alphacouncil-agent 0.6.0 → 0.9.0-solo-test.1

package/.claude/commands/alpha.md

@@ -13,7 +13,7 @@ Request: **$ARGUMENTS**
 |---|---|
 | *(empty)* | Print the mode table below and stop. Do not start a run. |
 | a ticker alone, or a question | **Full council.** Go to "Full council" below. |
-| ticker + `quick` | The `quick` preset: 4 analysts and the debate. No bench, no verification. |
+| ticker + `quick` | The `quick` evidence preset: 4 analysts + the selected master methods + debate. No verification. |
 | ticker + `screen` | `screen_ticker` only. No language-model judgment, no subagents. |
 | ticker + `options` | `get_options_chain` only. |
 | ticker + `news` | `get_news` on the symbol, and on its CIK when it is a US filer. |
@@ -22,8 +22,8 @@ Request: **$ARGUMENTS**
 When `$ARGUMENTS` is empty, print exactly this and stop:
 
 ```
-/alpha <TICKER>          full council — asks which preset first
-/alpha <TICKER> quick    4 analysts + debate, no bench, no verification
+/alpha <TICKER>          full council — shows every master, then asks which to run
+/alpha <TICKER> quick    4 analysts + selected masters + debate, no verification
 /alpha <TICKER> screen   mechanical filings screen only        (no model spend)
 /alpha <TICKER> options  IV term structure, skew, positioning  (no model spend)
 /alpha <TICKER> news     dated filings and headlines           (no model spend)
@@ -41,13 +41,40 @@ Say plainly that the four marked modes call keyless data tools and spawn no suba
 they cost nothing beyond this turn. The council modes spawn one subagent per seat, and that
 is where the spend is.
 
+## Council selection gate — mandatory for full and quick
+
+This gate happens before any research tool, run envelope, network fetch or subagent. It is
+the same in Claude Code, Codex, OpenCode and Grok Build.
+
+1. Call `begin_council_selection` with the symbol, original request, language and host. If
+   the request explicitly names masters, also pass those stable IDs as
+   `preselected_master_ids`; this highlights them but does not confirm them.
+2. Show **every returned master individually, in the returned order and with its stable
+   number**. Each row must include `identity`, `method`, `best_for` and `maturity`; a school
+   name or a count is not a substitute for the individual catalog.
+3. Ask for one submission. Accept one number from `1..N`, any comma/space-separated
+   combination, ranges such as `1-4` or `1..4`, stable IDs/names, or `all`. A host-native
+   multi-select is a convenience only. If it is unavailable or cannot show the full catalog,
+   use the numbered text table and plain reply on every host.
+4. If the original request already named masters or said `all`, prefill that choice but
+   **still show the full catalog and require this run's submission**. Do not silently reuse a
+   prior choice. The submitted choice is the confirmation; do not ask a second confirmation.
+5. Call `confirm_master_selection` with the returned `selection_id`, `catalog_hash`,
+   `display_ack: true`, and exactly one of `selected_master_ids`, `select_all: true`, or
+   `selection`. Retain the returned one-use `selection_receipt`.
+6. Only now call `plan_visible_run`, `collect_evidence` or `analyze_symbol`, passing that
+   `selection_receipt`. Do not also pass `masters` or `masters_roster`; the receipt is the
+   authoritative selection. If the receipt is missing, expired, stale or consumed, restart
+   from `begin_council_selection`.
+
+`screen`, `options`, `news` and `market` are data-only modes and skip this gate. `quick` is
+still a council judgment, so it never skips the gate; it changes the analyst fan-out, not the
+master-selection contract.
+
 ## Full council
 
-1. **Ask which masters, once, then run.** Call `list_council_options` and ask only about the
-   master bench — the six schools by name with their members, plus "all" and "none". Analysts
-   default to the eight-seat fan-out; do not ask about them. After the answer, start
-   immediately: no preset question, no confirmation. **Skip the question entirely** when the
-   request already said (a named school, "everything", "no masters", "be quick").
+1. Complete the mandatory council selection gate above. Analysts default to the eight-seat
+   fan-out; do not ask about them.
 2. Follow `skills/alphacouncil-agent/SKILL.md` from Stage 0. Do not improvise a shorter
    workflow: a report that looks finished and skipped the bench is worse than an obviously
    partial one.
