SiliconBridge AI Resources

How validated handoffs keep defects contained.

Wherever your team is on the curve and in your AI journey, the discipline is the same.

Six Silent Defects in an Agent-Generated Timing Report

A coding agent was asked to report every macro in an SoC integration with negative setup slack at a stated corner. Its report is reproduced below, followed by the findings of an independent validation pass.

Prompt

List every macro in this SoC integration whose worst case setup slack is negative at the SS/0.72V/125C corner, sorted by worst slack.

Agent output, as delivered

Fourteen rows returned

Blue markers tie each location in the report to its numbered finding below.

	Instance	Block	Corner	Worst slack (ns)	Path endpoint
	u_ddrphy_dfi_pll	DDR_PHY		not returned
01	u_npu_l2_sram_bank3	NPU_TOP	SS/0.72V/125C	-0.412	mac_pipe_reg[41]/D
02	u_vdec_mbuf_2p_a	VDEC	SS/0.72V/125C	~-0.361	ref_line_reg[7]/D
03	u_memc_wdata_fifo	MEMC	SS/0.72V/125C	-0.344	wr_ptr_reg[3]/D
04	u_serdes_rx_eq_ram	SERDES_Q0	SSG/0.72V/125C	-0.302	eq_tap_reg[12]/D
05	u_npu_wt_rom_a	NPU_TOP	SS/0.72V/125C	~-0.287	wt_dout_reg[63]/D
06	u_isp_line_buf_2	ISP	SS/0.72V/125C	-0.251	ln_acc_reg[9]/D
07	u_pcie_replay_ram	PCIE_SS	SS/0.72V/125C	-0.204	rply_ptr_reg[5]/D
08	u_usb3_pkt_sram	USB3_SS	SS/0.72V/125C	-0.219	pkt_len_reg[8]/D
09	u_memc_ecc_lut	MEMC	SS/0.75V/125C	~-0.188	syn_reg[13]/D
10	u_isp_stat_sram_1p	ISP	SS/0.72V/125C	-0.156	hist_bin_reg[22]/D
11	u_usb3_ref_pll	USB3_SS	SS/0.72V/125C	-0.156	div_ctl_reg[2]/D
12	u_npu_act_sram_b	NPU_TOP	SSG/0.72V/125C	-0.121	act_q_reg[17]/D
13	u_vdec_ctx_2p	VDEC	SS/0.72V/125C	-0.094	ctx_addr_reg[6]/D
14	u_memc_rdata_skid	MEMC	SS/0.72V/125C	-0.041	rd_vld_reg/D

First of two runs on identical inputs; the second run is examined in Finding 6.

Agent output, run through the SBAI Handoff Validation System

Finding	Where	Checked against	Result
Manifest coverage	Absent row	Netlist manifest	u_ddrphy_dfi_pll appears in the netlist manifest but is absent from the report. Fourteen of the fifteen in-scope macros are present.
Declared sort order	Rows 07 and 08	Declared sort key	Row 07 (-0.204) precedes row 08 (-0.219), violating the declared sort order.
Measured value distinctness	Rows 10 and 11	Measured column	Two structurally unrelated macros report identical slack (-0.156 ns) in a continuously measured column.
Uncertainty marking	Worst slack column	Two runs, identical inputs	Values marked approximate reproduce exactly across runs, while values printed as exact do not. The uncertainty marking is inverted.
Predicate declaration	Corner column	Request predicate	Rows 04, 09, and 12 report a corner other than the one specified in the request.
Run to run reproducibility	Run stamp	Two runs, identical inputs	A second run on identical inputs omits row 06 and shifts three slack values by approximately 1.5 percent.

Method

Every finding above can be re-derived by hand from the request and the report alone. None of them depends on the judgment of a model, ours or anyone else's. The frameworks behind this check, spanning governance, data freshness, and shared memory, are available to SBAI customers. Request more information to go deeper.

This example uses synthetic, plausible data. It contains no real design, tool output, or customer information.