alphacouncil-agent 0.6.0 → 0.8.0

package/.claude-plugin/marketplace.json

@@ -6,14 +6,14 @@
   },
   "metadata": {
     "description": "AlphaCouncil plugins for public-equity research.",
-    "version": "0.6.0"
+    "version": "0.8.0"
   },
   "plugins": [
     {
       "name": "alphacouncil-agent",
       "source": "./",
       "description": "Multi-agent equity research: evidence packets, bull/bear debate, portfolio-manager decision.",
-      "version": "0.6.0",
+      "version": "0.8.0",
       "author": {
         "name": "Zhao73"
       },

package/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "alphacouncil-agent",
-  "version": "0.6.0",
+  "version": "0.8.0",
   "description": "Multi-agent public-equity research workflow for Claude Code: spawns analyst workers, gathers sourced JSON evidence packets, runs bull/bear debate, and produces a portfolio-manager Buy/Overweight/Hold/Underweight/Sell decision with a full report.",
   "author": {
     "name": "Zhao73"

package/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "alphacouncil-agent",
-  "version": "0.6.0",
+  "version": "0.8.0",
   "description": "Host-visible or background public-equity research workflow with shared evidence packets.",
   "author": {
     "name": "Zhao73"

package/CHANGELOG.md

@@ -2,6 +2,95 @@
 
 Notable changes per release. Dates are UTC.
 
+## [0.8.0] — 2026-07-27
+
+Minor rather than patch: **the master bench changes from prompt voices to method models,
+and a defect is fixed that manufactured consensus in every prior run.**
+
+A user asked why the masters never appear in the report. The answer turned out to be three
+problems wearing one coat, and the third is the one that mattered.
+
+### Fixed
+
+- **An unrecognised master stance was silently rewritten as `cautious`.** The enum check
+  fell through to a default and the default was a real stance carrying real weight, so a
+  caller writing `avoid`, `long` or `neutral` got a seat that looked deliberate and voted.
+  A ten-seat run whose opinions ranged from avoid to long **recorded `cautious` ten times**
+  and rendered as a unanimity nobody produced. Worse in the other direction: a mistyped
+  `out_of_scope`, which carries zero weight, would have been promoted to a full vote.
+  Unmappable values now record as `out_of_scope` and warn; plausible synonyms are mapped;
+  the tool schema states the enum so a caller can see it before guessing.
+- **Master opinions were rendered nowhere.** `markdown.mjs` contained no reference to
+  `master`. Opinions were recorded, gated for completeness and weighted into the synthesis,
+  and a run could select ten lenses, pass every gate, and publish a report in which none of
+  them were readable.
+- **`record_*` echoed the entire run on every call.** Payloads grew with each recording and
+  a late call in a twenty-one-seat run passed 240k characters — context exhaustion arriving
+  exactly when a run was nearly finished. They now return progress and a path to
+  `status.json`.
+- **The Bluesky handle loader silently returned an empty list**, because `readFileSync` was
+  never imported and the `catch` swallowed the `ReferenceError`.
+
+### Added
+
+- **Persona v2: method models, not impressions.** A master is now an eligibility gate, a
+  scoring function with cited thresholds, an evidence slice and a narrative — in that order,
+  with the model confined to the last. `docs/persona-v2-spec.md`,
+  `schemas/persona-v2.schema.json`.
+- **Four pilot packs** — Buffett, Duan Yongping, Marks, Taleb — each with its own gate rather
+  than its own adjectives. On the real NOK facts the Buffett method declines (a 20-F filer
+  has no long-run series) while the Taleb method acts (an option chain exists). Different
+  methods see different companies, which twenty-one prompts over one brief never could.
+- **The admission bar is enforced in the loader.** A corpus below 25 propositions / 5 primary
+  sources / 5 decisions / 3 failures / 10 vetoes / 10 counterfactuals is downgraded to
+  `operator_lens` and carries its shortfall. All four pilots are honestly below it. A display
+  name reading as a person, a doctrine rule citing no grade A/B source, and a threshold
+  without provenance are each refused.
+- **Swap experiments.** Name swap: renaming a pack moves no verdict. Policy swap: Taleb's
+  policy under Buffett's name judges as Taleb. The differentiation diagnostic returns `none`
+  for four renamed copies of one method, which is what makes `effective` on the real four
+  worth anything.
+- **Memory with both clauses of the leak rule** — `public_at <= as_of AND
+  memory_created_at <= as_of`. The second clause stops a model reading its own diary. Undated
+  records are excluded rather than assumed harmless, and a postmortem cannot be written
+  before the horizon it judges.
+- **Enforced evidence slices.** The frozen fact pack is shared and unoverridable; what each
+  method may read on top of it is filtered in code, not requested in a prompt.
+
+### Changed
+
+- **The bench prints dissent first**, and a correlation note above every bench states that
+  the seats share a base model and an evidence brief, so their agreement is not independent
+  confirmation and the stance spread is not a vote count. Unanimity is reported as the
+  absence of independent dissent.
+- **`master_bench` is a required report section**, conditional on a bench having run, so
+  screen and quick modes are not failed for omitting one they never selected.
+- **Curated Bluesky accounts** are configurable by file or environment and ship **empty**:
+  unverified handles would be invented sources inside a tool that exists to refuse them. X
+  remains uncovered and the payload says so.
+
+### Wired into the live run
+
+`plan_visible_run` takes the deterministic pass before spawning anything. A seat whose
+method cannot reach the security is settled during planning and written straight into the
+run as an `out_of_scope` opinion — the completeness gate is satisfied and no agent is paid
+for a lens that had already declined. A seat that can look receives its settled verdict
+inside the prompt and is told to explain it rather than choose one; `record_master_opinion`
+then reconciles what came back, and a narrated stance contradicting the arithmetic does not
+win quietly — the deterministic verdict stands and the disagreement is kept as
+`narrated_stance`.
+
+Two distinctions the wiring forced. Absent grounding is not a screen that computed nothing:
+a run that was never measured falls back to v1 prompts instead of having its bench declined
+on missing data. And a declined seat cannot merely be skipped, because the completeness gate
+counts every selected master.
+
+### Still open
+
+The remaining seventeen masters, and `N_eff` — which needs an error-correlation matrix, which
+needs resolved ground truth that does not exist yet. Investment return is recorded as a
+long-run outcome and is never a gate.
+
 ## [0.6.0] — 2026-07-26
 
