Tvijo agent system: PRIZMA, RUNLEDGER and acceptance
Tvijo is developing a controlled workflow from a repository task to an accepted result, with permissions, resource limits and attributable evidence. PRIZMA describes the intended resource profile; RUNLEDGER connects the work to its checks a…
🎯 For Teams running AI agents who need to know what a task actually cost and whether it was accepted.
- Separate transport success, task acceptance and money admission
- Read the resource profile (PRIZMA) before the first model call, not after
- Keep unknown cost unknown instead of filling it with zero
- Recognise where an orchestration layer stops earning its keep
Tvijo is developing a controlled workflow from a repository task to an accepted result, with permissions, resource limits and attributable evidence. PRIZMA describes the intended resource profile; RUNLEDGER connects the work to its checks and costs. These components have different verification states, so this report does not declare an integrated customer service ready.
This snapshot describes inspected source and dated project reports. A configuration, a passing unit test, an installed component and a complete customer workflow are separate levels of evidence.
What each component does
| Component | Responsibility | Boundary |
|---|---|---|
| Executor ladder and router | Select an eligible execution route | A configured alias does not prove availability or the actual served model |
| PRIZMA | Resolve resource and review requirements | A returned plan does not enforce a deadline, permission or budget |
| Enrichment tiers | Bound preparation of a task before execution | More preparation must justify its cost and delay |
| Existing worker and orchestrator | Execute and recover bounded work | Dispatch, delivery and completed work require separate receipts |
| Independent checking | Compare an artifact with frozen acceptance criteria | Agreement between models alone does not establish correctness |
| RUNLEDGER | Link task, attempts, artifacts, verdicts and cost provenance | A successful provider response is not an accepted task |
| Human workspace | Show the result and decisions to the owner | Visibility is distinct from approval or deployment |
PRIZMA's source configuration defines iterations, depth, time, parallelism, reviewers and change volume. Permission to write must remain separate from review depth. Commercial subscription tiers also remain separate from execution rights.
Current evidence
| Area | Observed status on 10 November 2026 | What remains |
|---|---|---|
| PRIZMA resolver | Source available in draft PR 221 at 89252008b8bfed18b8a9d6f5d670d33044a6a18b | Validation defects and runtime enforcement require completion |
| PRIZMA unit suite | The PR author reports 16 passing tests | This editorial review did not rerun pytest; it used direct isolated module calls |
| Resolver boundary checks | Direct calls accepted inconsistent judge settings, missing write scope and invalid axis or budget values | Fix the boundaries and add regression coverage before activation |
| Enrichment selection | Simple repository-write input selected RC2 while the conditional criteria were not consulted | Make selection depend on the declared conditions |
| RUNLEDGER delivery | Implementation commits were reported by the operator's agent | Those local commits and their final verdict were not independently inspected here |
| Customer workflow | The frozen programme-v3 report explicitly leaves installed-client acceptance and customer mapping unresolved | Produce one attributable customer-like end-to-end receipt |
The direct module checks do not demonstrate a live production exploit, a deployed fix or complete repository acceptance. Their exact technical details belong in the private evidence appendix.
The minimum useful workflow
- Define one task, its allowed paths, expected artifact, acceptance checks and authorised budget.
- Resolve an eligible executor and a sufficient resource profile.
- Reserve and enforce limits through the existing controller.
- Execute in an isolated workspace and retain attempt identities.
- Run deterministic checks and independent review where required.
- Show the artifact, evidence, actual costs and unresolved questions to the owner.
- Record acceptance separately from publication or deployment.
For a small independent edit, extra preparation and several reviewers may add little value. Parallel workers are useful when their inputs, outputs and permitted paths are independent. A reviewer should assess a frozen artifact and must not approve its own changes.
