Publish the frozen expertise pre-registration and its hash

CONTRACTS.md

@@ -270,3 +270,102 @@ outside the band are removed from the RL mix and **kept in eval**.
 - Not a claim that the trained policy transfers to live traffic. It is trained
   on a fitted simulator and evaluated on held-out real calls; the gap is
   measured, published, and named as the central limitation.
+
+---
+
+# ADDENDUM — EXPERTISE (contracts/v1-expertise)
+
+Added 2026-09-03 per `docs/DECISIONS.md` D030. Everything above stays in force
+for the EDGAR population; this section governs the scholarly one. Frozen on the
+same terms: no change without a DECISIONS entry **and** measured evidence.
+
+## A1. Attestation tiers, restated for expertise
+
+| Tier | Meaning | Source |
+|---|---|---|
+| **A** | Authorship attested by a third party with a persistent identifier | OpenAlex work ↔ author, backed by a Crossref DOI |
+| **B** | Self-asserted to a registry | ORCID affiliations and self-added works |
+
+Note this **reverses D003's demotion of ORCID**, and only partly. ORCID is the
+identity spine - a stable, self-owned, CC0 identifier - while the *attestation*
+comes from the publication record, which the researcher does not control. A
+person cannot self-assert a DOI into existence.
+
+## A2. The attested expertise tuple
+
+```
+ExpertiseTuple = {
+  author_id:    str    # OpenAlex author identifier (disambiguated)
+  orcid:        str    # ORCID iD, when present
+  display_name: str
+  topic_id:     str    # OpenAlex topic
+  topic_label:  str
+  n_works:      int    # attested works by this author in this topic
+  first_year:   int
+  last_year:    int
+  institution:  str    # last attested affiliation, ROR-identified
+}
+```
+
+Published as pointers - OpenAlex IDs, ORCID iDs, topic IDs, DOIs and counts.
+Never as an assembled dossier. `docs/ETHICS.md` governs.
+
+## A3. Claim polarity, and how a false claim is CONSTRUCTED
+
+The reason this population is stronger than the employment one: **a false claim
+is constructible from the oracle rather than judged.**
+
+- **ATTESTED**: the author has `n_works >= MIN_WORKS` in the topic.
+- **CONSTRUCTED_FALSE**: the author has **zero** works in the topic, and the
+  topic is drawn from an **adjacent** field by the OpenAlex topic hierarchy -
+  a sibling under a shared ancestor, never hand-picked and never absurd.
+
+An implausible negative ("this cardiologist is not an expert in medieval
+poetry") measures nothing. Adjacency is mechanical: same domain, different
+field, sibling topic. The selection rule is code, versioned as
+`topic_adjacency/v1`, and no model participates.
+
+Tasks are balanced 50/50 by polarity. The balance is by construction, not by
+sampling luck.
+
+## A4. `MIN_WORKS` and the contamination bound
+
+`MIN_WORKS = 3`. One paper in a topic is a dabble and would make the ATTESTED
+side unfairly hard; a high threshold would only measure famous people.
+
+OpenAlex author disambiguation is imperfect and **its error rate is published by
+OpenAlex itself**. That rate is the contamination bound here, exactly as the
+same-human-two-CIK rate was for EDGAR (section 4.2), and expertise rates are
+reported raw and net of it. An author record that merges two humans manufactures
+both false ATTESTED and false CONSTRUCTED_FALSE claims, so the bound cuts both
+ways and is reported on both.
+
+## A5. The `expert_sourcing` cost profile
+
+Severity is a property of the error; cost is a property of the caller. For an
+expert-sourcing platform the asymmetry is extreme and is grounded in observed
+market facts rather than intuition:
+
+- A **miss** costs roughly one sourcing bounty. Observed range for a verified
+  expert hire: **$250-$15,000**.
+- A **fake expert admitted** - a `FALSE_MERGE` here - has cost this market a
+  security incident involving suspected state-actor infiltration, a breach with
+  class actions, and contaminated training data delivered to frontier labs.
+
+|  | `MISS` | `FALSE_MERGE` | `UNSURE_WRONG` | `ABSTAIN` | `STALE` |
+|---|---|---|---|---|---|
+| `expert_sourcing` | 1 | **150** | 30 | 0.5 | 2 |
+
+150 is deliberately the largest ratio in the registry, above `journalism`'s 100.
+It is a stated assumption, not a measurement, and - like every profile - the
+result that matters is **ranking stability across profiles**, not the number
+chosen here.
+
+## A6. What this addendum does not claim
+
+- Not a claim that any named person is or is not an expert. The instrument
+  measures a **provider's** verdict against an attestation record, never a
+  person's competence. See `docs/ETHICS.md` A-E1.
+- Not a claim that publication record equals expertise. It is what can be
+  attested. A brilliant engineer who never publishes is invisible here, and that
+  is a coverage gap, published as one.