 Minor rather than patch: **every screened number changes.** Anyone comparing a result from

package/data/social-handles.json

@@ -0,0 +1,25 @@
+{
+  "_readme": [
+    "Curated Bluesky accounts read by get_social_pulse. Reading a named account is public;",
+    "search is not, so a curated list is the only free path to a professional layer -- and a",
+    "curated list was the better design anyway, because cashtag search is where the",
+    "promotional accounts live.",
+    "",
+    "This ships EMPTY on purpose. Shipping a list of handles that were never verified to",
+    "exist would put invented sources into a tool whose entire job is refusing invented",
+    "sources. Add accounts you have opened and checked yourself.",
+    "",
+    "Override the file with ALPHACOUNCIL_SOCIAL_HANDLES (comma-separated) or pass handles",
+    "directly to get_social_pulse.",
+    "",
+    "Honest scope: this does not restore X/Twitter coverage. X has no free discovery channel",
+    "-- Nitter search is dead, the X API bills per post, xAI bills per call -- so the",
+    "professional FinTwit layer stays uncovered unless you supply your own credential.",
+    "Reddit is not a substitute for it and must never be reported as one."
+  ],
+  "handles": [],
+  "notes": {
+    "format": "Bluesky handle without the leading @, e.g. someone.bsky.social",
+    "max_read": 12
+  }
+}