@@ -61,9 +88,10 @@ is where the spend is.
 
 ## Quick mode owes the reader one sentence
 
-State in the output that no master lenses ran and nothing was cross-verified, so the
-confidence is lower than a standard run. Without it, a four-seat run reads like a full
-council result.
+Complete the same mandatory council selection gate, then use four analysts, the selected
+masters and the debate without verifier fan-out. State that evidence coverage and
+cross-verification are thinner than a standard run; list the selected masters so the quick
+result cannot be mistaken for the full bench.
 
 ## Screen mode
 

package/.claude-plugin/marketplace.json

@@ -6,14 +6,14 @@
   },
   "metadata": {
     "description": "AlphaCouncil plugins for public-equity research.",
-    "version": "0.6.0"
+    "version": "0.9.0-solo-test.1"
   },
   "plugins": [
     {
       "name": "alphacouncil-agent",
       "source": "./",
       "description": "Multi-agent equity research: evidence packets, bull/bear debate, portfolio-manager decision.",
-      "version": "0.6.0",
+      "version": "0.9.0-solo-test.1",
       "author": {
         "name": "Zhao73"
       },

package/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "alphacouncil-agent",
-  "version": "0.6.0",
+  "version": "0.9.0-solo-test.1",
   "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": "0.9.0-solo-test.1",
   "description": "Host-visible or background public-equity research workflow with shared evidence packets.",
   "author": {
     "name": "Zhao73"

package/.grok/commands/alpha.md

@@ -13,7 +13,7 @@ Request: **$ARGUMENTS**
 |---|---|
 | *(empty)* | Print the mode table below and stop. Do not start a run. |
 | a ticker alone, or a question | **Full council.** Go to "Full council" below. |
-| ticker + `quick` | The `quick` preset: 4 analysts and the debate. No bench, no verification. |
+| ticker + `quick` | The `quick` evidence preset: 4 analysts + the selected master methods + debate. No verification. |
 | ticker + `screen` | `screen_ticker` only. No language-model judgment, no subagents. |
 | ticker + `options` | `get_options_chain` only. |
 | ticker + `news` | `get_news` on the symbol, and on its CIK when it is a US filer. |
@@ -22,8 +22,8 @@ Request: **$ARGUMENTS**
 When `$ARGUMENTS` is empty, print exactly this and stop:
 
 ```
-/alpha <TICKER>          full council — asks which preset first
-/alpha <TICKER> quick    4 analysts + debate, no bench, no verification
+/alpha <TICKER>          full council — shows every master, then asks which to run
+/alpha <TICKER> quick    4 analysts + selected masters + debate, no verification
 /alpha <TICKER> screen   mechanical filings screen only        (no model spend)
 /alpha <TICKER> options  IV term structure, skew, positioning  (no model spend)
 /alpha <TICKER> news     dated filings and headlines           (no model spend)
@@ -41,13 +41,40 @@ Say plainly that the four marked modes call keyless data tools and spawn no suba
 they cost nothing beyond this turn. The council modes spawn one subagent per seat, and that
 is where the spend is.
 
+## Council selection gate — mandatory for full and quick
+
+This gate happens before any research tool, run envelope, network fetch or subagent. It is
+the same in Claude Code, Codex, OpenCode and Grok Build.
+
+1. Call `begin_council_selection` with the symbol, original request, language and host. If
+   the request explicitly names masters, also pass those stable IDs as
+   `preselected_master_ids`; this highlights them but does not confirm them.
+2. Show **every returned master individually, in the returned order and with its stable
+   number**. Each row must include `identity`, `method`, `best_for` and `maturity`; a school
+   name or a count is not a substitute for the individual catalog.
+3. Ask for one submission. Accept one number from `1..N`, any comma/space-separated
+   combination, ranges such as `1-4` or `1..4`, stable IDs/names, or `all`. A host-native
+   multi-select is a convenience only. If it is unavailable or cannot show the full catalog,
+   use the numbered text table and plain reply on every host.
+4. If the original request already named masters or said `all`, prefill that choice but
+   **still show the full catalog and require this run's submission**. Do not silently reuse a
+   prior choice. The submitted choice is the confirmation; do not ask a second confirmation.
