feat: separate Task, Read, and Effect audit stages

.gitignore

@@ -12,3 +12,8 @@ __pycache__/
 /candidates.jsonl
 /chats.jsonl
 /judgments.jsonl
+/task-source.jsonl
+/task-inventory-draft.jsonl
+/task-inventory.jsonl
+/read-attributions.jsonl
+/effect-judgments.jsonl

README.md

@@ -1,22 +1,22 @@
 # Context Tree Value Audit
 
-`context-tree-value-audit` 0.4.0 is an explicit-only Skill for task-first,
+`context-tree-value-audit` 0.5.0 is an explicit-only Skill for task-first,
 evidence-first analysis of Context Tree decision value for the current First
 Tree Runtime when its native historical evidence is supported. It reconstructs
 single-Agent-owned continuous Tasks from authorized Chats, records whether a
-Tree Read is observed or unresolved, and judges one optional Effect without a
-minimum sample gate.
+Tree Read is observed or unresolved, and judges zero or more independent
+Effects without a minimum sample gate.
 
 The 0.2 series renamed the installable Skill from
 `context-tree-insights` to `context-tree-value-audit`. Replace the old Skill
 directory during upgrade; do not install both names because they represent one
 explicit audit capability, not two independent workflows.
 
-Version 0.4 replaces the 0.3 Task judgment schema v2 with continuous episode
-schema v3. Do not reuse schema-v1 or schema-v2 `task-judgments.jsonl`:
-reconstruct the episodes and their ownership/objective/outcome anchors, then
-rerun the deterministic reporter. No migration or compatibility reader is
-provided.
+Version 0.5 replaces the combined schema-v3 Task judgment with three ordered
+schema-v4 artifacts: frozen Task inventory, Read attributions, and Effect
+judgments. Do not reuse schema-v1/v2/v3 judgment files. Reconstruct Tasks from
+the message-only projection, freeze the inventory digest, then rerun Read and
+Effect analysis. No migration or compatibility reader is provided.
 
 The audit core remains separate from First Tree core. Codex, Claude Code, and
 Claude Code TUI use their existing native local transcripts. Cursor and Kimi
@@ -40,9 +40,14 @@ Tree.
   counting unit.
 - Short continuations, status prompts, context-dependent questions, repeated
   review/fix requests, and phases of one delivery do not become separate Tasks.
-- Clear Tasks carry source-backed ownership, objective, and outcome anchors;
-  work owned by another Agent remains context until the audited Agent visibly
-  receives, accepts, or takes over an objective.
+- Clear Tasks carry source-backed objective and same-Agent outcome messages;
+  work owned by another Agent remains context until the audited Agent receives
+  or visibly accepts an objective. That acceptance is an analyst judgment from
+  work messages; the reporter does not infer the addressee of arbitrary prose.
+- Task reconstruction uses a message-only projection that excludes
+  collector-derived Reads, passages, Tree-mention indexes, receipts, choice
+  projections, and Effect judgments; original work-message content remains
+  unchanged. Its normalized inventory is frozen before Read attribution.
 - Local Runtime evidence is preflighted against authorized Chat and Agent IDs
   before complete recorded output is scanned.
 - Missing, cleaned, ambiguous, malformed, truncated, or unsupported traces are
@@ -59,8 +64,8 @@ Tree.
   duplicate, pending, or out-of-window results stay unresolved.
 - Read is only `observed` or `unresolved`; unresolved is never counted as
   unused.
-- Effect is optional and only `confirmed`, `constrained`, `redirected`, or
-  `conflicted`.
+- Each Task has zero or more Effects; each is only `confirmed`, `constrained`,
+  `redirected`, or `conflicted`.
 - A decision receipt may support an Effect but cannot create one by itself.
 - There is no fixed Task quota, task-type gate, or saturation state.
 - The output is a sampled evidence report, not causal proof, ROI, or a
@@ -184,23 +189,26 @@ tasks. Pin a reviewed commit or release when installing for another Agent.
 
