CounselStreamDocs

Reference

Data Dictionary

The 37 columns an export carries: what each states, its format, the values CounselStream reads, and when it is required.

Rules for every cell

  • Every cell holds a value, or n/a (this column does not apply to this row), or undeclared (not known yet; CounselStream then asks the system's owner).
  • undeclared is allowed in every column except kind, name, platform_id and snapshot_at.
  • An empty cell is not an answer. The record is still accepted; the empty cell is read as not yet known and counted as a defect in the file.
  • Dates are YYYY-MM-DD. Every timestamp is ISO 8601 in UTC and ends in Z, for example 2026-11-02T08:00:00Z. A timestamp with no zone is read as not yet known.
  • These are the default column names. If your export already names a column differently, keep your name and tell us at onboarding; CounselStream reads your name.

Identity

platform_id

Your platform's own stable identifier for the row: the same value in every export for the life of the row, never reused. CounselStream matches each export to the register on it, because names repeat.

Format
text
Values
Any stable identifier.
Required
Always, on every row, including trigger rows and the platform row.
Example
a3f1c2d4-5b6e-4f70-8a91-0c2d3e4f5a6b

name

The system's name as people know it. Names need not be unique, so a name is never used to match rows.

Format
free text
Values
Any text that is not empty.
Required
Always.
Example
order-summary-agent

kind

What the row is in your platform.

Format
one word
Values
  • agent a configured agent that can call tools
  • workflow a defined sequence of steps or agents
  • trigger a schedule, event or call that starts an agent or workflow
  • platform the shared runtime and model record
  • model, algorithm, pipeline a model or algorithm you host or train
Required
Always.
Example
agent

owner

The person accountable for how the system works. CounselStream addresses its questions about the system to this person.

Format
name, team, email
Values
  • A named person: not a shared mailbox or a system identity
  • undeclared no owner has been assigned yet
Required
Always.

last_editor

Who or what last changed the definition. Kept as provenance. It is never shown as the owner.

Format
email, or a system identity
Values
  • A person's email, or a system identity: whoever made the last change
  • undeclared not known
Required
Always.
Example
migration

snapshot_at

When the export ran.

Format
ISO 8601 timestamp in UTC ending in Z
Values
One value for each export, the same on every row.
Required
Always.
Example
2026-11-02T08:00:00Z

What it is

built_by

Who built the system.

Format
one word
Values
  • own_build built by your own code, prompts and tools
  • vendor_configured_by_us a purchased product you configure
  • vendor_as_delivered a purchased product used as it comes
Required
Always.
Example
own_build

shown_under_company_name

Whether people outside your company meet the system under your company's name.

Format
one word
Values
  • internal only your staff meet it
  • external people outside your company meet it
  • n/a a trigger row
Required
Always. n/a on a trigger row.
Example
external

definition_snapshot

The versioned definition of the whole system. CounselStream compares consecutive versions to show what changed.

Format
v<N> <identifier> by <who> at <timestamp>; skills <name@version, ...>; documents <name@version, ...>; build <commit>
Values
N: a whole number. identifier: your identifier for that version. skills and documents: the versions of prompts and instructions. build: the code commit, or unknown where it is not recorded.
Required
Always.

first_run_at

When the system first ran in production.

Format
ISO 8601 timestamp in UTC ending in Z, or the form shown under Values
Values
  • A timestamp: when it first ran
  • unknown, predates history (<timestamp>) the system ran before your history begins, and the timestamp is where your history starts
  • n/a the system has not run yet
Required
Always.
Example
2026-03-18T14:05:00Z

retired_at

When the system was retired. Never delete a retired system's row. Set this instead.

Format
ISO 8601 timestamp in UTC ending in Z
Values
  • A timestamp: the system is retired from that moment
  • n/a the system is live
Required
Always.
Example
n/a

runs_in_window

Whether the system ran in the period the export covers.

Format
N runs, M failed, window D d
Values
N, M and D are whole numbers. 0 runs means the system did not run in the period.
Required
Always.
Example
3 runs, 0 failed, window 30d

Where it runs

where_it_runs

For agents and workflows, the model providers and the route used to reach them. For the platform row, where the runtime is hosted.

Format
comma list of "provider via route"; platform row: runtime and region
Values
Provider names and routes as your router names them, including a gateway, a router or a cloud provider's catalogue.
  • direct the provider's own API
  • self_hosted a model you host yourself
  • n/a a trigger row
Required
Always. n/a on a trigger row.

model_route

Which models the system calls, as requested and as resolved by your router.

