omnius 1.0.673 → 1.0.675

package/docs/ADJUDICATION.md

@@ -0,0 +1,257 @@
+# Evidence-bound adjudication
+
+`adjudicate` resolves a genuine decision impasse. It does not replace normal
+engineering judgment. It creates a small, tools-free decision environment. It
+then runs independent constituent assessments and a final judge.
+
+The tool separates three things:
+
+1. Evidence is part of the caller-supplied admissible record.
+2. Arguments are claims about that evidence. Arguments are not evidence.
+3. Constituent findings are analysis. Agreement between constituents does not
+   create a new fact.
+
+The host validates all identifiers, citations, panel results, and the verdict.
+The tool holds the case when it cannot validate the required quorum or final
+contract. It does not guess a result.
+
+## When to use it
+
+Use `adjudicate` when all of these conditions are true:
+
+- The active task has one exact, unresolved decision.
+- Two or more outcomes remain materially plausible.
+- The decision affects the next action.
+- The available evidence conflicts or supports different risk tradeoffs.
+- A fresh, impartial context can assess the decision more reliably than the
+  main loop's large working context.
+
+Do not use it for an ordinary implementation choice, a factual lookup, or a
+way to avoid gathering missing evidence.
+
+## Minimum call
+
+```json
+{
+  "question": "Should the failed release be rolled back or repaired in place?",
+  "allowedOutcomes": ["rollback", "repair_in_place"],
+  "evidence": [
+    {
+      "id": "E1",
+      "kind": "test_result",
+      "source": "integration run 2026-08-31T18:00:00Z",
+      "content": "The post-deploy integration suite failed 14 of 80 tests."
+    },
+    {
+      "id": "E2",
+      "kind": "artifact",
+      "source": "release receipt sha256:abc",
+      "content": "The prior artifact matches the last known-good receipt."
+    }
+  ]
+}
+```
+
+The case framer derives three distinct constituent assignments by default. Set
+`panelSize` from 2 through 6 to change that count.
+
+## Detailed call
+
+```json
+{
+  "caseId": "release-impasse-2026-08-31",
+  "question": "Should the failed release be rolled back or repaired in place?",
+  "context": "The release is paused. No mutation is authorized during this decision.",
+  "allowedOutcomes": ["rollback", "repair_in_place"],
+  "burdenOfProof": "clear_and_convincing",
+  "decisionRules": [
+    "Prefer verified recovery evidence.",
+    "Do not infer a successful recovery from panel agreement."
+  ],
+  "evidence": [
+    {
+      "id": "E1",
+      "kind": "test_result",
+      "source": "integration run 2026-08-31T18:00:00Z",
+      "content": "The post-deploy integration suite failed 14 of 80 tests.",
+      "reliability": 0.98
+    },
+    {
+      "id": "E2",
+      "kind": "artifact",
+      "source": "release receipt sha256:abc",
+      "content": "The prior artifact matches the last known-good receipt."
+    },
+    {
+      "id": "E3",
+      "kind": "observation",
+      "source": "operations log",
+      "content": "No rollback rehearsal exists for the new migration."
+    }
+  ],
+  "arguments": [
+    {
+      "id": "A1",
+      "position": "rollback",
+      "claim": "The known-good artifact makes rollback more reversible.",
+      "citedEvidenceIds": ["E1", "E2"]
+    },
+    {
+      "id": "A2",
+      "position": "repair_in_place",
+      "claim": "Rollback has untested migration risk.",
+      "citedEvidenceIds": ["E3"]
+    }
+  ],
+  "constituents": [
+    {
+      "id": "correctness",
+      "label": "Correctness",
+      "question": "Which outcome is best supported by functional evidence?",
+      "evidenceIds": ["E1", "E2"],
+      "argumentIds": ["A1"],
+      "decisionRuleIds": ["rule-1", "rule-2"]
+    },
+    {
+      "id": "reversibility",
+      "label": "Reversibility",
+      "question": "Which outcome has the safest verified recovery path?",
+      "evidenceIds": ["E2", "E3"],
+      "argumentIds": ["A1", "A2"],
+      "decisionRuleIds": ["rule-1", "rule-2"]
+    },
+    {
+      "id": "risk",
+      "label": "Operational risk",
+      "question": "What unsupported risk remains for each outcome?",
+      "evidenceIds": ["E1", "E3"],
+      "argumentIds": ["A2"],
+      "decisionRuleIds": ["rule-1", "rule-2"]
+    }
+  ],
+  "quorum": 2,
+  "maxConcurrency": 3,
+  "timeoutMs": 120000,
+  "maxTokensPerConstituent": 2048,
+  "maxJudgeTokens": 3072
+}
+```
+
+`insufficient_evidence` is a reserved verdict. Do not include it in
+`allowedOutcomes`. The host makes it available to constituents and the judge.
+It produces a held case instead of a binding decision.
+
+## Execution phases
+
+### 1. Record admission
+
+The host validates the input before inference. It rejects duplicate IDs,
+unknown citations, invalid outcome names, incomplete manual assignment
+coverage, and a quorum that exceeds the panel size.
+
+The host creates a canonical case record and SHA-256 record hash.
+
+### 2. Case framing
+
+If the caller omits `constituents`, a tools-free case framer creates 2 through
+6 assignments. Each assignment has a distinct question and an explicit subset
+of evidence, arguments, and decision rules.
+
+The host rejects unknown IDs and incomplete evidence coverage. It retries one
+time with validation codes only. The invalid model output is not added to the
+repair prompt.
+
+### 3. Constituent fan-out
+
+The host runs assignments with bounded concurrency. Each constituent gets a
+fresh context that contains only:
+
+- the exact case question and allowed outcomes;
+- the stated burden;
+- its assignment;
+- its assigned decision rules;
+- its assigned evidence;
+- its assigned arguments.
+
+The constituent gets no tools. It has no mutation authority. It must return a
+public assessment, a recommendation, confidence, cited evidence IDs, claims,
+counterarguments, and uncertainties.
+
+The host rejects unassigned citations, unknown outcomes, identity mismatches,
+and invalid schemas. One strict repair attempt is allowed.
+
+### 4. Quorum
+
+The default quorum is two thirds of the requested panel, rounded up, with a
+minimum of two. A caller can set a stricter quorum.
+
+The tool holds the case if too few findings pass host validation. It does not
+invoke the judge without quorum.
+
+### 5. Judge synthesis
+
+The judge gets the immutable full case record and validated constituent
+findings. The prompt labels findings as analysis rather than evidence.
+
+The judge must:
+
+- apply the burden and decision rules;
+- cite only admitted evidence IDs;
+- consider every validated constituent;
+- accept or reject every constituent finding exactly once;
+- preserve material dissent;
+- select an allowed outcome or `insufficient_evidence`.
+
+The host validates this contract. It holds the case after two invalid judge
+responses.
+
+### 6. Durable receipt
+
+Artifacts are written atomically under:
+
+```text
+.omnius/adjudications/<case-id>/run-<timestamp>/
+  case.json
+  assignments.json
+  findings.json
+  verdict.json
+  receipt.json
+```
+
+`verdict.json` is absent when the case holds before a valid verdict. The
+receipt records hashes, quorum, validation failures, elapsed time, and final
+status.
+
+## CLI display
+
+The CLI opens one live block when record admission starts. It shows:
+
+- case identity and decision question;
+- record size, burden, and allowed outcomes;
+- constituent assignments;
+- separately attributed public assessment streams;
+- host validation results;
+- the judge's public rationale stream;
+- final verdict, citations, status, and receipt path.
+
+The block has a `read mode` row. Select it to expand all retained rows. The
+display retains a bounded amount of text. It strips terminal controls and uses
+the normal secret redactor. It does not display provider hidden reasoning.
+
+## Harness
+
+Run the deterministic impasse harness:
+
+```bash
+pnpm harness:adjudication
+```
+
+The harness does not load a model and does not use a GPU. It verifies parallel
+constituent overlap, streamed public assessments, judge synthesis, and receipt
+persistence. The focused automated tests also cover invalid citations, strict
+repair, partial valid quorum, automatic case framing, and no-judge behavior
+when quorum fails.
+
+The implementation lives in
+`packages/orchestrator/src/adjudication.ts`. The live projection lives in
+`packages/cli/src/tui/adjudication-live-block.ts`.

