Oldtruthshouldn'tbecomenewcontext.
Retrieval optimises for relevance. A superseded claim is still relevant — which is why it keeps arriving as if it were current. Canon resolves claim history in HydraDB before an answer is grounded.
100k / 1M / 5M
“Volume discounts: predefined breaks at 100k, 1M, and 5M monthly tokens for hosted volume.”
dsid_1214ee9ab5e44de487c800f7a4771d7d
250k / 2M / 10M
“predefined breaks at 250k, 2M, and 10M monthly tokens (old doc says 100k/1M/5M — that’s outdated).”
dsid_10559147d3014931898864f22c311e47
Real spans from Hosted enterprise_playbook.monthly_token_volume_discount_breakpoints — 511,958 EnterpriseRAG-Bench documents indexed.
The result
Same corpus. Same retriever. Same prompt.
Only the context topology changes. Across all 20 official conflicting-information questions, this is how often a document that still asserts the retired value reached the model as present-tense context.
14/20
BM25 baseline
The superseded document is retrieved and enters the context unmarked.
1/20
Canon grounding
The one remaining case is contested — no supersession is established, so nothing is dropped.
20/20
Current evidence retrieved
Up from 18/20: the graph pins current evidence the retriever ranked too low.
How it works
A temporal claim graph between retrieval and generation.
Enterprise knowledge is time-bound assertions, not a bag of facts. Canon models the difference explicitly.
Step 1
Retrieve
SQLite FTS5 BM25 over every indexed document. Identical for both arms — Canon does not change the retriever.
Step 2
Resolve in HydraDB
Candidate documents are walked back to their claim keys. The graph returns the current proposition, the retired ones, and the supersession chain that connects them.
Step 3
Rebuild the context
Documents that assert a superseded value are replaced by the next candidate from the same ranking. History stays queryable; it just stops arriving as present tense.
Step 4
Answer
The same model and the same prompt as the baseline. Only the context topology changed.
Why HydraDB
The graph decides which evidence is current.
Not a visualisation layer. Remove HydraDB and the answers change — three graph operations run on every question.
20
Claim keys
One mutable property of one entity.
43
Assertions
Each with an exact evidence span and source document.
19
Supersession edges
Canon events that replaced an earlier canon event.
19
Canon transitions
1 contested — order not established, no winner invented.
Ask it something, then ask what it used to be.
Every answer returns a state — CANON, CONTESTED or UNKNOWN — with the evidence behind it, the retired context that was filtered, and the HydraDB queries that produced it. Never prose alone.