Format
comma list of "requested -> resolved", with (legacy) after an entry kept for old definitions
Values
Model identifiers as your router names them.
  • n/a a trigger row, or a system with no model
Required
Always. n/a on a trigger row or a system with no model.

What starts it and what it can do

starters

What starts a run: an API call, a person's message, a schedule, a trigger, another agent or workflow. An entry that refers to another row gives that row's platform_id, never its name.

Format
list separated by semicolons, then "; observed: <kind>=<count>"
Values
A channel is written as your platform names it: a person's message or an event in a tool.
  • api another system called it through the API
  • mcp an agent or tool called it
  • none nothing starts it
  • cron <expression> (disabled) a schedule, and whether it is switched off
  • agent_run, workflow_run another agent or workflow started it
  • trigger <platform_id> that trigger row starts it; one entry for each trigger
  • step of <platform_id> that workflow uses this row as a step; one entry for each
  • delegated by <platform_id> that agent hands work to this row; one entry for each
  • tool of <platform_id> on a model, algorithm or pipeline row, that agent or workflow uses it as a tool; one entry for each
The observed part counts real runs by kind.
Required
Always.

takes_actions

Whether the system can act, not only answer.

Format
one word
Values
  • yes it can act
  • no it only returns answers
  • n/a does not apply to this row
  • undeclared not known yet, which is read as unknown and never as no
Required
Always.
Example
yes

tool_allowlist

The tools the system is allowed to call. A tool listed by several systems is one shared tool.

Format
comma list of tool names
Values
Tool names as your registry names them.
  • none the system has no tools
Required
Always. none where the system has no tools.

effects

What a tool call can change outside the system.

Format
comma list of flags, or a coverage statement
Values
Any of the five flags, spelled as your registry spells them.
  • money the tool can move, hold or alter money or its terms for a person
  • publish the tool can make content or a listing visible to others, or change what is surfaced or ranked
  • account_status the tool can change a person's access or standing
  • external_message the tool can send a message outside your company
  • datastore_write the tool can write to a data store
  • none the tool has no effect outside the system
  • undeclared for N of M tools; declared: <flags> kept as partial coverage
undeclared is never read as none.
Required
Always.

Human oversight

approval_gate

Whether a person approves before the output takes effect.

Format
comma list of "mode: tool", or none
Values
  • blocking: <tool> the output waits for a person at that tool
  • non-blocking: <tool> the person is shown the action and the output proceeds
  • none there is no approval step
Required
Always.
Example
blocking: request_approval

approval_record

Where approvals are recorded, and how many happened in the period the export covers.

Format
store names, then "in window: approved / total"
Values
Store or table names, and counts as whole numbers.
  • none no approval is recorded anywhere
Required
Always.

stop_mechanism

How the system is stopped, and by what call.

Format
free text naming the call
Values
The call, flag or action that stops the system.
  • unknown no way of stopping it has been identified
Required
Always.

Output

output_delivery

How the output leaves the system.

Format
comma list of transports
Values
Transports as your platform names them.
  • n/a a trigger row
Required
Always. n/a on a trigger row.

conversational

Whether people talk to the system, and through what channel.

Format
one word, with the channel in parentheses
Values
  • no people do not talk to it
  • yes (<channel>) people talk to it through that channel; several channels may be listed
Required
Always.
Example
yes (api)

generates_media

Whether the system generates content, and what kind.

Format
one word
Values
  • no it generates none
  • text, image, video, audio the kind it generates
Required
Always.
Example
no

output_type

What the system emits.

Format
one word, or "workflow <field names>"
Values
  • text free text
  • workflow <field names> structured fields, and the list is the schema
  • n/a the system emits nothing
Required
Always.
Example
text

person_bearing

Whether what the system produces concerns identifiable people, as the system's owner declares it.

Format
one word
Values
  • yes it does
  • no it does not
  • undeclared the owner has not said yet
The test is given to your team at onboarding.
Required
Always.
Example
undeclared

Memory, training and evaluation

data_in_out

Whether the system keeps memory across runs, and where. What personal information flows through the system is not asked of your engineers.

Format
undeclared; memory: <value>
Values
The first part is always undeclared.
  • memory: default the platform's standard memory
  • memory: none no memory
  • memory: <store name> a named store
Required
Always, for the memory part.
Example
undeclared; memory: none

training_eval_use

Whether data this system reads or produces is used to train a model, and how much evaluation has been run.