package/docs/persona-v2-spec.md

@@ -0,0 +1,278 @@
+# Persona v2 — Method Models, Not Impressions
+
+Specification for the 0.8.0 master rewrite. Background: `docs/roadmap.md`. Superseded
+sections of the earlier plan are noted in `docs/plans/0.8.0-master-models.md`.
+
+## Naming discipline
+
+These are **method models**, never people. Every user-facing surface says so.
+
+```
+Buffett Method Model          not "Warren Buffett"
+Duan Yongping Method Model    not "段永平"
+Marks Method Model            Taleb Method Model
+```
+
+A model whose source corpus does not clear the admission bar below is not named after a
+person at all. It is an **Operator Lens** — `Duan-inspired Operator Lens` — and the report
+says which of the two it is. Claiming to hold a living investor's judgment because a model
+can imitate the cadence is the specific dishonesty this document exists to prevent.
+
+## Why the previous design failed
+
+Twenty-one masters were twenty-one prompt voices over one base model and one evidence
+brief. Their errors correlate, so agreement between them is not confirmation. The published
+result is blunt: LLM errors run **>60% correlated**, majority voting inherits legitimacy
+from a theorem requiring independence, and with identical inputs debate is a **martingale** —
+expected correctness does not improve across rounds.
+
+The 0.7.0 work produced empirical support that nobody predicted. Recording ten opinions
+whose stances ranged from avoid to long, the run stored **`cautious` ten times**: the enum
+check fell through to a default, and the default was a real stance carrying real weight. A
+unanimity no master produced, manufactured by a normalizer.
+
+That is the load-bearing lesson for this spec: **a structured decision layer is only as
+honest as its rejection layer.** Silent coercion of an unrecognised value into a plausible
+one is the same failure as an imitation passing for a judgment — a shape that looks like a
+decision, holding no decision.
+
+## Six modules per model
+
+Capability is the first five. Voice is the sixth and touches nothing above it.
+
+| Module | File | Decides |
+|---|---|---|
+| Doctrine | `doctrine.jsonl` | what the method asserts, with sources |
+| Research policy | `research_policy.json` | what it goes and looks for, privately |
+| Decision policy | `decision_policy.json` | what it computes, vetoes, and refuses |
+| Tools | `tools.json` | which calculators it may call |
+| Memory | `memory_policy.json` | what it remembers and when it may read it |
+| Voice | `voice.<lang>.md` | how the finished verdict reads |
+
+**Voice may not participate in fact extraction, computation, or signal generation.** A
+system that lets style reach the decision will score "sounds like Omaha" as "judges like
+Omaha", which is the confusion the whole exercise is meant to end.
+
+## Doctrine entries are rules, not quotations
+
+```jsonc
+{
+  "rule_id": "buffett.circle_of_competence.01",
+  "claim": "Stop valuing a business whose model cannot be explained in a paragraph",
+  "scope": ["public_equity", "operating_business"],
+  "trigger": "business_model_explanation_failed",
+  "action": "out_of_scope",
+  "source_ids": ["buffett:S17", "buffett:S23"],
+  "confidence": "high",
+  "counterexamples": [],
+  "version": 1
+}
+```
+
+Every entry carries source, scope, trigger, action, counterexamples, confidence, version.
+An entry without `source_ids` cannot be loaded.
+
+### Source grades
+
+| Grade | What it is | May define |
+|---|---|---|
+| A | Signed letters, articles, formal talks, company filings | core decision rules |
+| B | Verifiable full interviews or meeting transcripts | core decision rules |
+| C | Direct quotation in reliable press | supporting rules only |
+| D | Third-party summary | staging area, never a rule |
+| E | Unsourced internet aphorism | rejected |
+
+Only **A/B** define a rule. The repository stores short summaries, method propositions,
+source URLs, dates, minimal necessary excerpts, and the boundary of applicability — not
+long copyrighted passages.
+
+### Admission bar for a named model
+
+- ≥25 sourced method propositions
+- ≥5 independent primary (A/B) sources
+- ≥5 real decision cases
+- ≥3 documented failures, misses, or changes of mind
+- ≥10 executable veto conditions
+- ≥10 counterfactual tests
+
+Errors and reversals matter more than aphorisms: they are what define the boundary of a
