CANON is a human-in-the-loop evidence review tool for teams that need to turn a messy literature or source set into a cautious, cited briefing.
It is not a truth engine, autonomous researcher, or replacement for expert review. It is designed to do a few narrow jobs well:
- find candidate evidence across a controlled corpus
- explain why each source was retrieved
- show whether support is strong, weak, mixed, or missing
- surface disagreement and source-quality concerns before drafting
- produce a cited answer that a human can accept, revise, or reject
The primary user is an analyst, policy researcher, strategy associate, graduate researcher, or product researcher who needs to answer questions such as:
- What does the literature say about this policy claim?
- Which sources support, qualify, or conflict with this conclusion?
- Are we relying on narrow, stale, low-quality, or one-sided evidence?
- Where should a human reviewer spend attention first?
CANON helps a reviewer move from a broad question to an evidence-backed first brief faster, while keeping uncertainty visible.
The product does not promise final truth. It promises a better review queue: ranked evidence, cited drafts, conflict notes, source diagnostics, and explicit limitations.
- A user asks a focused question.
- CANON retrieves candidate evidence from a named corpus.
- The system ranks evidence using relevance, source quality, citation signals, recency, diversity, conflict awareness, and safety checks.
- CANON returns a cited answer with support level, limitations, and conflict notes.
- A human reviewer inspects the evidence, edits the answer, marks weak or misleading citations, and decides what can be used.
Search quality is shaped by word choice. A new user may describe a topic in plain language while the corpus uses field-specific terms, methods, acronyms, or policy phrases. CANON should make that gap visible and help the user learn the language of the field.
The goal is not to hide retrieval sensitivity. The goal is to teach it.
This feature must be connected directly to semantic retrieval, not bolted on as a generic prompt-tip layer. It should use two linked signals:
- User-query semantic retrieval: compare the user's exact query against the corpus and expose which query terms, bigrams, and semantic matches drove or failed to drive retrieval.
- Result-neighborhood semantic retrieval: inspect the high-ranking semantic result set and extract the field phrases, recurring concepts, and semantic neighbors that the corpus itself uses for the same topic.
The product question is:
What did the user's wording retrieve, and what language do the retrieved results suggest the field uses instead?
CANON should also learn from prior review patterns. If users repeatedly accept evidence, phrases, or query variants around a topic, those patterns should become stronger suggestions in future sessions. If users reject a term as tangential, promotional, or misleading, that term should become less likely to appear as a suggested expansion.
This creates a controlled exploration loop:
- original query terms keep the anchor
- accepted result-neighborhood terms expand the user's vocabulary
- rejected terms reduce future drift
- lower-probability thematic terms can appear when they remain close to the original query and the accepted evidence neighborhood
When a user asks a question, CANON should show:
- matched query terms that helped retrieval
- important query terms that found little or no evidence
- field phrases that appear often in high-ranking evidence
- suggested alternate phrasings to try
- semantic neighbors that retrieve different evidence
- terms and phrases from the semantic result set that were not present in the original user query
- whether results changed materially under query rewrites
Example:
User query:
Do sanctions work?
CANON may suggest:
- economic sanctions effectiveness
- sanctions compliance
- sanctions enforcement
- coercive diplomacy
- targeted sanctions
- sanctions evasion
- regime change outcomes
- humanitarian effects of sanctions
The user learns that the field may not only ask "do sanctions work?" It may ask which sanctions, under what mechanism, against which outcome, and with which tradeoffs.
This should be presented as a learning aid, not as automatic query rewriting that silently changes the user's intent.
Definition of done:
- each answer shows a compact query-language panel
- low-match user terms are visible
- suggested field phrases are tied to retrieved evidence, not invented
- suggestions include evidence-result phrases discovered from the semantic neighborhood, not only terms from the original query
- exploratory suggestions are labeled by distance from the original query: close match, field synonym, adjacent concept, or exploratory tangent
- alternate phrasings can be run side by side
- the system reports whether the answer is stable or sensitive across rewrites
- user-selected query variants are kept in the review audit trail
Why this matters:
People learning a field also learn its vocabulary. If CANON exposes the terms that move retrieval, it becomes both an evidence tool and a topic-learning tool.
Help a reviewer produce a first-pass cited brief from a focused question in 15 minutes or less.
Target workflow:
- Ask one question.
- Review the top evidence list.
- Inspect support level, limitations, and conflict notes.
- Export or revise a short cited brief.
