Shadow Mode Is a Standalone, Read-Only Advisor That Translates a Case on Request
Context and Problem Statement
CERT/CC wants Vultron's view of cases it already tracks in another system, starting with VINCE: where a case stands, which CVD desiderata it has met or broken, and which moves are open from here. Other trackers, such as GHSA or Jira, raise the same need.
Planning for the VINCE shadow (Epic #5954) described it as a running Vultron node that kept a local copy of each VINCE case as CERT/CC's Case Manager and ran the protocol's behavior trees on it. That framing made three problems: the node's trees would emit activities that a server mode had to suppress (#5962), acts by participants who are not on Vultron had no truthful way into the local case (#5965), and the node needed a fail-closed server mode with storage and logging rules attached (#6336).
What is shadow mode, and what does its first version leave out?
Decision Drivers
- The first question to answer is whether a case from an outside system can be translated into Vultron's model and give useful insights. Nothing beyond that is needed to answer it.
- The first version must advise the people working a case, not act in it. A shadow that runs the protocol on a case is a second player at the board.
- Outside-system cases hold embargoed data, so every component that holds or sends that data is a place it can leak.
- Translation cost and the share of mappings that need judgment rather than mechanical translation are unknown until tried.
Considered Options
- A Vultron node in shadow mode: CERT/CC's actor runs its trees on a local replica, with outward delivery suppressed.
- A standalone tool that reads, translates and analyses one case on request, then discards it.
Decision Outcome
Chosen option: "A standalone tool that reads, translates and analyses one case on request, then discards it", because it answers the first question with no actor, no network presence and no act to fake.
- Advisor, not player. The shadow tells an analyst where a case stands and which moves are open or better. This version does not act in the case, does not write to the outside system, and does not send anything to the case's participants.
- Pull, and follow. An analyst names one case. The shadow reads it, translates it, analyses it and shows the result. For a live Vultron case, the shadow can follow the case's ledger as it grows and show the analysis again after each new entry.
- Ephemeral. The shadow holds a case and its translation in memory only. It writes neither to persistent storage, and it leaves none of either on the host after its process exits. Transient working files are allowed only if they are gone when the process exits. A later need for the case is met by reading it again.
- Standalone. The shadow uses the Vultron core as a library. It hosts no actor and has no inbox, outbox or activity delivery, so "delivers nothing" holds because no delivery path exists, not because a setting turns one off.
- One standing, with moves in bands. The shadow reads the translated case as it stands: each participant's states, the desiderata met or broken, and the moves open to each participant, each move named by its use case or action rule, never by its wire activity type, and labelled with who can make it. Each move gets an outcome shift for each vendor: the change in expected outcome from the chance value, solved over the same chance model as the desiderata baselines. The page does not pick one best move. It shows the harm-free moves first, then the trade-off moves, and collapses the unhelpful moves. In the harm-free and trade-off bands, it shades each move by its layer in the dominance order and marks the top moves. It marks each coordinator move that is leading among the coordinator's own moves. How a move gets its band is ADR-0200. Scoring a move's setting, such as an embargo's duration, is a later decision.
- Measure before building on it. Translating real-shaped cases, timing it, and sorting each mapping into translation or interpretation comes before the analysis is built on top.
- This is the first version. Every "does not" above describes the shadow this record decides. A later version that writes to an outside system, acts in a case, or sends to participants is a later decision, and that record supersedes the parts of this one it changes.
How a case is read from an outside system, and what its translation contains, is ADR-0191. How the shadow is presented and packaged is ADR-0192.
Consequences
- Good, because an act by a participant who is not on Vultron needs no door into a live case: it becomes an entry in a translation (#5965).
- Good, because no activity can leave the shadow, with no suppression code to get wrong (#5962).
- Good, because a long-lived store of outside-system cases does not exist to protect.
- Good, because the same translation is the first half of a later gateway between two outside systems, each translated to and from Vultron.
- Bad, because every request pays the full read and translation cost, which is unmeasured.
- Bad, because a pull tool cannot alert anyone when a case changes. A watcher that runs the same pull on a schedule can be added on top.
- Neutral, because the read-write gateway, where Vultron or an outside system does act through the protocol, is a separate later decision that this record leaves open.
Validation
specs/case-shadow.yaml states the checkable rules (no delivery or write path, no long-lived copy), each citing this record; its packaging rules cite ADR-0192.
The data rules for VINCE in ADR-0189 still apply.
Pros and Cons of the Options
A Vultron node in shadow mode
- Good, because the same deployment shape could later take part in cases as a peer.
- Bad, because its trees run as CERT/CC's Case Manager on a copy of a case CERT/CC runs elsewhere, so Vultron acts rather than advises.
- Bad, because every act by a participant outside Vultron must either be credited to CERT/CC or synthesized as that participant's own message.
- Bad, because no-delivery, storage and log rules all depend on a server mode that a half-configured server can get wrong.
A standalone tool that translates on request
- Good, because it answers the first question with the least machinery.
- Good, because it has nothing to suppress and nothing to keep.
- Bad, because translation cost is paid on every request.