+5. Call `confirm_master_selection` with the returned `selection_id`, `catalog_hash`,
+   `display_ack: true`, and exactly one of `selected_master_ids`, `select_all: true`, or
+   `selection`. Retain the returned one-use `selection_receipt`.
+6. Only now call `plan_visible_run`, `collect_evidence` or `analyze_symbol`, passing that
+   `selection_receipt`. Do not also pass `masters` or `masters_roster`; the receipt is the
+   authoritative selection. If the receipt is missing, expired, stale or consumed, restart
+   from `begin_council_selection`.
+
+`screen`, `options`, `news` and `market` are data-only modes and skip this gate. `quick` is
+still a council judgment, so it never skips the gate; it changes the analyst fan-out, not the
+master-selection contract.
+
 ## Full council
 
-1. **Ask which masters, once, then run.** Call `list_council_options` and ask only about the
-   master bench — the six schools by name with their members, plus "all" and "none". Analysts
-   default to the eight-seat fan-out; do not ask about them. After the answer, start
-   immediately: no preset question, no confirmation. **Skip the question entirely** when the
-   request already said (a named school, "everything", "no masters", "be quick").
+1. Complete the mandatory council selection gate above. Analysts default to the eight-seat
+   fan-out; do not ask about them.
 2. Follow `skills/alphacouncil-agent/SKILL.md` from Stage 0. Do not improvise a shorter
    workflow: a report that looks finished and skipped the bench is worse than an obviously
    partial one.
@@ -61,9 +88,10 @@ is where the spend is.
 
 ## Quick mode owes the reader one sentence
 
-State in the output that no master lenses ran and nothing was cross-verified, so the
-confidence is lower than a standard run. Without it, a four-seat run reads like a full
-council result.
+Complete the same mandatory council selection gate, then use four analysts, the selected
+masters and the debate without verifier fan-out. State that evidence coverage and
+cross-verification are thinner than a standard run; list the selected masters so the quick
+result cannot be mistaken for the full bench.
 
 ## Screen mode
 

package/.opencode/command/alpha.md

@@ -13,7 +13,7 @@ Request: **$ARGUMENTS**
 |---|---|
 | *(empty)* | Print the mode table below and stop. Do not start a run. |
 | a ticker alone, or a question | **Full council.** Go to "Full council" below. |
-| ticker + `quick` | The `quick` preset: 4 analysts and the debate. No bench, no verification. |
+| ticker + `quick` | The `quick` evidence preset: 4 analysts + the selected master methods + debate. No verification. |
 | ticker + `screen` | `screen_ticker` only. No language-model judgment, no subagents. |
 | ticker + `options` | `get_options_chain` only. |
 | ticker + `news` | `get_news` on the symbol, and on its CIK when it is a US filer. |
@@ -22,8 +22,8 @@ Request: **$ARGUMENTS**
 When `$ARGUMENTS` is empty, print exactly this and stop:
 
 ```
-/alpha <TICKER>          full council — asks which preset first
-/alpha <TICKER> quick    4 analysts + debate, no bench, no verification
+/alpha <TICKER>          full council — shows every master, then asks which to run
+/alpha <TICKER> quick    4 analysts + selected masters + debate, no verification
 /alpha <TICKER> screen   mechanical filings screen only        (no model spend)
 /alpha <TICKER> options  IV term structure, skew, positioning  (no model spend)
 /alpha <TICKER> news     dated filings and headlines           (no model spend)
@@ -41,13 +41,40 @@ Say plainly that the four marked modes call keyless data tools and spawn no suba
 they cost nothing beyond this turn. The council modes spawn one subagent per seat, and that
 is where the spend is.
 
+## Council selection gate — mandatory for full and quick
+
+This gate happens before any research tool, run envelope, network fetch or subagent. It is
+the same in Claude Code, Codex, OpenCode and Grok Build.
+
+1. Call `begin_council_selection` with the symbol, original request, language and host. If
+   the request explicitly names masters, also pass those stable IDs as
+   `preselected_master_ids`; this highlights them but does not confirm them.
+2. Show **every returned master individually, in the returned order and with its stable
+   number**. Each row must include `identity`, `method`, `best_for` and `maturity`; a school
+   name or a count is not a substitute for the individual catalog.