Format
<training part>; eval runs N, judge reports N
Values
The training part is one of the three below. N values are whole numbers.
  • n/a no data from this system trains a model
  • yes <dataset identifiers> it does, and which datasets
  • planned it is planned
Required
Always.
Example
n/a; eval runs 0, judge reports 0

last_eval

The state of evaluation, and the safeguards switched on for the system.

Format
no eval run; no judge report; assignments N; then each safeguard as "<name> <true or false>"
Values
  • true the safeguard is on
  • false it is off
  • Blank: not known
Required
Always.

eval_metrics

The performance the system is expected to meet and the latest measured value, for each metric, with the date.

Format
entries separated by semicolons: metric[attribute]=value (target) @ date. [attribute] is optional and names what the result compares people across. Write (none) where there is no target.
Values
Any metric name, with the unit in the name where it is not a fraction. Numbers as your tool reports them.
  • not_measured the system's output concerns identifiable people and no scored evaluation exists
  • n/a its output does not concern identifiable people; also every trigger row and the platform row
  • undeclared not known yet
The test for whether a system's output concerns identifiable people is given to your team at onboarding.
Required
Always.
Example
pass_rate=0.91 (none) @ 2026-10-01

group_comparison

Whether the system's performance has been compared across groups of people, by whom, and with what measure. Measured results go in eval_metrics.

Format
measured <measure> by <own, vendor or independent_auditor>, repeated with semicolons if several; or not_measured; or n/a
Values
  • by own your team ran the comparison, and eval_metrics has a matching [attribute] entry
  • by vendor the model or product vendor ran it
  • by independent_auditor an outside auditor ran it
  • not_measured, n/a, undeclared as in eval_metrics
Required
Always.

logs

Where the system's logs are, what each location holds, and how long each is kept.

Format
three locations separated by pipes, with retention in days in parentheses on the log group
Values
A governance page | a run history query | a log group with its retention as (N d).
  • none there is no location for that position
Required
Always.

Models you trained or changed

adaptation_method

How each model behind the system was adapted. Write one value for each model on the row, in the same order as model_route. For one model, pick the first value in the order given under Values that applies.

Format
one word for each model_route entry, separated by commas
Values
  • trained_from_scratch a model whose weights you initialised and trained yourself
  • fine_tuned you started from published or vendor weights and changed them by further training
  • distilled a smaller model trained to imitate a larger model's outputs
  • classical_ml a statistical or machine-learning model that is not a language model, trained by you from your own data
  • prompt_or_retrieval_only the weights are unchanged, and you configured a system prompt, instructions, retrieved documents, tools or guardrails
  • none_used_as_is a model called with none of those configured
  • rules_or_statistical a row with no model call whose logic is fixed rules or formulas
  • n/a a trigger row, or a workflow row that calls no model of its own
On the platform row, write one value for each model the platform calls, in the order of model_route.
Required
Always. Never empty.
Example
prompt_or_retrieval_only

base_model

The published or vendor model that a fine-tuned or distilled model started from.

Format
<provider>/<model identifier>, as your model registry names it
Values
Any identifier.
  • n/a adaptation_method is anything other than fine_tuned or distilled
Required
When adaptation_method is fine_tuned or distilled. Otherwise n/a. One entry for each model, in model_route order.

weights_source_and_licence

Where the model's weights came from and under what licence.

Format
<source>; <licence>
Values
The licence is its published title; internal for weights you trained from scratch; vendor_terms for vendor_hosted. The source is one of:
  • own_training weights you trained from scratch
  • vendor_download downloaded from a vendor under a commercial licence
  • open_weights published weights under an open or community licence
  • vendor_hosted a vendor's fine-tuning service whose weights you never hold
Required
When adaptation_method is fine_tuned, distilled, trained_from_scratch or classical_ml. Otherwise n/a.

training_run_id

The identifier of the training run that produced the model version in service, so that the training record can be found.

Format
text, as your training pipeline or the vendor names runs
Values
Any run identifier.
Required
When adaptation_method is fine_tuned, distilled, trained_from_scratch or classical_ml. Otherwise n/a.

training_data_refs

The identifiers of the datasets the model was trained or fine-tuned on. Identifiers only. Whether a dataset holds personal information is not asked of your engineers.

Format
comma list of dataset identifiers, as your data platform names them
Values
Any identifiers.
Required
When adaptation_method is fine_tuned, distilled, trained_from_scratch or classical_ml. Otherwise n/a.

These pages describe a data contract: what to send and how it is read as data. They do not describe how any legal determination is reached. Determinations are made by privacy counsel inside CounselStream.