CONTRACTS.sha256

@@ -1 +1 @@
-6e4488c0fde3fad1bc31f7819e87ff1714ccc6e12d39f9b41cca2daf0cacb423
+5960979dcc884ca4c6869c371bf499b0953b7ee8a378a28e76d241b16d264305

PRE-REGISTRATION-EXPERTISE.md

@@ -0,0 +1,182 @@
+# PRE-REGISTRATION — EXPERTISE
+
+Frozen 2026-09-03, before `src/titer/corpus/scholar.py` exists and before a
+single credit is spent on this population. A hash is published to the public
+Hugging Face repository so the ordering is provable by someone other than its
+author.
+
+**This text is never edited.** It supersedes nothing: `docs/PRE-REGISTRATION.md`
+stays exactly as frozen, and its H1-H4 remain live for the EDGAR population -
+unfinished, not falsified. See `docs/DECISIONS.md` D030.
+
+**Written knowing how the first one went.** The employment pre-registration was
+honest and still produced two designs that could not answer their own questions:
+H2 sampled randomly when colliding names are 15% of the population (D027), and
+H1 sampled randomly when 272 of 299 observations then landed in one elapsed-time
+bin. Both were caught by looking at where observations fell, not by reading a
+summary. The sampling rules below are stricter because of it.
+
+---
+
+## 0. Fixed before data
+
+- The hypotheses, estimators, falsification conditions and sampling design here.
+- The attestation tiers and the topic taxonomy added to `CONTRACTS.md`.
+- The `expert_sourcing` cost profile and its ratios.
+- The rule that **every stratified result is reported per stratum and never
+  pooled into a marginal rate**.
+- The power rule in section 3, unchanged from the original: it fixes what may be
+  claimed at each achievable `n` before any `n` is known.
+
+Not fixed: the achievable `n`. It depends on measured cost per task.
+
+---
+
+## 1. Design
+
+**The negative control is constructible, and that is why this population is
+better than the last one.** A person's attested topics come from their authored
+works. A *false* expertise claim is therefore not a judgement call - it is a
+topic in which they have provably never published. Every task is a balanced
+discrimination between an attested claim and a constructed one, drawn from the
+same oracle. The employment corpus could never do this: there was no way to
+manufacture a verifiable non-employment.
+
+**Paired.** Every arm receives the identical task set in the identical order,
+each rendered in the shape its own API documents (D025). Differences are
+within-task.
+
+**Stratified by construction, not by hope.** Two axes, both fixed here:
+
+1. **Claim polarity** - attested vs constructed-false, balanced 50/50.
+2. **Name-collision degree**, using the provider-facing normalization (D027),
+   with equal allocation across `unique / low / medium / high`.
+
+Simple random sampling is **prohibited** for E1 and E2. It is what broke H1 and
+H2, and the failure mode - a bin holding 90% of the observations - is invisible
+in a summary statistic.
+
+**One measurement window**, recorded to the day.
+
+---
+
+## 2. Hypotheses
+
+### E1 - Providers affirm expertise that is not attested
+
+**Claim.** Asked whether a named person has expertise in a topic, providers
+affirm constructed-false claims at a non-zero rate, and that rate rises with
+name-collision degree.
+
+**Measurement.** Balanced attested / constructed-false tasks. Outcome classes
+reuse `CONTRACTS.md` section 4; a confident affirmation of a constructed-false
+claim is a `FALSE_MERGE` - **a fake expert admitted**.
+
+**Estimator.** Rate per (polarity x collision band) cell with Wilson intervals;
+paired bootstrap for provider differences, 10,000 resamples, seed 11. Expected
+loss under all reportable profiles including `expert_sourcing`.
+
+**Falsified if.** The false-affirmation rate is indistinguishable from zero at
+the achieved `n`, or flat in collision degree. Either is published as
+prominently as a positive.
+
+**Pre-committed nuisance.** A constructed-false topic must be *plausible* -
+drawn from an adjacent field, not an absurd one - or the task measures nothing.