+3. Ask for one submission. Accept one number from `1..N`, any comma/space-separated
+   combination, ranges such as `1-4` or `1..4`, stable IDs/names, or `all`. A host-native
+   multi-select is a convenience only. If it is unavailable or cannot show the full catalog,
+   use the numbered text table and plain reply on every host.
+4. If the original request already named masters or said `all`, prefill that choice but
+   **still show the full catalog and require this run's submission**. Do not silently reuse a
+   prior choice. The submitted choice is the confirmation; do not ask a second confirmation.
+5. Call `confirm_master_selection` with the returned `selection_id`, `catalog_hash`,
+   `display_ack: true`, and exactly one of `selected_master_ids`, `select_all: true`, or
+   `selection`. Retain the returned one-use `selection_receipt`.
+6. Only now call `plan_visible_run`, `collect_evidence` or `analyze_symbol`, passing that
+   `selection_receipt`. Do not also pass `masters` or `masters_roster`; the receipt is the
+   authoritative selection. If the receipt is missing, expired, stale or consumed, restart
+   from `begin_council_selection`.
+
+`screen`, `options`, `news` and `market` are data-only modes and skip this gate. `quick` is
+still a council judgment, so it never skips the gate; it changes the analyst fan-out, not the
+master-selection contract.
+
 ## Full council
 
-1. **Ask which masters, once, then run.** Call `list_council_options` and ask only about the
-   master bench — the six schools by name with their members, plus "all" and "none". Analysts
-   default to the eight-seat fan-out; do not ask about them. After the answer, start
-   immediately: no preset question, no confirmation. **Skip the question entirely** when the
-   request already said (a named school, "everything", "no masters", "be quick").
+1. Complete the mandatory council selection gate above. Analysts default to the eight-seat
+   fan-out; do not ask about them.
 2. Follow `skills/alphacouncil-agent/SKILL.md` from Stage 0. Do not improvise a shorter
    workflow: a report that looks finished and skipped the bench is worse than an obviously
    partial one.
@@ -61,9 +88,10 @@ is where the spend is.
 
 ## Quick mode owes the reader one sentence
 
-State in the output that no master lenses ran and nothing was cross-verified, so the
-confidence is lower than a standard run. Without it, a four-seat run reads like a full
-council result.
+Complete the same mandatory council selection gate, then use four analysts, the selected
+masters and the debate without verifier fan-out. State that evidence coverage and
+cross-verification are thinner than a standard run; list the selected masters so the quick
+result cannot be mistaken for the full bench.
 
 ## Screen mode
 

package/AGENTS.md

@@ -13,6 +13,16 @@ Run `npm run check` after any code or prompt change.
 The MCP server is the load-bearing integration on every host: it reads `personas/` directly,
 so a host that ignores the generated agent files still gets correct prompts.
 
+Every host also follows the same mandatory master-selection protocol for a full or quick
+council. Call `begin_council_selection`, display every returned entry with its number,
+identity, method and `best_for`, collect a numbered/range/ID/`all` submission, then call
+`confirm_master_selection` with `display_ack: true`. Existing names in the request are only
+a prefill; the full catalog is still shown. Only the returned one-use `selection_receipt`
+may authorize `plan_visible_run`, `collect_evidence` or `analyze_symbol`. A host-native
+multi-select is optional UI sugar; the numbered text fallback is mandatory on Claude Code,
+Codex, OpenCode and Grok Build. Data-only `screen`, `options`, `news` and `market` modes are
+the only `/alpha` routes that skip this gate.
+
 | Host | Config | Agents | Skills |
 |---|---|---|---|
 | Claude Code | `.claude-plugin/plugin.json`, `.mcp.json` | `.claude/agents/alphacouncil-*.md` | `skills/` via the plugin manifest |
@@ -76,8 +86,8 @@ rather than four in a menu of a hundred.
