venutian-antfarm

Feedback Guide

Behavioral feedback is recorded and queried via ops/feedback-log.sh. The script is the single entry point for all ledger writes and reads — agents never edit the ledger directly.

Two tracks exist: governance and core agents issue feedback directly; specialist agents propose it for supervisor review.

Direct Feedback (Governance and Core Agents)

`reprimand`

Records a negative behavioral signal. All fields are required.

ops/feedback-log.sh reprimand \
  --issuer <agent> \
  --subject <agent> \
  --domain <domain> \
  --severity low|medium|high \
  --description "What happened" \
  --evidence "Concrete reference" \
  [--item <id>]

Example:

ops/feedback-log.sh reprimand \
  --issuer cro \
  --subject compliance-auditor \
  --domain security \
  --severity medium \
  --description "Skipped evidence collection before closing finding" \
  --evidence "Finding F-017 closed with no attached evidence block" \
  --item 38

`kudo`

Records a positive behavioral signal. --severity is not required.

ops/feedback-log.sh kudo \
  --issuer <agent> \
  --subject <agent> \
  --domain <domain> \
  --description "What was good" \
  --evidence "Concrete reference" \
  [--item <id>]

Both subcommands print reward_id=<R-NNN|K-NNN> on success and emit a reward-issued metric event. If a new signal conflicts with an existing one (same subject, same item, opposing type), a tension entry is auto-generated.

Specialist Proposals

Specialists cannot issue feedback directly. Instead they recommend, and the routed supervisor formalizes or rejects.

`recommend`

Proposes feedback. The script resolves the supervisor from fleet-config.json (pathways.declared.feedback) and records a P-NNN proposal with a 7-day escalation deadline.

ops/feedback-log.sh recommend \
  --issuer <specialist-agent> \
  --subject <agent> \
  --type kudo|reprimand \
  --domain <domain> \
  --description "..." \
  --evidence "..." \
  [--severity low|medium|high]   # required when --type reprimand
  [--item <id>]

Prints proposal_id=P-NNN on success.

`formalize`

Supervisor approves a pending proposal. Converts the proposal to a K-NNN or R-NNN entry and marks the proposal as formalized.

ops/feedback-log.sh formalize P-001 --issuer <supervisor-agent>

The issuer must match the supervisor recorded on the proposal. Prints reward_id=<K|R>-NNN.

`reject`

Supervisor declines a pending proposal.

ops/feedback-log.sh reject P-001 --issuer <supervisor-agent> --reason "Insufficient evidence"

--reason is required. Marks the proposal as rejected — <reason>.

Escalation

`check-escalations`

Scans the ledger for pending proposals whose deadline has passed and escalates them to the next agent in the governance pathway. Each escalation extends the deadline by 7 days, updates the supervisor, adds a finding to the register, and emits a feedback-escalated metric.

ops/feedback-log.sh check-escalations

Run this periodically — the SM typically calls it before or during a retro. No arguments needed.

Queries

`profile`

Human-readable behavioral summary for an agent: signal counts, breakdown by domain and tier, 5 most recent entries, pending proposal count, and weighted score.

ops/feedback-log.sh profile <agent>

Example output:

compliance-auditor: 4 kudos, 2 reprimands, 1 tensions (1 open)
By domain:
  security (3)
  process (2)
  delivery (1)
Recent:
  R-004 [reprimand] 2026-03-20 — cro / security
  K-005 [kudo] 2026-03-22 — coo / delivery
  ...
Weighted: net=3.2 (kudos=4.0, reprimands=-0.8, 5 of 6 recent)

`score`

Machine-parseable version of the weighted score. Useful for automation or dashboard scripting.

ops/feedback-log.sh score <agent>

Output lines: net=, kudos=, reprimands=, signals=, recent=.

`tensions`

Lists open tension entries — cases where opposing feedback (a kudo and a reprimand) exists for the same subject on the same item. Optionally filter by item.

ops/feedback-log.sh tensions
ops/feedback-log.sh tensions --item 38

The SM should surface open tensions during retros. The CRO reviews patterns during compliance audits.

Weighted Scoring

The score is not a simple count. Each signal is weighted across three axes:

Tier multiplier — governance (1.0) > core (0.8) > specialist (0.5). Feedback from a governance agent carries more weight.
Type multiplier — reprimands (1.5) are weighted higher than kudos (1.0). A single reprimand has more pull than a kudo.
Domain multiplier — configurable per domain and per agent in fleet-config.json under rewards.weighting.domain_multipliers. Defaults to 1.0.

Decay applies a step function: signals older than cliff_items accepted items drop to post_cliff_multiplier (default: beyond 10 accepted items, weight drops to 0.25x). This prevents old signals from dominating the score indefinitely.

Defaults come from the script. Override them in fleet-config.json:

"rewards": {
  "weighting": {
    "tier_multipliers": { "governance": 1.0, "core": 0.8, "specialist": 0.5 },
    "type_multipliers": { "reprimand": 1.5, "kudo": 1.0 },
    "domain_multipliers": { "security": 1.2, "_default": 1.0 },
    "decay": { "cliff_items": 10, "post_cliff_multiplier": 0.25 }
  },
  "escalation_deadline_days": 7
}

Subcommand Reference

Subcommand	Who	What it does
`reprimand`	Governance / core	Record a negative signal directly
`kudo`	Governance / core	Record a positive signal directly
`recommend`	Specialist	Propose feedback, route to supervisor
`formalize`	Supervisor	Approve a pending proposal
`reject`	Supervisor	Decline a pending proposal
`check-escalations`	Any (SM by convention)	Auto-escalate stale proposals
`profile`	Any	Human-readable behavioral summary
`score`	Any	Machine-parseable weighted score
`tensions`	Any	List open conflicting-signal entries

docs/GOVERNANCE-FLOORS.md — How behavioral feedback ties into floor enforcement.
.claude/COLLABORATION.md — Collaboration protocol and agent roles.
fleet-config.json — Pathway declarations and scoring configuration.

This site is open source. Improve this page.

venutian-antfarm

Feedback Guide

Direct Feedback (Governance and Core Agents)

reprimand

kudo

Specialist Proposals

recommend

formalize

reject