package/docs/DISCOVERY.json

@@ -28266,6 +28266,36 @@
         "api.tools"
       ]
     },
+    {
+      "id": "guide.adjudication-uppercase",
+      "kind": "guide",
+      "title": "Evidence-bound adjudication",
+      "summary": "adjudicate resolves a genuine decision impasse. It does not replace normal engineering judgment. It creates a small, tools-free decision environment. It then runs independent constituent assessments and a final judge.",
+      "keywords": [
+        "ADJUDICATION",
+        "md"
+      ],
+      "maturity": "stable",
+      "audiences": [
+        "user",
+        "integrator",
+        "coding-agent"
+      ],
+      "layer": "documentation",
+      "interfaces": [
+        {
+          "type": "file",
+          "target": "docs/ADJUDICATION.md"
+        }
+      ],
+      "references": [
+        {
+          "type": "documentation",
+          "target": "docs/ADJUDICATION.md",
+          "relation": "canonical-artifact"
+        }
+      ]
+    },
     {
       "id": "guide.agent-memory-index",
       "kind": "guide",

package/docs/DISCOVERY.md

@@ -403,6 +403,7 @@ Daemon equivalents are `GET /v1/discovery/bootstrap`, `GET /v1/discovery?q=<inte
 
 | ID | Title | Summary |
 | --- | --- | --- |
+| `guide.adjudication-uppercase` | Evidence-bound adjudication | adjudicate resolves a genuine decision impasse. It does not replace normal engineering judgment. It creates a small, tools-free decision environment. It then runs independent constituent assessments and a final judge. |
 | `guide.agent-memory-index` | Agent Memory Index | Use the Omnius docs skills when an agent needs to explore the documentation corpus instead of loading the whole docs tree. |
 | `guide.agent-memory-index-uppercase` | Agent-Explorable Documentation | Omnius documentation is exposed to agents through project-local AIWG-style bundles under .aiwg/addons/. |
 | `guide.architecture-agent-system-map` | Omnius Agent System Map | Use this page when you need to understand how a user-visible behavior travels through Omnius, where its state lives, and which package owns a change. For a specific task recipe, search the generated catalog first: |

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.673",
+  "version": "1.0.675",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.673",
+      "version": "1.0.675",
       "bundleDependencies": [
         "image-to-ascii"
       ],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.673",
+  "version": "1.0.675",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/library.js",