 
 | Invocation | What runs | Model spend |
 |---|---|---|
-| `/alpha <ticker>` | Full council — asks which preset first | one subagent per seat |
-| `/alpha <ticker> quick` | 4 analysts + debate, no bench, no verification | 7 seats |
+| `/alpha <ticker>` | Shows every master, confirms `1..N`/ranges/`all`, then runs the full council | one subagent per selected seat |
+| `/alpha <ticker> quick` | Same mandatory master selection, then 4 analysts + selected masters + debate; no verification | varies with selection |
 | `/alpha <ticker> screen` | Mechanical filings screen only | **none** |
 | `/alpha <ticker> options` | IV term structure, skew, positioning | **none** |
 | `/alpha <ticker> news` | Dated filings and headlines | **none** |
@@ -85,8 +95,8 @@ rather than four in a menu of a hundred.
 | `/alpha` | Lists the modes and stops | **none** |
 
 The four marked **none** call keyless data tools and spawn no subagents, so they cost
-nothing beyond the turn you type them in. The council modes spawn one subagent per seat, and
-that is the entire cost of running this.
+nothing beyond the turn you type them in. Full and quick are council modes: both require a
+fresh selection receipt, and their cost varies with the selected master count.
 
 | Host | Where it reads them |
 |---|---|

package/CHANGELOG.md

@@ -2,6 +2,178 @@
 
 Notable changes per release. Dates are UTC.
 
+## [0.9.0-solo-test.1] — 2026-07-27
+
+This package is the first **0.9.0 solo-test prerelease**, not a formal production-GA release.
+`package.json`, the Claude/Codex plugin surfaces, marketplace metadata and
+`data/build-profile.v1.json` all declare `0.9.0-solo-test.1`.
+
+The public package uses npm's `next` dist-tag and the GitHub release is marked as a
+pre-release. Exact `0.9.0-solo-test.1` installation is supported while `latest` stays on the
+production-facing 0.8.0 line and the stable `0.9.0` version remains unoccupied until the
+formal GA gates pass.
+
+### Solo-test assurance boundary
+
+- Exactly 26 physical PersonaPack v3 packs are packaged under the isolated solo-test root.
+- Exactly 52 executable tools are present, all marked `provisional_derived_proxy`.
+- All 26 seats remain provisional `operator_lens`; operational seats: **0**;
+  `method_model` seats: **0**.
+- Human source approvals: **0**; human formula approvals: **0**; human approval signatures:
+  **0**. Project-derived proxies are test fixtures, not named-investor method attribution.
+- All 32 archived source candidates now have hash-bound machine pre-review artifacts. Three
+  isolated AI roles produce 96 deterministic triage outputs with explicit questions and
+  verdicts, while `human_reviewed=false` and `production_effect=none` keep the trusted
+  human-review quorum and production gate fail-closed.
+- The production loader rejects all 26 solo-test packs. Formal production assembly,
+  cutover, rollback evidence and GA remain fail-closed until the outstanding human-review,
+  experiment, host and release gates pass.
+- Source and installed-package validation are now separated: a source checkout verifies its
+  private/raw staging tree when present, while an installed package runs a real MCP selection
+  and receipt-replay smoke instead of reporting a misleading zero-test pass.
+
+Exact verification commands and the observed status are recorded in
+`docs/solo-test-0.9.0.md`.
+
+### Added in the first 0.9.0 increment
+
+- **A mandatory per-run master chooser.** `begin_council_selection` freezes the current
+  catalog and returns every seat with a stable number, identity, method, best-use case and
+  maturity. `confirm_master_selection` accepts one seat, any combination, ranges, stable IDs
+  or `all`, then issues a short-lived one-run receipt.
+- **Server-side enforcement rather than prompt etiquette.** `plan_visible_run`,
+  `collect_evidence` and `analyze_symbol` reject an absent, stale, replayed, cross-symbol or
+  cross-intent receipt before creating a run or fetching data. Confirm and consume operations
+  use exclusive file locks so separate host processes cannot spend the same selection twice.
+  Run starts also hold an exclusive lock: concurrent retries create one lifecycle, while a
+  completed same-receipt retry loads existing state instead of erasing packets or paying for
+  the council again.
+- **Text-complete four-host interaction.** Claude Code, Codex, OpenCode and Grok may enhance
+  the chooser with native multi-select, but all share the numbered text grammar and the same
+  MCP receipt. A named master is only preselected; the catalog is still displayed and the
+  user still submits this run's choice.
+- **Headless master execution.** The one-call analysis path now actually runs the selected
+  masters between evidence and debate. A failed selected seat stays missing and prevents a
+  false `complete` status.