Roles proposed for the next integration task
Sol coordinates and integrates; Luna performs bounded inventory and documentation; Terra implements and checks scoped changes; Astra performs the final independent read-only review. This is a proposed allocation, not proof that a native swarm is running. Actual dispatch, model identity and receipts must be recorded.
Cost and recovery
Keep estimated price, authorised cap, reservation, measured provider usage, final billing evidence and allocated subscription cost distinct. Unknown cost remains unknown. A provider export covering an account or key window cannot by itself price one task.
The task budget includes preparation, failed attempts, child work and review. A retry must not reset the lifetime budget. After interruption, recovery must reconcile completed actions and uncertain outcomes before retrying.
How to keep the design small
Reuse the existing worker, budget controller and event journal. Complete one task class before introducing more adapters, resource classes or dashboards. Measure accepted results, end-to-end time, human intervention and total attributable cost against a simple baseline.
Repository-wide line counts and database counts are inventory signals. They do not measure the size or usefulness of this particular orchestration layer without a defined scope.
Product boundary
The proposed customer offer is a bounded task workflow with evidence and spending control. Assisted onboarding and a narrow paid pilot can validate demand. Subscription pricing, included capacity and support remain hypotheses until measured; no commercial availability is established by this report.
Sources and limits
The following are first-party project sources inspected on 10 November 2026. GitHub sources may require repository access; this is not an independently reproducible public benchmark.
- PRIZMA source and tests, draft PR 221, pinned head 89252008b8bfed18b8a9d6f5d670d33044a6a18b.
- Programme-v3 delivery report.
- Model-court decision.
- Wiki editorial workflow, rendered in the browser during this review.
Companion Lab manuscript: “Four ways agent measurements can mislead”. Its public address is intentionally omitted until the publication route is verified. Human editorial review is pending.
Update 2026-11-10: two claims replaced by receipts
- PRIZMA resolver.
pytest -q tests/test_royorcjud_prizma.py→ 16 passed, rc 0, commit89252008b. This supersedes the earlier wording that the PR author reported the tests: the run was repeated in this session. - RUNLEDGER. The vertical slice is committed as
49d633a2a. Its report records 3 independently accepted attempts out of 8 preserved, with the cost of an accepted result and the route-failure share left UNKNOWN rather than filled with zero, and money admission returning2(incomplete) instead of passing. - Unchanged. AC-7 (production routing and model quarantine) is still NOT_RUN: it was moved to a separate stage by the contract delta, so the routing decision lives in configuration, not in the code path.
Sources
Source links and observation dates are recorded below. Review status is stated for each material.
🔗 Article connections
Open the full knowledge graph ↗Published article links. Select a node to open its article at the heading.
Show the connection graph
Article links · keyboard / text view
Same topic
- Read agent communication receipts without mistaking delivery for a human read
- Audit an AI technology radar against primary sources before promoting a story
- Run an AI coding harness evaluation offline before any live pilot
- DSHREM on a phone: SSH, tmux, and safe task handoff
- DSHREM owner practice: cwd, worktrees, WIP, and root boundaries
- DSHREM economy swarm: checkpoints, routes, and real cost evidence
Next steps
Related reading
Read agent communication receipts without mistaking delivery for a human read
Find the current stage of a swarm message without confusing durable storage, delivery, runtime receipt and viewing. This guide connects local checks, the private observation screen and distinct receipts, so a delivery acknowledgement does not become an unsupported claim that an action completed.
AI and process foundationAudit an AI technology radar against primary sources before promoting a story
Check a technology news item against its primary source, version and event date before deciding to adopt it. This guide follows a fixed ten-item sample, separates facts from interpretation and forecasts, and preserves limitations and the condition for checking the claim again.
AI and process foundationRun an AI coding harness evaluation offline before any live pilot
Check the coding-agent evaluation path without a model request, then read eight real engineering pilot slots. This guide connects commands, independent artifact acceptance, full elapsed time and expense receipts, explaining why a saved response differs from a newly accepted result.
💬 Comments