SOURCES · FRESHNESS · EVIDENCE GATE · SELECTIVE DELIVERY
A scheduled report is not yet analysis.
A useful one changes a decision.
In the Digi Anton AI lab, recurring analytical reviews are assembled by agents and local models. Experience showed that automation which dutifully sends a text every week can be useless or even harmful. Below is how a route works when the report rests on fresh sources, separates facts from interpretation, passes an evidence check and stays honestly silent when nothing material has changed — and what of this actually works today and what does not yet.
Problem
Regularity is easily mistaken for usefulness
An analytical report is valuable not because it arrives on time, but because afterwards a person makes a different decision, sees an uncertainty they had missed, or no longer needs to re-check the sources by hand. A report that arrives on schedule but retells stale material, fills empty sections with generalities or presents a collection failure as “no news” is worse than no report at all: it creates a false sense of control and still requires manual verification. That is why the route is built not around a schedule and a template, but around the question “what changed, how well is it supported and what follows from it”.
Route
From source to delivery in seven stages
Primary and appropriate sources
Primary sources wherever possible: official announcements, documents, data. Each direction needs several independent source families unless a primary source has an objective monopoly — single-source collection wipes out a direction at the first network failure.
Explicit freshness and explicit failure
The collector extracts dated events rather than checking whether landing pages load. Every result carries a date, and a failed attempt carries a failure state. An access error does not count as substantive evidence, and an old cache is never passed off as a fresh review.
A local model works on a bounded slice
The model receives only the material collected for its direction, a clear task and an output format. The model actually used must match the one assigned: silently substituting one model for another before publication is not acceptable.
Fact, interpretation and practical relevance
For every item, separately: what the source says, how the model reads it and why it may matter to the reader. Evidence strength, limitations and uncertainty are stated explicitly.
An evidence and usefulness check
One executable gate before delivery: are there fresh supported items with direct links, are any sources failing, are there unsupported claims, boilerplate or technical lines.
Pending delta and last successful state
A change that has not yet passed the check or has not been delivered is kept as pending and does not become the new comparison baseline. A source failure does not erase the last successful state.
Selective — with the right to stay silent
Substantive material is delivered. Technical noise, duplicates, empty issues and routine “all OK” messages are suppressed. When nothing material has changed, the correct result is a short verified statement or silence, not an invented review.
What went wrong
Broad automation that failed acceptance
A collection error became “no updates”
Prompts suggested writing “no updates” on empty input and did not distinguish an absence of events from a collection failure. The reader could not tell that a direction had not actually been checked.
Old data was presented as the current review
When discovery timed out, an issue was assembled from a stale cache and presented as fresh weekly material.
The template mattered more than the facts
A fixed structure pushed the model to fill empty sections. In one case a static template was sent with no collected items at all and every source failing.
Service lines reached the reader
Any program output, including a technical status, could end up in delivery. A “do not send” flag and the absence of evidence did not block sending.
A successful exit was taken as news
A program’s successful exit code was treated as a material change — with no facts and no links.
There was no single gate
Several independent sending paths existed, and some ignored the delivery block. Some directions had been silently merged into neighbouring ones and had no analysis of their own.
What was done
Stop the noise first, then accept one contour at a time
The review in early September 2026 was unambiguous: no direction worked as a useful end-to-end route, and the result was rejected outright. The first action was not to polish the texts but to stop: all automatic analytical senders were disabled, and internal collection continued without delivery. Mandatory conditions were then set: one shared executable delivery gate with no bypass paths; blocking on source failures, a delivery hold, no supported or fresh evidence, the wrong model, unsupported claims or boilerplate; separate states for “no news” and “data not received”, neither of which produces filler; a dedicated route for each direction; and for each item — what changed, why it matters, how well it is supported and what follows. Each direction is switched on separately, only after a no-send control run and an independent review.
Accepted contour
One bounded weekly delta contour
In late September 2026 one bounded contour was accepted under these rules — a weekly delta for the migration topic. What is confirmed: a local model produced a section from real collector output with direct verifiable links and concluded that there were no material changes; the gate accepted the section with no technical noise and no artificially invented actions. An undelivered delta is now kept as pending and does not become the comparison baseline until a successful check; a source failure no longer deletes the last successful state. A minor change found in one of the sources survived re-collection in the pending state and was confirmed only after it passed the acceptance check run; automatic baseline acknowledgement in production nevertheless remains blocked until the whole weekly package passes the check. Source access errors were not counted as evidence. An independent cloud audit accepted the contour, noting that part of the adjacent subject matter is explicitly left outside the declared automatic coverage.