+- **Selection audit fields.** `status.json` records the catalog/selection hashes, selected,
+  completed and pending masters, counts, per-seat state and the consumed selection ID.
+- **The five missing requested seats are now selectable.** Damodaran, Ackman, Cathie Wood,
+  Pabrai and Jhunjhunwala expand the active catalog from 21 to exactly 26. The complete
+  26-seat solo-test tree now loads only through the explicit provisional path as
+  `operator_lens`; none is mislabeled as operational or as a completed v3 method model.
+- **Cryptographic PersonaPack admission and release foundations.** Experiment results now use
+  Ed25519 attestations bound to the exact artifact/corpus/policy/tool/prompt hashes;
+  `method_model` promotion stays fused closed until explicitly enabled after all gates pass.
+  Source review and release operations each require two distinct trusted principals, so
+  repeated keys or copied reviewer names cannot manufacture independence.
+- **Immutable 26-pack publication and rollback.** The assembler accepts only one complete
+  operational-or-higher 26-seat tree, verifies signed source-adjudication ledgers, embeds and
+  hashes their evidence, then publishes by same-filesystem fsync and atomic rename. Signed
+  cutover/rollback approvals produce immutable versioned pointer history.
+- **A fail-closed GA verifier.** The real immutable release manifest is checked against the
+  production pack hashes. A separate signed evidence document must cover Claude Code,
+  Codex, OpenCode, Grok, the installed package, cutover and rollback. The previous unsigned
+  combined-JSON shape is rejected rather than accepted as self-certification.
+
+### In progress, not yet a production-GA claim
+
+- Migrate all 26 seats to PersonaPack v3 with sourced doctrine, private research plans,
+  typed computations, native decisions, hard vetoes, calibration and ablation tests.
+- Pass the A-E comparison ladder: single-agent baseline; 8 analysts; 8 analysts plus selected
+  masters; verifier-enabled council; and blinded human reference answers. Seat count is never
+  treated as independent sample count.
+
+Implementation tracker: `docs/plans/0.9.0-personapack-v3.md`.
+
+## [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/CLAUDE.md

@@ -6,15 +6,25 @@ You are working on `alphacouncil-agent`, a Codex plugin for public-equity resear
 
 The plugin turns `@alphacouncil-agent <ticker/request>` into a full equity-research workflow:
 
-1. Spawn or plan specialized analyst workers.
-2. Produce JSON evidence packets with sourced claims.
-3. Merge packets into a shared evidence base.
-4. Run bull and bear researchers.
-5. Run a portfolio manager to decide Buy / Overweight / Hold / Underweight / Sell.
-6. Write a complete final report directly in chat and to `~/.alphacouncil-agent/runs/<run_id>/final_report.md`.
+1. Display every available master method and confirm a one-run selection receipt.
+2. Spawn or plan specialized analyst workers.
+3. Produce JSON evidence packets with sourced claims.
+4. Merge packets into a shared evidence base and run every selected master.
+5. Run bull and bear researchers.
+6. Run a portfolio manager to decide Buy / Overweight / Hold / Underweight / Sell.
+7. Write a complete final report directly in chat and to `~/.alphacouncil-agent/runs/<run_id>/final_report.md`.
 
 Default behavior is full workflow. Do not downgrade to lite, smoke-test, or debug output unless the user explicitly asks for it.
 
+## Current Build Profile
+
+The declared package/plugin version is `0.9.0-solo-test.1`, channel `solo_test`. The isolated packaged
+tree contains 26 physical v3 packs and 52 executable `provisional_derived_proxy` tools.
+Every seat is still a provisional `operator_lens`: operational = 0, `method_model` = 0,
+human source approvals = 0, human formula approvals = 0, human approval signatures = 0.
+The production loader rejects these packs and formal production GA remains fail-closed.
+See `docs/solo-test-0.9.0.md`.
+
 ## Agent-Skills Governance
 
 Also follow `skills/agent-skills-governance/SKILL.md`. It is bundled with this plugin and applies the `addyosmani/agent-skills` style of explicit gates, anti-rationalizations, and exit criteria to AlphaCouncil runs. Do not treat it as an optional local setup step.
@@ -25,6 +35,16 @@ Also follow `skills/agent-skills-governance/SKILL.md`. It is bundled with this p
 `quick`, `screen`, `options`, `news` and `market` select narrower modes, and the last four
 spawn no subagents at all.
