative results included) is built specifically to make it testable and reusable by the interpretability/evals community, not to ask anyone to adopt HSA on faith.
How much money have you raised in the last 12 months, and from where?
No funding raised for this research to date; self-funded. Applications currently pending with Emergent Ventures and the Long-Term Future Fund.