 ## Pipeline
 
-The Skill orchestrates four stages:
+The Skill orchestrates six operational steps across three analysis stages:
 
 1. `export-chats` resolves explicit authorization and exports visible records.
 2. `collect` maps authorized Chats to supported native local evidence,
    classifies every in-window Tree-read attempt into a conserving four-state
    grammar, reconstructs exact or read-only-composite evidence plus visible
    choices, and distinguishes local default-branch matches from unverified
    sources. Unsupported Runtime history produces unresolved Reads.
-3. The Agent reconstructs continuous Task episodes with explicit
-   ownership/objective/outcome anchors and writes one observed/unresolved Read
-   plus at most one Effect in schema-v3 `task-judgments.jsonl`.
-4. `report` validates episode ownership and anchors, weak fragment-only
-   objectives, source ownership, windows, cross-Chat linkage, Read/choice
-   timing, deduplication, and aggregate conservation, then creates
-   `evidence.jsonl` and `REPORT.md`. An optional hash-anchored reviewed
-   baseline is shown separately, so a current collector gap cannot erase
-   previously reviewed positive cases or silently import them into the rerun.
+3. `task-source` removes every collector-derived Tree/Read/Effect projection
+   and keeps only authorized work messages. It does not redact literal
+   Tree/Read/Effect discussion from the work itself.
+4. The Agent reconstructs Tasks from that projection; `freeze-tasks` validates
+   and writes a digest-bound `task-inventory.jsonl`.
+5. The Agent writes digest-bound Read attributions, then zero-or-more Effect
+   judgments for each clear Task without changing the frozen inventory.
+6. `report` validates Task sources, weak fragment-only objectives, source
+   ownership, windows, cross-Chat linkage, Read/choice timing, inventory
+   digests, deduplication, multi-Effect conservation, and creates
+   `evidence.jsonl` and `REPORT.md`. An optional hash-anchored reviewed baseline
+   remains separate from the current rerun.
 
 There is no default time window. `--days` is an optional data-acquisition
 bound. Every available Task in the authorized bound is reported; sample size

evals/manual-behavior-checklist.md

@@ -4,7 +4,7 @@ Use this checklist before admitting a `context-tree-value-audit` revision. Run i
 in designated First Tree Agent workspaces against disposable or sanitized
 records. Never commit generated artifacts.
 
-For a 0.4.0 upgrade, confirm the installed payload exposes only
+For a 0.5.0 upgrade, confirm the installed payload exposes only
 `$context-tree-value-audit`; the superseded `$context-tree-insights` directory
 must not remain as a second callable Skill. Exercise the README's minimal
 move/copy/compare flow and confirm the old payload is outside both Skill
@@ -170,7 +170,7 @@ Pass when:
 - malformed receipt is omitted with `context_decision_invalid`;
 - no malformed receipt blocks the message, Chat export, or audit.
 
-## 5. Continuous Task reconstruction
+## 5. Pure Task reconstruction and freeze
 
 Use one Chat containing one objective across planning, implementation, review,
 QA, correction, and final delivery, followed by a genuinely independent
@@ -184,29 +184,32 @@ one or more leading Agent mentions; and concrete objectives that contain
 
 Pass when:
 
+- `task-source.jsonl` contains authorized work messages but no
+  collector-derived Reads, passages, Tree-mention indexes, decision receipts,
+  choice projections, or Effect judgments;
+- original work-message content remains byte-for-byte intact even when the
+  work itself discusses Tree, Read, or Effect, and those literal terms are not
+  treated as audit evidence during Task reconstruction;
 - all phases and continuations for the first deliverable remain one Task;
 - it produces two Task rows with separate source messages and windows;
-- clear Tasks have audited-Agent ownership, objective, material object scope,
-  independent outcome, bounded source fragments, one primary deliverable, and
-  explicit ownership/objective/outcome anchors;
+- clear Tasks have a concrete objective, independent same-Agent outcome,
+  bounded source fragments, and explicit objective/outcome sources;
+- material scope and primary deliverable may clarify the boundary but are not
+  required for research, diagnosis, review, or decision Tasks;
 - weak prompt variants and weak normalized objectives are merged or excluded,
   while concrete `continue` objectives remain eligible;
-- a weak assignment cannot borrow another sender's concrete objective, and one
-  Agent message cannot simultaneously serve as objective and terminal outcome;
-- an earlier weak objective anchor cannot admit a Read that precedes the
-  earliest ownership-compatible concrete objective anchor;
-- an earlier ownership anchor from a sender incompatible with the ownership
-  kind cannot admit a Read before compatible ownership is established;
-- every outcome anchor is a non-empty current-Agent message, so a later human
+- one Agent message cannot simultaneously serve as objective and terminal
+  outcome;
+- every outcome source is a non-empty current-Agent message, so a later human
   follow-up cannot extend the episode or become the bound Effect outcome;
-- work owned by another Agent remains context until a visible assignment,
-  transfer, or acceptance;
+- work owned by another Agent remains context until the audited Agent receives
+  or visibly accepts an objective;
 - a candidate missing any clear-Task gate is excluded with the deterministic
   structured exclusion kind;
-- excluded candidates contain no episode, Read, or Effect judgment;
-- Reads and choices outside the established episode are rejected;
-- ownership/objective/outcome anchors copied across Tasks are rejected;
-- one Read or choice copied into both Tasks is rejected.
+- excluded candidates contain no clear-Task sources, Read, or Effect judgment;
+- `freeze-tasks` writes one normalized inventory digest to every row;
+- changing any frozen Task row invalidates the digest;
+- Read/Effect fields in a Task inventory draft are rejected.
 
