|
2 | 2 |
|
3 | 3 | ## Unreleased |
4 | 4 |
|
| 5 | +- **A declaration cannot discharge a category it does not cover, and a |
| 6 | + published schema keeps its bytes.** Three follow-ups to the monotone |
| 7 | + declaration rule, each a defect that shipped with it |
| 8 | + ([#409](https://github.com/ThreeMoonsLab/agents-shipgate/issues/409), |
| 9 | + [#411](https://github.com/ThreeMoonsLab/agents-shipgate/pull/411)). |
| 10 | + |
| 11 | + *Effects are risk-ordered; their obligations are not.* The monotone |
| 12 | + comparison read the effect rank alone, so declaring `financial_write` over an |
| 13 | + inferred `external_communication` read as escalation and stayed silent — |
| 14 | + `financial_write` requires approval, audit, and idempotency but **not** |
| 15 | + confirmation, which is exactly what communicating outward requires. The |
| 16 | + action reported pass-eligible with no gap and no |
| 17 | + `SHIP-ACTION-EXTERNAL-COMMUNICATION-AUDIT-MISSING` while its external-write |
| 18 | + risk tag sat untouched in the same report. A declaration now accounts for an |
| 19 | + observation only when it ranks at or above it *and* obliges at least that |
| 20 | + observation's built-in controls, from the new `BUILTIN_EFFECT_OBLIGATIONS` |
| 21 | + table — pinned to the branches it mirrors by a test that walks each entry |
| 22 | + through a real scan. Coverage also reads every policy-eligible claim rather |
| 23 | + than the `effect` field alone, so a `risk_tags: [financial_action]` entry |
| 24 | + accounts for an inferred `financial_write` exactly as the control evaluator |
| 25 | + already treats it. |
| 26 | + |
| 27 | + *The two comparators disagreed.* The declaration rule compared `_EFFECT_RANK` |
| 28 | + while `_non_authoritative_effect_escalation_support` compared |
| 29 | + `ACTION_EFFECT_RANK`, and the two order `write` and `privileged_data_access` |
| 30 | + oppositely — so a declaration could read as covered in one place and raise |
| 31 | + `mixed_policy_evidence` in the other, a verdict no override could reach. Both |
| 32 | + now call `declaration_covers`, which requires *both* orders to agree; nothing |
| 33 | + that gated before stops gating. |
| 34 | + |
| 35 | + *Four published schema documents were mutated in place.* The new |
| 36 | + `declaration_below_inferred_evidence` value was written into |
| 37 | + `packet-schema.v0.12.json`, `verifier-schema.v0.9.json`, |
| 38 | + `capability-lock-schema.v0.6.json`, and `capability-lock-diff-schema.v0.7.json` |
| 39 | + while they kept their version identifiers — and the two capability-lock |
| 40 | + documents had no successor version at all, so a consumer pinned to any of the |
| 41 | + four rejected artifacts that document is supposed to describe. |
| 42 | + `generate_schemas.py --check` cannot catch this: it proves committed == |
| 43 | + generated, never that a content change moved the version. All four are |
| 44 | + restored byte-for-byte, the capability lock advances `0.6` → `0.7` and its |
| 45 | + diff `0.7` → `0.8`, and a lock written under `0.6` is advanced on read rather |
| 46 | + than rejected — the normalizer handled only `0.1`–`0.4`, so the bump would |
| 47 | + otherwise have orphaned every committed `capabilities.lock.json`. |
| 48 | + |
| 49 | + Also: the row now names *every* uncovered observation rather than the |
| 50 | + strongest one alone, so a reviewer is not asked to acknowledge evidence the |
| 51 | + row never showed them; and the published remedy is true of its state — |
| 52 | + "raise the effect" is not printed where the declared effect already outranks |
| 53 | + the observation. |
| 54 | + |
5 | 55 | - **A declaration weaker than the evidence inferred for it is no longer |
6 | 56 | silent.** Declaring `effect: read` on a tool this scanner itself tagged |
7 | 57 | `external_write` was accepted with zero findings: the pre-existing |
|
0 commit comments