Story-2075 Walkthrough

Change: CE-16 of the Context Engine (Epic-2049, RFD pramaan-functions/docs/notes/rfd-context-engine.md §14.3 row CE-16; §7 EngineState, §8 the tool catalog). CE-15 (Story-2074) wired Mira to consume the ContextBundle and stamp context_provenance on turn.completed. CE-16 is the measurement: a non-empty set of fixture packs (agents/mira/evals/datasets/context/*.yaml) and a Why+Cite grade that switches its bar on engine_state — the RFD's honesty mechanism, exercised.

What the RFD asks for

Two clauses drive the grade:

§7: “the eval harness switches grading rules on [engine_state]”, and “a cited fact whose span is not in the pack fails closed regardless of engine state.” Canned spans live only in eval fixtures — never in a production stub.
§8 (the cite(span_id) row): “eval harness grading Why+Cite mechanically | document_pages byte re-verification.”

The switch

`engine_state`	The bar	Why
`live` (bundle served)	Why + byte-verified Cite	The engine served real record: every produced value must carry a one-sentence producer rationale (Why, §4) and a `source_span.v1` whose quote byte-verifies.
`partial` / `stubbed` / no bundle	frame-only closure	Not-yet-backfilled / stub / frame-only turn — grading degrades gracefully, a Cite is not required (honest empties, never canned content).
ANY state, span present	fail-closed fabrication check	A span that does not byte-verify is fabrication and reds the pack even when stubbed. This invariant never degrades.

Byte re-verification — the offline `cite(span_id)`

verify_span() is the whole mechanical Cite grade: it resolves a span against the fixture's canned documents (the offline document_pages) and checks the quote is byte-present at the anchored coordinates — byte-exact, no normalization (“a court reads the bytes”, §4). With char_start/char_end the sliced unit must equal the quote; without them the quote must be byte-present in the block or page. A stale or forged span returns verified=False → fabrication.

The golden command set over the packs

Five packs, realistic Indian-litigation matter shells, one per honesty state:

Pack	engine_state	Matter
`live-supply-breach`	`live`	Ferro-alloy supply-contract arbitration — GRN + ledger spans
`live-cheque-dishonour`	`live`	s.138 NI Act cheque bounce — return memo + demand notice (page anchor)
`partial-backfill`	`partial`	RFCTLARR land-acquisition compensation — award span, bar dropped
`stubbed-empty`	`stubbed`	Carriage-of-goods stub pre-fetch — honest empty
`no-bundle`	—	Demurrage frame-only turn — null-hash provenance marker

Each pack's golden command runs through the REAL command route (run_command) with the pack's bundle threaded in, an honest stubbed panel (no key, no egress). The turn's context_provenance is asserted to mirror the served engine_state, then the switch grades Why+Cite.

The switch flips both ways (the proof it measures something)

Mirroring the adversarial suite's RED-on-removal proof, test_context_evals.py takes a real live pack and:

strips its evidence, re-grades as live → RED (the Why+Cite bar is unmet);
flips the same empty pack to stubbed → GREEN (the bar drops). That contrast is the engine_state switch, proven;
corrupts a span's quote at live/partial/stubbed → RED at every state (fail-closed never degrades).

Run it

cd ~/projects/pramaan-io/pramaan-agents/agents/mira
uv run python -m evals.run context      # OFFLINE — no key (the blocking CI gate)
uv run pytest tests/test_context_evals.py -q

Offline, deterministic, keyless — the shape the RFD grading contract names (“deterministic invariant checks run on EVERY PR”). A red grade blocks merge (.github/workflows/ci.yml, the Mira context fixture-pack gate). The live-model / judge lane over the same packs is the separate, non-blocking sample.

Acceptance

Criterion	Where
fixture-pack eval run green	`run.py context` — 5 packs held / 0 failures; `test_context_evals.py` (13 tests)
grading switches on engine_state	`context.grade_pack` — live = Why+Cite, degraded = frame-only closure; both directions proven

Closes pramaan-io/pramaan-spec#2075.