Amendment — 2026-11-13
Rule 1 was replaced, and rule 7 was added. Rule 1 read: "Advisor, not player. The shadow tells an analyst where a case stands and which moves are open or better. It never acts in the case, never writes to the outside system, and never sends anything to the case's participants." It now reads: "Advisor, not player. The shadow tells an analyst where a case stands and which moves are open or better. This version does not act in the case, does not write to the outside system, and does not send anything to the case's participants." Rule 7 reads: "This is the first version. Every "does not" above describes the shadow this record decides. A later version that writes to an outside system, acts in a case, or sends to participants is a later decision, and that record supersedes the parts of this one it changes."
The maintainer asked that the record state what this version does, not what the shadow will never do: a later version may write to VINCE.
- Source: #6336 (PR #6364 review)
Amendment — 2026-11-13 (revision 3)
Rule 5 was replaced. It read: "Insights come in levels. The first level reads the translated case as it stands: each participant's states, the desiderata met or broken, and the moves the state machines allow. The second level ranks each allowed move by the score the case would have if that move came next. A multi-step search or simulation of future play is a later decision." It now reads: "One standing, scored. The shadow reads the translated case as it stands: each participant's states, the desiderata met or broken, and the moves open to each participant, each move named by its use case or action rule, never by its wire activity type, and labelled with who can make it. Each move is scored by how far it changes the case's expected outcome from here, solved over the same chance model as the desiderata baselines. The page leads with the case's best move, then the moves the coordinator can make now. Which single number stands for a multi-vendor case's outcome is a later decision, made after a comparison of candidate aggregates against cases that would make the advice work against the collective good. Scoring a move's setting, such as an embargo's duration, is a later decision too."
Planning for #5966 found that ranking a move by the score of the history with that move added credits almost nothing, because most desiderata are orderings that one move does not settle: notifying a vendor makes "vendor aware before public" more likely without deciding it. The change in expected outcome credits it, and computing it needs only the legal moves the first level already lists, so the two levels are one. The maintainer asked for the advice of a card player's coach: the best move for the case, whoever makes it, and the moves open to the coordinator now, optimized for the case rather than for one participant.
- Source: #5966
Amendment — 2026-11-16 (revision 4)
Rule 5 was replaced. It read: "One standing, scored. The shadow reads the translated case as it stands: each participant's states, the desiderata met or broken, and the moves open to each participant, each move named by its use case or action rule, never by its wire activity type, and labelled with who can make it. Each move is scored by how far it changes the case's expected outcome from here, solved over the same chance model as the desiderata baselines. The page leads with the case's best move, then the moves the coordinator can make now. Which single number stands for a multi-vendor case's outcome is a later decision, made after a comparison of candidate aggregates against cases that would make the advice work against the collective good. Scoring a move's setting, such as an embargo's duration, is a later decision too." It now reads: "One standing, with moves in bands. The shadow reads the translated case as it stands: each participant's states, the desiderata met or broken, and the moves open to each participant, each move named by its use case or action rule, never by its wire activity type, and labelled with who can make it. Each move gets an outcome shift for each vendor: the change in expected outcome from the chance value, solved over the same chance model as the desiderata baselines. The page does not pick one best move. It shows the harm-free moves first, then the trade-off moves, and collapses the unhelpful moves. In the harm-free and trade-off bands, it shades each move by its layer in the dominance order and marks the top moves. It marks each coordinator move that is leading among the coordinator's own moves. How a move gets its band is ADR-0200. Scoring a move's setting, such as an embargo's duration, is a later decision."
The maintainer asked for options, not one answer. Many moves are good at a given point in a case, and some are clearly better than others. One case score would reduce all the desiderata to one number, and one best move hides the other good moves. Thus the page puts the moves in bands and shows the trade-offs between them. A CVD case is a cooperative game, so a move that harms no vendor is in a higher band than a move that trades one vendor's loss for another's gain.
- Source: #6450
Amendment — 2026-11-16 (revision 5)
Rules 2 and 3 were replaced. Rule 2 read: "Pull. An analyst names one case. The shadow reads it, translates it, analyses it and shows the result. Nothing runs between requests." It now reads: "Pull, and follow. An analyst names one case. The shadow reads it, translates it, analyses it and shows the result. For a live Vultron case, the shadow can follow the case's ledger as it grows and show the analysis again after each new entry." Rule 3 read: "Ephemeral. A translation is built for the request that needs it and discarded afterwards. The shadow keeps no long-lived copy of an outside-system case or its translation. Transient working files within a request are allowed. A later need for the case is met by translating it again." It now reads: "Ephemeral. The shadow holds a case and its translation in memory only. It writes neither to persistent storage, and it leaves none of either on the host after its process exits. Transient working files are allowed only if they are gone when the process exits. A later need for the case is met by reading it again."
The maintainer asked that the shadow follow a live Vultron case, so the dashboard moves with the case. The maintainer also said what the ephemeral rule guards against: a copy of embargoed data left on the host. Holding a case in memory while the process runs does not leave one, so the rule now names the host and the process, not the request. The rule applies to every case the shadow reads, an outside-system case included.
- Source: #6370
More Information
Planned in #6336, with the maintainer, on 2026-11-12. The VINCE shadow's data-handling rules are ADR-0189.
Generated spec requirements: case-shadow.yaml SHD-01, SHD-04.