+Adjacency is defined by the OpenAlex topic hierarchy, mechanically, never by
+hand-picking.
+
+### E2 - A capability constraint does not disambiguate
+
+**Claim.** Given a colliding name plus a capability constraint ("the Jane Chen
+who published on X"), providers do not resolve to the right person more often
+than with the name alone.
+
+**Measurement.** Colliding-name tasks only, four arms: name alone; name + a
+current-state filter the API accepts; name + capability in free text; name +
+capability oracle-supplied (ceiling). This is the D029 experiment.
+
+**Falsified if.** The free-text capability arm beats the name-alone arm - the
+semantic query field already absorbs biographical context, and the architectural
+gap D029 asserts does not exist in practice.
+
+### E3 - The free signals are informative but uncalibrated
+
+**Claim.** Provider-returned confidence on an expertise claim carries
+information about correctness and is not calibrated as shipped.
+
+**Measurement.** Reliability diagram, 10 equal-width bins, **empty bins reported
+as empty, never merged**. ECE and Brier before and after isotonic recalibration
+fitted on a disjoint split. Out-of-range confidences are an error, not silently
+dropped.
+
+**Falsified if.** Raw ECE is already low, or recalibration cannot improve Brier
+beyond its interval.
+
+### E4 - Coverage differs across populations
+
+**Claim.** The same provider resolves the academic population and the SEC
+executive population at materially different rates.
+
+**Measurement.** Identical instruments over both corpora, reported side by side,
+never pooled.
+
+**Falsified if.** The paired difference interval contains zero.
+
+**Why it is worth reporting either way.** A difference is a coverage finding
+about who an index knows. No difference is evidence that resolution quality is a
+property of the method rather than the population.
+
+---
+
+## 3. Power, and what may be claimed at each n
+
+Unchanged from the employment pre-registration, because the rule is what matters
+and it must not be renegotiated per study.
+
+| Achieved 95% half-width | What may be said |
+|---|---|
+| <= 0.05 | comparisons may be reported as rankings |
+| 0.05 - 0.10 | intervals only; **no ranking language anywhere**, including the README and any abstract |
+| > 0.10 | state plainly that this budget cannot separate the arms, publish the intervals, and publish the `n` each hypothesis would have required |
+
+`scripts/power.py` selects the branch from a **measured** cost per task, before
+the measurement runs. That branch fired as `cannot_separate` on the employment
+study and was honoured; it is honoured here too.
+
+---
+
+## 4. Stopping and multiplicity
+
+- **No optional stopping.** The task set is drawn once, per stratum, by seed.
+- **No arm is dropped for performing badly** - only for access or Terms reasons,
+  recorded with the quoted clause.
+- **Primary claims: E1 under `expert_sourcing`, and E2.** Everything else is
+  secondary and labelled so.
+- **Ploid is not measured** and no claim about it is made from this study.
+  D029 remains a hypothesis (`docs/DECISIONS.md`).
+
+---
+
+## 5. Published regardless of outcome
+
+1. Every exclusion rate, including the OpenAlex author-disambiguation error rate
+   used as the contamination bound.
+2. The achieved `n` per stratum and which section 3 branch fired.
+3. The full per-stratum table. **No pooled marginal rate is ever published for a
+   stratified result.**
+4. Any falsified hypothesis.
+5. The negative-control construction, in full, so a reader can attack it.
+6. Whether the web floor is a valid floor this time - the last one was a parser
+   artefact and was withdrawn (`docs/BENCHMARK.md`).
+
+---
+
+## 6. Reversal clause
+
+Superseded only by a dated `docs/DECISIONS.md` entry stating what changed, why,
+what evidence forced it, what was rejected, and **what would have been reported
+under this plan**. A supersession that cannot state the counterfactual is not
+permitted.
+
+If provider Terms prohibit named comparative publication, the naming changes and
+nothing else does.