+method. Below the bar, the model is an Operator Lens and is labelled as one.
+
+## Evidence: shared facts, private research
+
+1. **Frozen fact pack** — price, filed figures, filing dates, share count, option chain.
+   Identical for every model and **not overridable**. Four models must not discover four
+   market caps in the name of independence.
+2. **Private research** — each model then runs its own query plan. Buffett-method looks for
+   moat, capital allocation and owner earnings; Duan-method for user value, operating
+   culture and opportunity cost; Marks-method for consensus, cycle position and risk
+   premium; Taleb-method for fragility, leverage, liquidity and convexity.
+   **Private material is invisible to the other seats until after first submission.**
+3. **Independent recomputation** — each model recomputes at least two load-bearing figures
+   from the frozen pack. It may disagree with a derived number; it may not silently
+   overwrite a filed one.
+
+This is the only lever that produces disagreement from something other than tone.
+
+## Anonymous first round
+
+First-round output carries no name and no style: evidence selected, computation performed,
+decision, confidence, abandonment condition, and where it is most likely wrong. Identity
+and voice are attached afterwards, before debate.
+
+## Structured decision, with a rejection layer
+
+```jsonc
+{
+  "stance": "constructive",
+  "action": "watch",
+  "confidence": 0.68,
+  "method_fit": 0.91,
+  "circle_of_competence": "inside",
+  "required_evidence_complete": false,
+  "vetoes_triggered": [],
+  "price_conditions": [],
+  "position_cap": 0,
+  "belief_updates": [],
+  "memory_ids_used": [],
+  "source_ids": []
+}
+```
+
+**Position size is never chosen by the model.** Deterministic code sets it from historical
+calibration, evidence quality, verification results, error correlation with other seats,
+portfolio limits, and triggered vetoes.
+
+Every enumerated field is validated on the way in. An unrecognised value is **rejected or
+recorded as a zero-weight abstention with a warning — never coerced into a neighbouring
+valid value.** This rule exists because its absence already produced a fake consensus once.
+
+## Memory, with a time boundary
+
+| Layer | Written when | Mutable by a run |
+|---|---|---|
+| Doctrine | source review + version release | no |
+| Episodic | every judgment | append only |
+| Belief | on belief change, pointing at episodes and sources | yes |
+| Postmortem | only after the prediction horizon expires | append only |
+| Working | during a run, archived after | no |
+
+**The leak rule.** In a run at `as_of = T`, a model may read only memories where
+`public_at <= T` **and** `memory_created_at <= T`. Without the second clause a model reads
+the future through its own diary, and every backtest built on it is fiction.
+
+## Proving it is not a performance
+
+Diagnostics that can return "this bench is decorative".
+
+| Experiment | Change | Expected if real |
+|---|---|---|
+| **Name swap** | names only | **no material change** |
+| **Policy swap** | load Taleb policy under the Buffett name | follows the policy, not the name |
+| **Evidence swap** | private packs only | explicable parts move, vetoes hold |
+| **Memory ablation** | drop episodic / postmortem / reflection / private retrieval | isolates which layer adds value |
+| **Voice removal** | neutral JSON, blind raters | method still identifiable |
+| **Model cross-over** | each persona on several models | separates persona from model |
+| **Counterfactual** | −40% price, 2× debt, margin drop, integrity event, convex→concave payoff | each model moves by its own rules |
+
+Name swap and policy swap are the cheapest and the most decisive. If a name change moves
+the verdict, the system is acting.
+
+`variance = company + persona + model + persona×model`. **If the model effect exceeds the
+persona effect, no personality has been built.**
+
+### N_eff is deferred, deliberately
+
+`N_eff ≈ (Σw)² / (wᵀ C w)` needs an error-correlation matrix `C`, which needs resolved
+ground truth on many cases. It is therefore **not a v1 deliverable** and must not be printed
+before the postmortem corpus exists. Until then the report discloses correlation
+qualitatively and refuses to publish a tally.
+
+## Pilot: four models, chosen to be hard
+
+| Model | Must independently develop | Private inputs |