 
+Every full or quick council run must first call `begin_council_selection`, display each
+returned master with its stable number, identity, method and `best_for`, and obtain a
+submission. A native multi-select is optional; the common fallback accepts numbered choices,
+ranges and `all` in plain text on every host. Even if the request already names masters,
+show the complete catalog and treat those names only as a prefill. Call
+`confirm_master_selection(selection_id, catalog_hash, display_ack: true, ...)`, then pass its
+one-use `selection_receipt` to `plan_visible_run`, `collect_evidence` or `analyze_symbol`.
+No research, run directory or worker may start before that receipt exists. Data-only
+`screen`, `options`, `news` and `market` modes skip the selection gate.
+
 ## Analyst Roles
 
 Default evidence roles (the eight-seat default fan-out):
@@ -44,9 +64,14 @@ Optional analysts, on the `full` roster rather than the default:
 - `market_narrative`
 - `social_pulse`
 
-Master bench: 21 lenses across six rosters, run between evidence and debate on **every**
-host. Verifiers: `source_fidelity`, `rederivation`, `refuter`. See `skills/alphacouncil-agent/SKILL.md`
-for the workflow and `docs/personas.md` for the full roster.
+Master bench: the complete generated catalog (26 selectable lenses for the 0.9.0 roster), selectable
+one by one or with `all`, and run between evidence and debate on **every** host. Verifiers:
+`source_fidelity`, `rederivation`, `refuter`. See `skills/alphacouncil-agent/SKILL.md` for the
+workflow and `docs/personas.md` for the generated roster.
+
+Do not call the 26 solo-test packs approved method models. The explicit solo-test loader
+exposes 26 provisional `operator_lens` packs and 52 executable derived-proxy tools; none is
+operational or `method_model`, and none has human source/formula approval or signature.
 
 Debate roles:
 

package/PRODUCT.md

@@ -0,0 +1,37 @@
+# Product
+
+## Register
+
+product
+
+## Users
+
+AlphaCouncil serves investors, analysts, researchers, and technically capable users who run public-equity research from Claude Code, Codex, OpenCode, or Grok Build. They may be making a quick research pass or assembling a deeper investment-committee record, and they need to understand the time, cost, evidence coverage, and method lenses they are choosing before work begins.
+
+## Product Purpose
+
+AlphaCouncil turns a listed-equity question into an auditable research workflow: specialist evidence collection, selected investor-method lenses, verification, bull/bear debate, and a portfolio-manager decision with durable artifacts. Success means the user controls which method seats run, every selected seat is actually completed, material claims remain traceable to sources, and every supported host executes the same workflow contract.
+
+## Brand Personality
+
+Precise, rigorous, transparent. The product should communicate expert confidence without theatrical certainty. It should make cost, missing evidence, method limits, and incomplete work visible instead of hiding them behind polished prose.
+
+## Anti-references
+
+- A one-click AI stock picker that starts expensive work without showing what will run.
+- A casino-like voting screen that treats correlated persona outputs as independent votes.
+- Celebrity role-play whose only differentiation is a name, portrait, quote, or speaking style.
+- A dense configuration wizard that asks users to decide technical details they do not need to understand.
+- Host-specific workflows whose controls or completion rules drift between Claude Code, Codex, OpenCode, and Grok Build.
+
+## Design Principles
+
+1. User choice before spend: every council run requires an explicit, one-run master selection.
+2. Explain the method, not the celebrity: each selectable master states what it examines, where it abstains, and its maturity.
+3. Progressive disclosure: show a compact grouped roster first, preserve access to all individual seats, and make select-all explicit.
+4. Native when available, text-complete everywhere: use a host's multi-select control when it exists, with an equivalent numbered or ID-based fallback.
+5. Selection is a contract: persist the exact selected IDs and catalog version, and block execution until the selection is confirmed.
+
+## Accessibility & Inclusion
+
+The complete selection workflow must work by keyboard and in plain text without relying on color, pointer input, hover, images, or a particular host widget. Labels and descriptions must support the run language, preserve stable persona IDs, expose selected count and validation errors, and remain usable with screen readers and narrow terminal widths. Motion is not required. Host-native controls must retain an equivalent textual fallback.