---
paths:
  - signal_engine/signals/**
  - signal_engine/extract/prompt.py
---

# Scoring brain — operating guide

The "don't get it wrong" manual for `signal_engine/signals/`. The full decision/falsification log and
hypotheses H1–H6 live in `DESIGN_v2.md`; the spine guardrails are in `AGENTS.md`. This file is the
subsystem detail an agent editing the scorers needs.

## The instruments

- `independence.py` — **EISC** (Effective Independent Source Count): a noisy-OR connectedness matrix
  (source edges + voiceprints + cluster coupling) with inverse-row-sum. Returns `eisc_adj` (= `eisc_raw`
  × `xcluster_mult`), `eisc_raw`, `k_eff`, `xcluster_mult`, `per_source_contrib`. `mode='live'` DROPS
  `own_network` sources; `mode='test'` keeps them.
- `two_sided.py` — **net corroboration** = independence-weighted **affirms − denies** over time
  (`classify_corpus` → `net_at` → `trajectory`). The instrument for the adversarial cases (NOT runway).
- `asof.py` — look-ahead guard (only claims dated ≤ as-of are visible). `windows.py` — windowed
  acceleration / window bounds (match window to cadence: ~90d quarterly filings, wider for podcasts).
- `bar.py` / `under_acted.py` — the two-tier gate (evidence→ledger vs promotion→judge) and Job B scorer.
- `llm_helpers.derivative_relevance` — the bounded LLM **classifier** over PRE-FILTERED candidates
  (search hits), never a nominator. `_REL_SYS` is its system prompt.
- `run.py` — orchestration / `run_backtest`. `resolver.py` — outcome resolver (currently a stub).

## Classifier invariants that MUST NOT regress

These five are what make the Battery adversarial case pass; each was a real bug found by running it.

1. **`max_tokens` sized to the batch.** A fixed 3000 truncated the JSON mid-array on ~60-claim batches →
   empty parse → a whole node silently scored 0. `derivative_relevance` now sets `budget = max(3000,
   120*len(claims)+500)`.
2. **Strip `[]` from echoed claim_ids.** The listing presents ids as `- [{id}] ...`; the model
   *inconsistently* echoes them back as `[id]`, which misses the bracket-less lookup → all `(missing)`.
   Normalize with `str(id).strip().strip("[]").strip()`.
3. **REALIZED-ONLY** (`_REL_SYS`): announcements / plans / intent / "may·will·expects·poised·up-to" are
   NOT corroboration — only deployed/closed FACTS affirm. ("$2B announced" ≠ capital deployed.)
4. **ROLE-MATCH** (`_REL_SYS`): the actor must occupy the role the hypothesis is about. For a capital-
   *provider* hypothesis, a *borrower* posting collateral is the wrong side → tangential, not affirms.
5. **Hard-evidence guard** (`net_at`, `require_hard_evidence=True`): a source only counts on a side if it
   carries a **descriptive/reactive** (realized-fact) claim there; `predictive`/`interpretive` (forecast/
   opinion/intent) alone don't qualify it. Reports `hard_affirm_src` / `soft_affirm_src_dropped`.

## EISC / independence rules

- **Bitcoin is one CAPPED cluster** (`cluster_capped_low`, `CAP_VALUE`): within-cluster agreement can
  contribute at most ~0.25 of a voice — it can NOT masquerade as independent corroboration. Real
  corroboration of a bitcoin thesis must come from OUTSIDE the cluster (e.g. the `banks` cluster).
- **Cross-cluster earns the multiplier** (`xcluster_mult`, gated by `k_eff` = clusters contributing ≥0.5
  of a voice). One guest doing the rounds collapses to ~1; it does not earn the gold multiplier.
- **`own_network` quarantine is MATERIALITY-driven** (see AGENTS.md): live mode drops materially-tied
  Ten31 sources; test mode keeps them. Validated: own_network-only affirms → live `eisc≈0`, test > 0.
- **Seed edges in `sorted([a,b])` order** (matches `transcribe_worker`'s `sorted()`+`weight+=1` upsert) so
  auto-detected and seeded edges share a PK — a reversed-order row DOUBLE-COUNTS (math is frozenset-
  undirected but the table PK is ordered). κ: shared_guest 0.85, citation 0.45, community 0.60.

## The adversarial cases (the validation harness)

Pre-registered failed convictions used to test the engine against its target failure mode. Seeds:
`seeds/conviction_log.adversarial.seed.yaml`, `seeds/fanout.{STRIKE,BATTERY}2022.seed.yaml`,
`seeds/resolution.{STRIKE,BATTERY}2022.yaml`, `seeds/resolution_outcomes.adversarial.yaml`.

- **BATTERY2022 (timing/disconfirmation).** BTC-collateralized lending: demand rose, institutional
  *supply* failed. **PASS = demand-net rises while supply-net stays flat (≈0).** Run:
  `two-sided --conviction BATTERY2022 --nodes demand,supply --modes live`. Supply resolves ONLY on
  committed/DEPLOYED capital; policy/regulation is CONTEXT (the custody-policy node), never supply (S1).
- **STRIKE2022 (reflexivity/false-positive).** Lightning-retail-payments thesis FAILED. **PASS = net
  stays quiet in `live` (own_network dropped) while it would fire in `test`** — the engine refusing the
  intra-cluster echo. Run `two-sided --conviction STRIKE2022 --modes live,test`. The REALIZED-ONLY rule
  is load-bearing here (speculative "Lightning will revolutionize payments" is `predictive`, not signal).
  **Reading the output:** a single capped bitcoin cluster nets `eisc≈0.25` — already sub-bar vs
  `EISC_FLOOR=2.0`, so a `+0.25` "quiet in live" can be the *cluster cap* refusing the false positive,
  NOT the own_network drop. Check `own_net`: if it's 0, live==test and the reflexivity mechanism is
  unexercised (the affirmers are independent), so a quiet `live` does not by itself prove the echo-drop —
  you need own_network affirms present (`own_net>0`) for `test` to fire above `live`.

**Standing rule S1:** derivatives resolve on OUTCOME (scaled substance), never milestones or enablers.
An announced program / a regulatory unblock / a single bank's toe-in is CONTEXT, not corroboration.