+|---|---|---|
+| Buffett | circle of competence, moat, owner earnings, capital allocation | 10y financials, capex split, buyback/M&A record, pricing-power evidence |
+| Duan Yongping | business model, 本分 culture, user value, stop-list, opportunity cost, default inaction | product/user evidence, management behaviour timeline, culture events, candidate comparator |
+| Marks | cycle position, consensus, price-implied expectations, permanent loss | credit spreads, liquidity, valuation percentile, sentiment and positioning |
+| Taleb | fragility, tail risk, convexity, ruin, barbell | option chain, liability structure, liquidity stress, extreme scenarios, payoff shape |
+
+Buffett and Duan Yongping are chosen **because they are similar**. If the system can only
+separate those two by tone, Persona v2 has failed and expansion stops.
+
+## Acceptance gates for the pilot
+
+Recalibrate after the first ten cases.
+
+| Metric | Bar |
+|---|---|
+| Method propositions with sources | 100% |
+| Unsupported direct quotations | 0 |
+| Look-ahead leaks | 0 |
+| Structured-policy adherence | ≥95% |
+| Citations that actually support the claim | ≥95% |
+| Counterfactual direction correct | ≥85% |
+| Decision stability on repeated identical input | ≥80% |
+| Blind method identification after name+voice removal | >70% (chance = 25%) |
+| Error-correlation reduction vs. the 21 prompt masters | ≥15% |
+| Final report pass rate | ≥95% |
+| Runs still failing after auto-repair | <5% |
+
+If groups D/E do not beat B/C on factual accuracy, citation support, calibration, or unique
+information contribution, **stop and do not build the remaining seventeen.**
+
+### Control groups
+
+A: single strong model · B: current 8 analysts · C: 8 analysts + 21 prompt masters ·
+D: 8 analysts + 4 Persona v2 · E: D + verifiers · **F: human reference — optional.**
+
+F requires two independent raters plus an adjudicator and is likely unobtainable for a
+single-maintainer project. It is explicitly not allowed to block conclusions from A–E.
+
+## Returns are not a validation metric
+
+Investment return is a **lagging, low-power, confounded** indicator and cannot substitute
+for source accuracy, look-ahead control, method adherence, or risk calibration.
+
+- 36 historical plus 12 shadow cases cannot separate skill from luck, and correlated seats
+  reduce the effective sample further.
+- "Right for the wrong reason" is indistinguishable from skill by return alone.
+- Return is the metric look-ahead leakage inflates most and hides in best.
+- The product claim is *this is what the method would say*, not *this makes money*. A
+  method model that declines while the stock triples is not thereby wrong; a filter has a
+  known opportunity cost.
+
+Returns are recorded in the ledger as a long-run outcome. They are never a gate.
+
+## Phases
+
+| Phase | Output | Days |
+|---|---|---|
+| 0 | this spec, `schemas/persona-v2.schema.json`, case spec, source policy | 3–5 |
+| 1 | four source corpora | 5–7 |
+| 2 | `personas-v2/`, `policies/`, `master-tools/`, private research, anonymous round 1 | 7–10 |
+| 3 | `memory/` with time filtering and postmortem gating | 5–7 |
+| 4 | 36 frozen cases, 12 shadow cases, groups A–E, all swap experiments | 7–10 |
+| 5 | shadow running, no live trading | 8–12 wks |
+| 6 | expand only by missing method, not by fame | — |
+
+Benchmarks are frozen in Phase 0 so success criteria cannot be invented afterwards to
+flatter the result.
+
+## No fine-tuning in phase one
+
+Order is: sourced doctrine → independent retrieval → tools → decision policy → memory →
+evaluation → *then* consider tuning. Fine-tuning learns cadence fastest and judgment
+slowest, and twenty-one LoRAs over one base model still share their errors. If it ever
+happens, the method adapter and the voice adapter are separate artifacts and the method
+adapter faces the same ablations.
+
+## Unresolved
+
+- No open-source project has demonstrated faithful reproduction of a real investor's full
+  judgment. This spec does not claim to be first; it claims to be falsifiable.
+- The swap-experiment battery is this project's construction from standard ablation
+  practice, not an industry standard.
+- Letta has marked its older server generation legacy; pin a current version before
+  depending on its memory model.
+- `N_eff` and group F are both deferred, for different reasons, and neither may gate v1.