PRE-REGISTRATION-EXPERTISE.sha256

@@ -0,0 +1 @@
+94342e4c8eaf78aa8db5b91eb1928d57b5c87ea8a13b42e3edc8807b7c129053

README.md

@@ -15,10 +15,34 @@ pretty_name: "titer — EDGAR officer corpus (pre-registration placeholder)"
 
 > **THERE IS NO DATA IN THIS REPOSITORY YET.**
 >
-> This repository currently contains only a **frozen pre-registration and its
-> hash**, published so that the ordering — analysis plan fixed *before* any
-> measurement — is provable by someone other than its author. The corpus
-> described below is scheduled for a later release.
+> This repository contains **frozen pre-registrations and their hashes**,
+> published so that the ordering — analysis plan fixed *before* any measurement
+> — is provable by someone other than its author. Corpora are scheduled for
+> later releases.
+>
+> There are now **two** studies. The second does not replace the first: the
+> original pre-registration is byte-identical to its first publication and its
+> hypotheses remain live and unfinished.
+
+## The two studies
+
+| | Population | Question | Frozen |
+|---|---|---|---|
+| **Employment** | SEC Section 16 officers and directors | How often does a people index resolve the wrong human, and how stale is it? | 2026-09-02 |
+| **Expertise** | Scholarly authors (OpenAlex / ORCID / Crossref) | How often does a provider affirm expertise that is not attested? | 2026-09-03 |
+
+The second study exists because of what the first one found about itself. Two of
+its designs could not answer their own questions — random sampling put 272 of
+299 observations into a single elapsed-time bin, and 251 of 299 into a single
+name-collision band. Both were caught by looking at where observations landed
+rather than at a summary statistic. **Simple random sampling is prohibited in
+the expertise pre-registration as a direct result.**
+
+The expertise population is also stronger in one specific way: a **false claim is
+constructible**. A person's attested topics come from their authored works, so a
+false expertise claim is a topic with zero attested works, drawn from an adjacent
+field by a mechanical rule. The employment corpus could never manufacture a
+verifiable non-employment.
 
 ## Why a pre-registration is published before the data
 
@@ -45,8 +69,10 @@ repository are locally forgeable, so they are not evidence. These hashes are.
 
 | File | What |
 |---|---|
-| `PRE-REGISTRATION.md` | The frozen analysis plan. Never edited; superseded only by dated decision entries carrying a reversal clause. |
+| `PRE-REGISTRATION.md` | The frozen employment analysis plan. Never edited; superseded only by dated decision entries carrying a reversal clause. |
 | `PRE-REGISTRATION.sha256` | Its hash, as locked in the working repository. |
+| `PRE-REGISTRATION-EXPERTISE.md` | The frozen expertise analysis plan. |
+| `PRE-REGISTRATION-EXPERTISE.sha256` | Its hash. |
 | `CONTRACTS.md` | The frozen definitions the pre-registration depends on: attestation tiers, the attested tuple, closed taxonomies, outcome classes, cost profiles, reward, split rules. |
 | `CONTRACTS.sha256` | Its hash. |