 ## 6. Cross-Chat handoff merge
 
@@ -228,31 +231,36 @@ Pass when:
 
 - the first is `observed` with valid Read IDs;
 - the second is `unresolved` with a reason;
-- unresolved Read has no Read IDs and a null Effect;
+- unresolved Read has no Read IDs and an empty Effect list;
 - no `not_observed`, `unused`, or negative-value state is emitted;
-- Reads outside the established episode, Task window, or source Chats are
+- Reads outside the frozen Task window or source Chats are
   rejected;
+- one Read cannot be attributed to different Tasks;
 - unresolved Tasks appear in coverage counts, never an unused denominator.
 
-## 8. Optional Effect
+## 8. Zero-or-more Effects
 
 Prepare valid examples of `confirmed`, `constrained`, `redirected`, and
-`conflicted`, plus observed-Read Tasks with no Effect.
+`conflicted`, including one Task with two distinct Effects supported by the
+same Read, plus observed-Read Tasks with no Effect.
 
 Pass when:
 
-- unknown Effects, multiple Effects, confidence tiers, and numeric weights are
-  rejected;
+- unknown Effects, confidence tiers, and numeric weights are rejected;
+- one Task can carry multiple Effects and the report separately counts Effect
+  Tasks and total Effects;
 - every Effect has observed Task Reads, later same-Agent choices, an outcome
-  anchor bound to the episode outcome, and a summary;
-- the bound Effect outcome anchor is not earlier than any cited Read completion
+  message inside the Task, and a summary;
+- the Effect outcome message is not earlier than any cited Read completion
   or choice;
 - post-choice Reads and out-of-window choices are rejected;
-- a null Effect requires one short `effect_reason`;
+- an empty Effect list requires one short `effect_reason`;
 - a decision receipt alone does not create an Effect;
 - superseded task types, sampling fields, `verified` / `probable`, rubrics,
   and support fields are rejected;
-- there is at most one Effect per Task.
+- the same Read may support multiple distinct choices;
+- one choice cannot be reused across Effects;
+- unresolved Reads cannot have Effects.
 
 ## 9. Sample handling
 
@@ -270,22 +278,23 @@ Pass when:
 
 Pass when the report includes:
 
-- a complete clear Task inventory with ownership, objective, scope, primary
-  deliverable, outcome, Read, Effect, and evidence summary;
-- every clear Task's ownership/objective/outcome anchors and boundary rationale;
+- a complete frozen Task inventory with objective, optional scope/deliverable,
+  outcome, source IDs, Read, zero-or-more Effects, and evidence summary;
+- the Task inventory digest and every clear Task's objective/outcome sources;
 - a complete excluded-candidate inventory with exclusion kind, observed scope,
   and reason;
 - observed and unresolved Read Tasks;
-- Effect Tasks and observed Reads without an Effect;
+- Effect Tasks, total Effects, and observed Reads without an Effect;
 - the four-Effect distribution;
 - authorized Chat/message/trace coverage and gaps;
 - the four-class in-window Tree-read attempt conservation table.
 
 Verify:
 
 - observed + unresolved = clear Tasks;
-- Effect + observed Read without Effect = observed Read Tasks;
-- the four Effect counts sum to Effect Tasks;
+- Effect Tasks + observed Read without Effect = observed Read Tasks;
+- the four Effect counts sum to total Effects;
+- total Effects is greater than or equal to Effect Tasks;
 - the report does not output a global effectiveness rate;
 - Read counts, receipts, and unresolved gaps are not represented as causal
   value or non-value.

scripts/validate_skill.py

@@ -33,6 +33,11 @@
     "chats.jsonl",
     "judgments.jsonl",
     "task-judgments.jsonl",
+    "task-source.jsonl",
+    "task-inventory-draft.jsonl",
+    "task-inventory.jsonl",
+    "read-attributions.jsonl",
+    "effect-judgments.jsonl",
 }