Definition of done:
- at least 80% of test questions return three or more relevant evidence items
- every answer includes citations, limitations, and support assessment
- weak-support answers are cautious or abstain
- a reviewer can trace each major sentence back to retrieved evidence
Why this matters:
This is the simplest product promise. CANON should make the first brief faster without hiding the review work that remains.
Turn retrieval results into a review queue that tells the human where to spend attention first.
Target workflow:
- Ask a question or upload a draft claim.
- CANON groups evidence into support, qualification, conflict, and weak match.
- The reviewer opens the highest-risk items first.
- The reviewer marks evidence as accepted, rejected, needs review, or missing.
Definition of done:
- every retrieved item has an explanation trace
- conflict and weak-support items are visible above routine supporting evidence
- review decisions can be exported with the answer
- the system records why evidence was accepted or rejected
Why this matters:
The product is not just "answer generation." Its real value is helping a person review faster and with fewer hidden assumptions.
Use CANON as a claim-boundary checker for memos, grant drafts, literature reviews, policy briefs, or product strategy documents.
This is the unique use case: instead of starting with a question, the user starts with a draft paragraph or claim. CANON checks whether the available corpus actually supports the wording.
Target workflow:
- Paste a draft claim or paragraph.
- CANON extracts the major claims.
- For each claim, CANON retrieves supporting, qualifying, and conflicting evidence.
- The system flags statements that look too strong for the available evidence.
- The human rewrites, narrows, cites, or removes the claim.
Definition of done:
- extracted claims are shown separately from the draft text
- each claim receives a support status: supported, qualified, contested, weak, or not found
- at least one suggested safer wording is generated for weak or contested claims
- the final output preserves a review trail linking claims to evidence
Why this matters:
Most teams do not only need answers. They need to know when their own draft is outrunning the evidence. CANON can become useful at the moment before a memo, brief, or proposal leaves the team.
CANON should make the first pass over a source set more efficient. Its output should help the user see which sources deserve review, why they were selected, and what risk each source carries.
Success looks like:
- the top evidence contains genuinely relevant sources
- each source has an explanation trace
- weak, noisy, or off-topic evidence is visible
- reviewers can quickly identify what to read next
CANON should produce a short, cautious answer grounded in retrieved evidence. The draft should cite the evidence it depends on and avoid strong conclusions when support is thin.
Success looks like:
- every substantive sentence is tied to cited evidence
- the answer distinguishes support from qualification or disagreement
- unsupported claims are omitted or marked as limitations
- the draft is useful enough for a human to revise rather than start from zero
CANON should be especially good at saying, "this is not settled." For real knowledge work, that may be more valuable than a smooth answer.
Success looks like:
- conflicting evidence is surfaced instead of averaged away
- low-support answers trigger abstention or caution
- source limitations are part of the response
- users can inspect why the system did not make a stronger claim
CANON should leave behind enough trace information for another reviewer to understand the answer.
Success looks like:
- retrieval policy and score contributors are visible
- cited chunks can be traced back to source records
- generated reports capture corpus limitations
- portfolio/demo outputs are reproducible from documented commands
A human reviewer remains responsible for:
- final conclusions
- citation validity
- domain interpretation
- deciding whether evidence is representative
- handling sensitive, legal, medical, financial, or policy implications
- approving any answer used externally
CANON should assist the reviewer, not bypass them.
- Corpus coverage may be incomplete or biased.
- OpenAlex-derived metadata does not guarantee full-text coverage.
- Retrieval can miss relevant evidence.
- Source-quality signals are proxies, not expert judgments.
- Citation counts and author prominence can reinforce existing field bias.
- Generated answers can still misread or overstate source material.
- Current faithfulness checks are technical aids, not a substitute for human verification.
- Public benchmark validation and expert labels are still needed before broad claims.
The strongest initial use case is not "AI research assistant for everything." It is:
Evidence briefing copilot for literature-backed decisions, with built-in citation review, disagreement surfacing, and claim limits.
Good first users:
- policy and think-tank analysts
- academic research assistants
- market and strategy teams reviewing public literature
- product teams evaluating technical or regulatory claims
- grant, diligence, or memo writers who need cited evidence fast
Poor first users:
- anyone expecting fully autonomous conclusions
- teams without a reviewer who can inspect citations
- high-stakes domains without expert oversight
- broad web search use cases where corpus boundaries are unknown
For GitHub, present CANON as a product-minded evidence workbench:
- focused problem: cited evidence briefing from controlled corpora
- visible workflow: retrieve, compare, draft, inspect, decide
- honest limits: human approval required
- technical depth: evaluation gates, traceable ranking, safety checks, data cards
- product discipline: the system knows when not to overclaim