package/docs/personas.md

@@ -39,9 +39,9 @@ To add a role, add one file. It appears in the tool schema automatically.
 | `master_dalio` | master | yes | masters-adversarial, masters-core | deep | Dalio Lens | debt-cycle, machine-view, regime | original |
 | `master_duan_yongping` | master | yes | masters-value, masters-core | deep | Duan Yongping Lens | business-model, stop-doing-list, corporate-culture | ai-berkshire (MIT) |
 | `master_lynch` | master | yes | masters-value-classic, masters-core | deep | Peter Lynch Lens | category-first, peg, story-in-two-minutes | original |
+| `master_forensic_short` | master | yes | masters-adversarial, masters-core | deep | Forensic Short Seller Lens | forensic-accounting, crowding, borrow | original |
 | `master_li_lu` | master | yes | masters-value, masters-core | deep | Li Lu Lens | ten-year-certainty, civilization-trend, management-integrity | ai-berkshire (MIT) |
 | `master_marks` | master | yes | masters-value-classic, masters-core | deep | Howard Marks Lens | second-level-thinking, cycle-position, risk-is-permanent-loss | original |
-| `master_short_seller` | master | yes | masters-adversarial, masters-core | deep | Short Seller Lens | forensic-accounting, crowding, borrow | original |
 | `master_burry` | master | yes | masters-adversarial, masters-core | deep | Michael Burry Lens | primary-documents, contrarian, structural-short | original |
 | `master_klarman` | master | yes | masters-value-classic, masters-core | deep | Klarman Lens | absolute-return, cash-as-option, bottom-up | original |
 | `master_taleb` | master | yes | masters-options, masters-core | deep | Taleb Lens (Convexity and Tails) | tail-risk, convexity, options | original |