Honest boundary
What this result does not mean
Not all analytics — one bounded contour has been accepted. The full weekly package remains incomplete and was not accepted by this review.
Not full autonomy — automatic confirmation of a new comparison baseline in production will not trigger until the whole weekly package passes the check. Meanwhile, the pending delta is preserved and does not become the comparison baseline.
Not full topic coverage — some adjacent questions are deliberately left outside automatic monitoring, and this is recorded explicitly rather than hidden.
No guarantee for other directions — the rules are the same, but each direction must pass its own acceptance.
What genuinely exists now — a confirmed capability: a local model produces a verified section from real sources, and the route can correctly say “no material changes” without losing the pending delta.
Roles
Who is responsible for report content
Local model
Extracts facts, groups them and drafts an interpretation within the slice it was given. It is a worker, not an editor entitled to decide what the reader does not need to know.
Codex
Sets the route, context, format and acceptance criteria and checks grounding in sources. A weak model answer is a signal to repair the route, context, prompt or decomposition — not to silently rewrite the report itself.
Evidence gate
Checks links, freshness, source state and the absence of technical text. Its decision does not depend on how confident the model sounds.
Independent audit
Used for important, disputed or final results — for example, accepting a new contour. It is not a mandatory step for every issue, and its findings are checked too.
Reader and decision owner
Sees every material item with evidence-strength labels and can verify any of them independently. Major forks — changing the route, moving work elsewhere, switching delivery on — are decided by a person.
Two failures — change the approach
After two material failures of the same route on the same task class it is not repeated: the model, decomposition, context or acceptance method is changed.
Breadth
Do not compress the field to a single conclusion
The opposite mistake is a report that keeps only one “most important” item. If the period contains other material events, hypotheses or negative results, they are included — even early, disputed or narrowly applicable ones — with a clear label for evidence strength, maturity and limitations. Only duplicates, irrelevant collection noise and unsupported claims are excluded, and important exclusions are disclosed. Report length follows the real evidence: brevity is not a goal in itself, but inventing topics when material is thin is not allowed either. A local model’s weakness does not justify narrowing the visible field — it is a reason to repair the route.
How to read this article
Fact, interpretation and practical takeaway
What the records confirm
Early broad automation failed acceptance: source failures, stale cache, filler and technical statuses reached delivery. Later one bounded weekly delta contour was accepted; the full weekly package remains incomplete.
What follows
The quality of an autonomous report is determined not by the model or the schedule, but by whether the route distinguishes failure from an absence of events, preserves state and is able to block delivery.
What to do
Stop the noise first, then introduce a single gate, separate fact from interpretation, keep the pending delta and switch directions on one at a time after a no-send check.
For a company
When an automated report is genuinely useful
The same approach applies to industry monitoring, regulatory-change reviews, competitive intelligence and internal management analytics. The test for any recurring report is simple: does it change a decision, expose uncertainty, cite verifiable evidence and reduce manual re-checking? If a report merely arrives on schedule, its value is zero and its hidden cost is the time of the people who read and re-check it. Practical requirements for such a contour: every claim is tied to a source and a date; a collection failure is visible as a failure; “no changes” is a valid, verified result; neither a local nor a cloud model decides for the reader what they should know; and rollout and expansion proceed one direction at a time after a control run. How models and checks are distributed in this scheme is covered in the article on routing and quality gates.
Boundaries
What this approach does not promise
It does not make all analytics autonomous and does not guarantee that every direction will pass acceptance. It is a working practice in an experimental lab: one bounded contour has been accepted, the full weekly package remains incomplete, and some topics are deliberately left outside automatic coverage. What holds is the order in which a source, explicit freshness, bounded analysis, separation of fact and interpretation, an evidence gate and preserved state produce a report that can be trusted — or honest silence. Related reading: Digi Anton, local models, Codex and quality gates and shared AI-agent memory.
How routing and quality gates work →