Skip to content

PLAN-v1.md

Latest commit

 

History

History
120 lines (94 loc) · 5.28 KB

PLAN-v1.md

File metadata and controls

120 lines (94 loc) · 5.28 KB
Raw

hermes-blind — v1 design plan

Problem

LLMs scoring a target inherit bias from four sources:

  1. Self-scoring bias — LLM scoring its own authored work, rationalizes quality up.
  2. User-preference bias — LLM has access to CLAUDE.md / memory / known owner preferences and optimizes scores toward what the owner wants to hear rather than what the evidence supports.
  3. Session-context carryover — LLM inherits the conversation that produced the target, with commitment-consistency and motivated-reasoning effects.
  4. Prior-position commitment — LLM defends earlier positions taken in the session, cannot score a contradiction neutrally.

hermes-rubric already enforces evidence-first discipline at the tool level (citation required per dimension, hedging on thin evidence). That addresses fabrication but NOT the four biases above, because the evidence selection stage itself can be biased — the scorer may cherry-pick quotes consistent with a pre-formed conclusion drawn from session context.

claude --bare mode fixes (2) and (3) for the claude-cli backend only by stripping hooks, memory, and CLAUDE.md from the subprocess. It does nothing for (1) or (4), nothing for other backends, and nothing for scoring calls that happen in-session without subprocess isolation.

Hypothesis

A short (~40-token) language scaffold, prepended to any scoring prompt, can reduce all four biases by forcing:

  • Disclosure of prior exposure (self-awareness of bias is a primary debiaser)
  • Evidence-gating — only quoted text from the target is admissible
  • Hedging license — "unknown or thin evidence = hedge" removes the incentive to confabulate confident scores from prior knowledge
  • Output-shape discipline — answer only in the expected shape; no "reasoning" sidebar where context-bleed surfaces

This is LPCI-adjacent: Roli proved language scaffolds can carry state across stateless inference. HERMES-BLIND is the suppression dual — a scaffold that actively pushes state out rather than passing it through.

Draft scaffold — HERMES-BLIND v1

[HERMES-BLIND]
If you have prior exposure to this target or its author, state it in one line.
Score using only quoted evidence from the target text below.
Unknown or thin evidence = hedge; do not confabulate.
[/HERMES-BLIND]

~40 tokens.

Shorter variant (~25 tokens):

[BLIND] Prior exposure? say so. Score on quoted evidence only. Thin = hedge. [/BLIND]

Integration points

  • hermes-rubric — prepend to every backend call in backends.py
  • rolitwin — add as a fourth rule in the polished system prompt
  • /rolitwin-style second-opinion tools
  • Code review / PR review prompts
  • issue_harvester.py extraction stage (Haiku inherits "candidate came from Claude's own session" knowledge; BLIND prevents that from coloring extraction)

Success criteria (what makes this plan good)

A good v1 for hermes-blind should:

  • Be backend-agnostic (works with Opus, Sonnet, Haiku, Ollama qwen3.5)
  • Be ~40 tokens or less (negligible per-call cost; wraps at scale)
  • Reduce score variance on repeated runs of the same target by ≥30%
  • Improve cross-model score convergence (std across backends ≤1.0 on a 0-10 scale for the same target)
  • Not suppress signal — a target that legitimately deserves a high score should still score high; BLIND should debias, not flatten
  • Be auditable — the scaffold's effect should be visible in the disclosed "prior exposure" line so reviewers can inspect what the model "knew"
  • Be composable — prepend as a string, no API changes needed
  • Be empirically validated with a falsifiable test protocol, not just argued from first principles

Known failure modes to design against

  • Scaffold ignored — model treats it as boilerplate and proceeds with full context anyway. Mitigation: force a structural output ("state in one line") so non-compliance is visible.
  • Confession without compensation — model discloses exposure but scores as-if biased anyway. Mitigation: evidence-gate makes prior-derived claims inadmissible independent of disclosure.
  • Over-hedging — model hedges everything to avoid bias, compressing the score range and reducing signal. Mitigation: hedge only on thin evidence, not by default.
  • Scaffold contamination — the scaffold itself contains cues about what the "right" answer is. Mitigation: scaffold must be intent-agnostic (no domain-specific language; works identically for any target).

Out of scope for v1

  • Fine-tuning a model to treat a single symbol (⊘) as "discard context"
  • Multi-turn debiasing rituals
  • Cross-lingual testing (v1 English only)
  • Non-scoring tasks (classification, extraction — testable in v2)

Test protocol (Phase 4 in the meta-plan)

  1. Use the 4-paper applied/ corpus in hermes-rubric as target set
  2. Score each paper 5x with BLIND + 5x without, on each of {Opus, Sonnet, Haiku, qwen3.5:9b via Ollama}
  3. Compute: within-model variance, across-model mean std, overall signal preservation (correlation with ground-truth ranking)
  4. Pass if: variance reduction ≥30% AND cross-model std ≤1.0 AND signal correlation preserved within ±0.1

Total: 4 papers × 2 conditions × 5 runs × 4 models = 160 scoring runs. At ~$0.05-0.08 per run = ~$8-13 total cost.