package/docs/plans/0.8.0-master-models.md

@@ -0,0 +1,155 @@
+# 0.8.0 — Masters Become Methods
+
+Complete implementation plan. Background and citations: `docs/roadmap.md`.
+
+## The defect, stated precisely
+
+Three separate problems that turn out to be one problem.
+
+**1. The bench is invisible.** `mcp/lib/markdown.mjs` contains zero occurrences of
+`master`. Grepping a completed run for a master's own words returns nothing outside
+`master_<id>.json`. Ten lenses ran, ten were gated for completeness, ten were weighted in
+synthesis, and the reader saw none of them.
+
+**2. The report is whatever the manager retypes.** `finalReportMarkdown()` is
+`return manager.report_markdown`. Eight evidence packets exist on disk and none of them
+reach the final report except by the manager writing them out again by hand. The quality
+gate checks that the string `market_data` appears somewhere — which the literal five
+characters satisfy.
+
+**3. The seats are not independent, and the report implies they are.** A tally of
+`6 avoid / 2 neutral / 2 long` invites arithmetic that the underlying samples do not
+support. Same base model, same evidence brief, same context. Agreement is the expected
+outcome, not a finding.
+
+These are one problem because the third is invisible while the first two hold. You cannot
+see that ten masters said the same thing in ten voices if you cannot see what they said.
+
+## Design
+
+### Persona frontmatter gains four fields
+
+The existing format (`personas/masters/<roster>/<id>.md`, JSON frontmatter, validated in
+`mcp/lib/personas/registry.mjs`) already carries `philosophy_tags`, `disqualifiers`,
+`default_weight`, `model_tier`. Four additions, all optional so unmigrated personas keep
+working:
+
+```jsonc
+{
+  "evidence_grade": "A",          // A|B|C — how much verifiable public methodology exists
+  "eligibility": {                 // deterministic gate, evaluated before any model call
+    "requires": ["screen.rules_computed >= 4", "filer.structured_financials"],
+    "on_fail": "out_of_scope"
+  },
+  "scoring": {                     // deterministic rules; every threshold cites a source
+    "max_score": 20,
+    "rules": [
+      {
+        "id": "roe_10y",
+        "metric": "screen.metrics.roe_10y",
+        "op": ">=", "value": 0.15, "points": 3,
+        "provenance": "1987 shareholder letter; 'above-average return on equity'"
+      }
+    ]
+  },
+  "evidence_slice": ["earnings_deep_dive", "insider_sec", "ib_event_analysis"]
+}
+```
+
+`evidence_grade` is the honesty valve. Grade A means the thresholds are quotable. Grade C
+means the public record is thin, the rule set is deliberately short, and the report says
+so instead of padding it to look uniform.
+
+### Four layers, model confined to the last
+
+| Layer | Where it runs | Can it set `stance`? |
+|---|---|---|
+| 1. Eligibility gate | `masters/eligibility.mjs`, pure JS | Yes — `out_of_scope`, no model call |
+| 2. Scoring function | `masters/scoring.mjs`, pure JS | Yes — stance derives from score bands |
+| 3. Evidence slice | `masters/slice.mjs`, pure JS | No — selects inputs |
+| 4. Narrative | subagent prompt | **No** — explains a fixed verdict |
+
+Layer 1 is the one that matters most for correctness. A master whose method cannot reach a
+security should say so and stop. In the NOK run every master received the same brief
+including "mechanical screen 0/7 computable" and every master wrote an essay anyway. A
+deterministic Buffett returns `out_of_scope` there, costs nothing, and is more informative
+than the essay.
+
+### Information asymmetry
+
+Layer 3 is the only lever that produces disagreement from a cause other than tone. Default
+slices:
+
+| Roster | Sees | Does not see |
+|---|---|---|
+| value, value-classic | multi-year financials, filings, governance | price action, options |
+| adversarial | price, positioning, macro, narrative | long-run ROE series |
+| quant | factor and price series | management commentary |
+| options | volatility surface, positioning | long-run financials |
+
+A master that never sees the price cannot anchor on the drawdown. That is the point.
+
+### Reporting changes
+
+- **The tally is removed.** Replaced by a deterministic score table: master, eligibility,
+  score/max, stance, evidence grade.
+- **Correlation is disclosed as a run property**, naming the shared base model and shared
+  evidence, and stating that agreement between seats is not independent confirmation.
+- **Minority report is promoted** to its own section above the concurring seats, with the
+  cause of divergence attributed (information slice vs. method vs. unexplained).
+
+### `council_diagnostics`
+
+A tool that can prove the bench decorative:
+
+- **self-consistency** — one master, repeated runs, same inputs
+- **pairwise agreement** — different masters, same inputs
+- **slice sensitivity** — one master, different evidence slices
+
+If pairwise agreement ≈ self-consistency, personas produced no differentiation. The verdict
+lands in `report_quality.json` as `master_differentiation: none|weak|effective` and is
+printed in the report. A run that cannot distinguish its own masters must not present them
+as a bench.
+
+## Work order
+
+Each step ends green on `npm run check`.
+
+1. **Visibility** (folded-in 0.7.0) — `renderMasterMarkdown`, `master_<id>.md`,
+   `all_agents.md` section, `artifact_index.md` entries.
+2. **Report contract** — `master_bench` section in `REPORT_SECTIONS`, required only when
+   `run.masters` is non-empty. Analyst work log generated from packets. Gate checks
+   substance and a scoped `<task>:S*` citation, not a bare task name.
+3. **Response slimming** — `record_*` returns a compact ack. Full state stays in
+   `status.json`.
+4. **Schema** — four frontmatter fields, registry validation, all existing personas still
+   load.
+5. **Deterministic core** — `eligibility.mjs`, `scoring.mjs`, `slice.mjs`, wired into
+   `plan_visible_run` and `record_master_opinion`.
+6. **Migrate masters by grade** — Buffett, Munger, Graham, Taleb, Simons, Thorp, Asness
+   first (grade A/B). Thin ones stay grade C with short rule sets.
+7. **Reporting** — score table, correlation disclosure, minority report.
+8. **Diagnostics** — `council_diagnostics` tool, `master_differentiation` in quality JSON.
+9. **Release** — CHANGELOG, version 0.8.0, npm publish, GitHub release.
+
+## Tests
+
+- master opinions render into all three artifacts
+- a run selecting a bench whose report omits `master_bench` fails the gate
+- an analyst work log with the task name and no substance fails the gate
+- eligibility gate returns `out_of_scope` without a model call when inputs are missing
+- scoring is pure: same inputs, same score, no network
+- evidence slices exclude what they claim to exclude
+- pairwise agreement ≈ self-consistency ⇒ `master_differentiation: none`
+- every persona with a `scoring` block has provenance on every threshold
+
+## What this does not fix
+
+Stated in the report, not just here.
+
+- **Shared base model.** Per-master vendor routing is supported but not required; without
+  it the seats remain correlated and the report says so.
+- **Thresholds are human choices.** Determinism relocates the uncertainty into constants.
+  Provenance is the mitigation; it is not a guarantee of correctness.
+- **Thin public methodology for some masters.** Grade C is a label, not a defect to be
+  papered over. Twelve real methods beat twenty-one accents.